// Copyright 2021 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include void xnn_f16_f32_vcvt_ukernel__sse2_int16_x32( size_t n, const void* input, float* output, const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(uint16_t) == 0); assert(input != NULL); assert(output != NULL); const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse_int16.sign_mask); const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int16.exp_offset); const __m128 vexp_scale = _mm_load_ps(params->sse_int16.exp_scale); const __m128i vmagic_mask = _mm_load_si128((const __m128i*) params->sse_int16.magic_mask); const __m128 vmagic_bias = _mm_load_ps(params->sse_int16.magic_bias); const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int16.denorm_cutoff); const uint16_t* i = (const uint16_t*) input; for (; n >= 32 * sizeof(uint16_t); n -= 32 * sizeof(uint16_t)) { const __m128i vh0 = _mm_loadu_si128((const __m128i*) i); const __m128i vh1 = _mm_loadu_si128((const __m128i*) (i + 8)); const __m128i vh2 = _mm_loadu_si128((const __m128i*) (i + 16)); const __m128i vh3 = _mm_loadu_si128((const __m128i*) (i + 24)); i += 32; const __m128i vsign0 = _mm_and_si128(vh0, vsign_mask); const __m128i vsign1 = _mm_and_si128(vh1, vsign_mask); const __m128i vsign2 = _mm_and_si128(vh2, vsign_mask); const __m128i vsign3 = _mm_and_si128(vh3, vsign_mask); const __m128i vnonsign0 = _mm_xor_si128(vh0, vsign0); const __m128i vnonsign1 = _mm_xor_si128(vh1, vsign1); const __m128i vnonsign2 = _mm_xor_si128(vh2, vsign2); const __m128i vnonsign3 = _mm_xor_si128(vh3, vsign3); const __m128i vprenorm0 = _mm_slli_epi16(vnonsign0, 13); const __m128i vprenorm1 = _mm_add_epi16(_mm_srli_epi16(vnonsign0, 3), vexp_offset); const __m128i vprenorm2 = _mm_slli_epi16(vnonsign1, 13); const __m128i vprenorm3 = _mm_add_epi16(_mm_srli_epi16(vnonsign1, 3), vexp_offset); const __m128i vprenorm4 = _mm_slli_epi16(vnonsign2, 13); const __m128i vprenorm5 = _mm_add_epi16(_mm_srli_epi16(vnonsign2, 3), vexp_offset); const __m128i vprenorm6 = _mm_slli_epi16(vnonsign3, 13); const __m128i vprenorm7 = _mm_add_epi16(_mm_srli_epi16(vnonsign3, 3), vexp_offset); const __m128i vnorm0 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm0, vprenorm1)), vexp_scale)); const __m128i vnorm1 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm0, vprenorm1)), vexp_scale)); const __m128i vnorm2 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm2, vprenorm3)), vexp_scale)); const __m128i vnorm3 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm2, vprenorm3)), vexp_scale)); const __m128i vnorm4 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm4, vprenorm5)), vexp_scale)); const __m128i vnorm5 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm4, vprenorm5)), vexp_scale)); const __m128i vnorm6 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm6, vprenorm7)), vexp_scale)); const __m128i vnorm7 = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm6, vprenorm7)), vexp_scale)); const __m128i vdenorm0 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign0, vmagic_mask)), vmagic_bias)); const __m128i vdenorm1 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign0, vmagic_mask)), vmagic_bias)); const __m128i vdenorm2 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign1, vmagic_mask)), vmagic_bias)); const __m128i vdenorm3 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign1, vmagic_mask)), vmagic_bias)); const __m128i vdenorm4 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign2, vmagic_mask)), vmagic_bias)); const __m128i vdenorm5 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign2, vmagic_mask)), vmagic_bias)); const __m128i vdenorm6 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign3, vmagic_mask)), vmagic_bias)); const __m128i vdenorm7 = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign3, vmagic_mask)), vmagic_bias)); const __m128i vmask0 = _mm_cmpgt_epi16(vnonsign0, vdenorm_cutoff); const __m128i vmask1 = _mm_cmpgt_epi16(vnonsign1, vdenorm_cutoff); const __m128i vmask2 = _mm_cmpgt_epi16(vnonsign2, vdenorm_cutoff); const __m128i vmask3 = _mm_cmpgt_epi16(vnonsign3, vdenorm_cutoff); const __m128i vxmask0 = _mm_unpacklo_epi16(vmask0, vmask0); const __m128i vf0 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign0), _mm_or_si128(_mm_and_si128(vxmask0, vnorm0), _mm_andnot_si128(vxmask0, vdenorm0))); const __m128i vxmask1 = _mm_unpackhi_epi16(vmask0, vmask0); const __m128i vf1 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign0), _mm_or_si128(_mm_and_si128(vxmask1, vnorm1), _mm_andnot_si128(vxmask1, vdenorm1))); const __m128i vxmask2 = _mm_unpacklo_epi16(vmask1, vmask1); const __m128i vf2 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign1), _mm_or_si128(_mm_and_si128(vxmask2, vnorm2), _mm_andnot_si128(vxmask2, vdenorm2))); const __m128i vxmask3 = _mm_unpackhi_epi16(vmask1, vmask1); const __m128i vf3 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign1), _mm_or_si128(_mm_and_si128(vxmask3, vnorm3), _mm_andnot_si128(vxmask3, vdenorm3))); const __m128i vxmask4 = _mm_unpacklo_epi16(vmask2, vmask2); const __m128i vf4 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign2), _mm_or_si128(_mm_and_si128(vxmask4, vnorm4), _mm_andnot_si128(vxmask4, vdenorm4))); const __m128i vxmask5 = _mm_unpackhi_epi16(vmask2, vmask2); const __m128i vf5 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign2), _mm_or_si128(_mm_and_si128(vxmask5, vnorm5), _mm_andnot_si128(vxmask5, vdenorm5))); const __m128i vxmask6 = _mm_unpacklo_epi16(vmask3, vmask3); const __m128i vf6 = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign3), _mm_or_si128(_mm_and_si128(vxmask6, vnorm6), _mm_andnot_si128(vxmask6, vdenorm6))); const __m128i vxmask7 = _mm_unpackhi_epi16(vmask3, vmask3); const __m128i vf7 = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign3), _mm_or_si128(_mm_and_si128(vxmask7, vnorm7), _mm_andnot_si128(vxmask7, vdenorm7))); _mm_storeu_ps(output, _mm_castsi128_ps(vf0)); _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf1)); _mm_storeu_ps(output + 8, _mm_castsi128_ps(vf2)); _mm_storeu_ps(output + 12, _mm_castsi128_ps(vf3)); _mm_storeu_ps(output + 16, _mm_castsi128_ps(vf4)); _mm_storeu_ps(output + 20, _mm_castsi128_ps(vf5)); _mm_storeu_ps(output + 24, _mm_castsi128_ps(vf6)); _mm_storeu_ps(output + 28, _mm_castsi128_ps(vf7)); output += 32; } for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { const __m128i vh = _mm_loadu_si128((const __m128i*) i); i += 8; const __m128i vsign = _mm_and_si128(vh, vsign_mask); const __m128i vnonsign = _mm_xor_si128(vh, vsign); const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13); const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset); const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale)); const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale)); const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias)); const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias)); const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff); const __m128i vxmask_lo = _mm_unpacklo_epi16(vmask, vmask); const __m128i vf_lo = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign), _mm_or_si128(_mm_and_si128(vxmask_lo, vnorm_lo), _mm_andnot_si128(vxmask_lo, vdenorm_lo))); const __m128i vxmask_hi = _mm_unpackhi_epi16(vmask, vmask); const __m128i vf_hi = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign), _mm_or_si128(_mm_and_si128(vxmask_hi, vnorm_hi), _mm_andnot_si128(vxmask_hi, vdenorm_hi))); _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo)); _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi)); output += 8; } if XNN_UNPREDICTABLE(n != 0) { const __m128i vh = _mm_loadu_si128((const __m128i*) i); const __m128i vsign = _mm_and_si128(vh, vsign_mask); const __m128i vnonsign = _mm_xor_si128(vh, vsign); const __m128i vprenorm_lo = _mm_slli_epi16(vnonsign, 13); const __m128i vprenorm_hi = _mm_add_epi16(_mm_srli_epi16(vnonsign, 3), vexp_offset); const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale)); const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vprenorm_lo, vprenorm_hi)), vexp_scale)); const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpacklo_epi16(vnonsign, vmagic_mask)), vmagic_bias)); const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_unpackhi_epi16(vnonsign, vmagic_mask)), vmagic_bias)); const __m128i vmask = _mm_cmpgt_epi16(vnonsign, vdenorm_cutoff); const __m128i vxmask_lo = _mm_unpacklo_epi16(vmask, vmask); __m128i vf = _mm_or_si128(_mm_unpacklo_epi16(_mm_setzero_si128(), vsign), _mm_or_si128(_mm_and_si128(vxmask_lo, vnorm_lo), _mm_andnot_si128(vxmask_lo, vdenorm_lo))); if (n & (4 * sizeof(uint16_t))) { _mm_storeu_ps(output, _mm_castsi128_ps(vf)); output += 4; const __m128i vxmask_hi = _mm_unpackhi_epi16(vmask, vmask); vf = _mm_or_si128(_mm_unpackhi_epi16(_mm_setzero_si128(), vsign), _mm_or_si128(_mm_and_si128(vxmask_hi, vnorm_hi), _mm_andnot_si128(vxmask_hi, vdenorm_hi))); } if (n & (2 * sizeof(uint16_t))) { _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf)); output += 2; vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf))); } if (n & (1 * sizeof(uint16_t))) { _mm_store_ss(output, _mm_castsi128_ps(vf)); } } } void xnn_f16_vabs_ukernel__sse2_x16( size_t n, const void* input, void* output, const union xnn_f16_abs_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(uint16_t) == 0); assert(input != NULL); assert(output != NULL); const uint16_t* i = (const uint16_t*) input; uint16_t* o = (uint16_t*) output; const __m128i vnonsign_mask = _mm_load_si128((const __m128i*) params->sse.nonsign_mask); for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) { __m128i vacc0 = _mm_loadu_si128((const __m128i*) i); __m128i vacc1 = _mm_loadu_si128((const __m128i*) (i + 8)); i += 16; vacc0 = _mm_and_si128(vacc0, vnonsign_mask); vacc1 = _mm_and_si128(vacc1, vnonsign_mask); _mm_storeu_si128((__m128i*) o, vacc0); _mm_storeu_si128((__m128i*) (o + 8), vacc1); o += 16; } for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { __m128i vacc = _mm_loadu_si128((const __m128i*) i); i += 8; vacc = _mm_and_si128(vacc, vnonsign_mask); _mm_storeu_si128((__m128i*) o, vacc); o += 8; } if XNN_UNLIKELY(n != 0) { __m128i vacc = _mm_loadu_si128((const __m128i*) i); vacc = _mm_and_si128(vacc, vnonsign_mask); if (n & (4 * sizeof(uint16_t))) { _mm_storel_epi64((__m128i*) o, vacc); o += 4; vacc = _mm_unpackhi_epi64(vacc, vacc); } if (n & (2 * sizeof(uint16_t))) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vacc)); o += 2; vacc = _mm_srli_epi64(vacc, 32); } if (n & (1 * sizeof(uint16_t))) { *o = (uint16_t) _mm_extract_epi16(vacc, 0); } } } void xnn_f16_vneg_ukernel__sse2_x16( size_t n, const void* input, void* output, const union xnn_f16_neg_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(uint16_t) == 0); assert(input != NULL); assert(output != NULL); const uint16_t* i = (const uint16_t*) input; uint16_t* o = (uint16_t*) output; const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse.sign_mask); for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) { __m128i vacc0 = _mm_loadu_si128((const __m128i*) i); __m128i vacc1 = _mm_loadu_si128((const __m128i*) (i + 8)); i += 16; vacc0 = _mm_xor_si128(vacc0, vsign_mask); vacc1 = _mm_xor_si128(vacc1, vsign_mask); _mm_storeu_si128((__m128i*) o, vacc0); _mm_storeu_si128((__m128i*) (o + 8), vacc1); o += 16; } for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { __m128i vacc = _mm_loadu_si128((const __m128i*) i); i += 8; vacc = _mm_xor_si128(vacc, vsign_mask); _mm_storeu_si128((__m128i*) o, vacc); o += 8; } if XNN_UNLIKELY(n != 0) { __m128i vacc = _mm_loadu_si128((const __m128i*) i); vacc = _mm_xor_si128(vacc, vsign_mask); if (n & (4 * sizeof(uint16_t))) { _mm_storel_epi64((__m128i*) o, vacc); o += 4; vacc = _mm_unpackhi_epi64(vacc, vacc); } if (n & (2 * sizeof(uint16_t))) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vacc)); o += 2; vacc = _mm_srli_epi64(vacc, 32); } if (n & (1 * sizeof(uint16_t))) { *o = (uint16_t) _mm_extract_epi16(vacc, 0); } } } void xnn_f32_argmaxpool_ukernel_4x__sse2_c4( size_t output_pixels, size_t pooling_elements, size_t channels, const float** input, size_t input_offset, float* output, uint32_t* index, size_t input_increment, size_t output_increment) XNN_OOB_READS { assert(output_pixels != 0); assert(pooling_elements != 0); assert(pooling_elements <= 4); assert(channels != 0); do { const float* i0 = input[0]; const float* i1 = input[1]; const float* i2 = input[2]; const float* i3 = input[3]; i0 = (const float*) ((uintptr_t) i0 + input_offset); i1 = (const float*) ((uintptr_t) i1 + input_offset); i2 = (const float*) ((uintptr_t) i2 + input_offset); i3 = (const float*) ((uintptr_t) i3 + input_offset); if (pooling_elements < 2) { i1 = i0; } if (pooling_elements <= 2) { i2 = i0; } if (pooling_elements != 4) { i3 = i0; } size_t c = channels; for (; c >= 4; c -= 4) { const __m128 vi0 = _mm_loadu_ps(i0); i0 += 4; const __m128 vi1 = _mm_loadu_ps(i1); i1 += 4; const __m128 vi2 = _mm_loadu_ps(i2); i2 += 4; const __m128 vi3 = _mm_loadu_ps(i3); i3 += 4; __m128 vmax = vi0; __m128i vidx = _mm_setzero_si128(); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1))); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2))); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3))); _mm_storeu_ps(output, vmax); output += 4; _mm_storeu_si128((__m128i*) index, vidx); index += 4; } if (c != 0) { const __m128 vi0 = _mm_loadu_ps(i0); const __m128 vi1 = _mm_loadu_ps(i1); const __m128 vi2 = _mm_loadu_ps(i2); const __m128 vi3 = _mm_loadu_ps(i3); __m128 vmax = vi0; __m128i vidx = _mm_setzero_si128(); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1))); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2))); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3))); if (c & 2) { _mm_storel_pi((__m64*) output, vmax); _mm_storel_epi64((__m128i*) index, vidx); vmax = _mm_movehl_ps(vmax, vmax); vidx = _mm_unpackhi_epi64(vidx, vidx); output += 2; index += 2; } if (c & 1) { _mm_store_ss(output, vmax); *index = (uint32_t) _mm_cvtsi128_si32(vidx); output += 1; index += 1; } } input = (const float**) ((uintptr_t) input + input_increment); output = (float*) ((uintptr_t) output + output_increment); } while (--output_pixels != 0); } void xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4( size_t output_pixels, size_t pooling_elements, size_t channels, const float** input, size_t input_offset, float* accumulation_buffer, uint32_t* index_buffer, float* output, uint32_t* index, size_t input_increment, size_t output_increment) XNN_OOB_READS { assert(output_pixels != 0); assert(pooling_elements != 0); assert(pooling_elements > 9); assert(channels != 0); do { { float* ab = accumulation_buffer; uint32_t* ib = index_buffer; const float* i0 = *input++; const float* i1 = *input++; const float* i2 = *input++; const float* i3 = *input++; const float* i4 = *input++; const float* i5 = *input++; const float* i6 = *input++; const float* i7 = *input++; const float* i8 = *input++; i0 = (const float*) ((uintptr_t) i0 + input_offset); i1 = (const float*) ((uintptr_t) i1 + input_offset); i2 = (const float*) ((uintptr_t) i2 + input_offset); i3 = (const float*) ((uintptr_t) i3 + input_offset); i4 = (const float*) ((uintptr_t) i4 + input_offset); i5 = (const float*) ((uintptr_t) i5 + input_offset); i6 = (const float*) ((uintptr_t) i6 + input_offset); i7 = (const float*) ((uintptr_t) i7 + input_offset); i8 = (const float*) ((uintptr_t) i8 + input_offset); for (size_t c = 0; c < channels; c += 4) { const __m128 vi0 = _mm_loadu_ps(i0); i0 += 4; const __m128 vi1 = _mm_loadu_ps(i1); i1 += 4; const __m128 vi2 = _mm_loadu_ps(i2); i2 += 4; const __m128 vi3 = _mm_loadu_ps(i3); i3 += 4; const __m128 vi4 = _mm_loadu_ps(i4); i4 += 4; const __m128 vi5 = _mm_loadu_ps(i5); i5 += 4; const __m128 vi6 = _mm_loadu_ps(i6); i6 += 4; const __m128 vi7 = _mm_loadu_ps(i7); i7 += 4; const __m128 vi8 = _mm_loadu_ps(i8); i8 += 4; __m128 vmax = vi0; __m128i vidx = _mm_setzero_si128(); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1))); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2))); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3))); const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); vmax = _mm_max_ps(vi4, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4))); const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); vmax = _mm_max_ps(vi5, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5))); const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); vmax = _mm_max_ps(vi6, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6))); const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); vmax = _mm_max_ps(vi7, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7))); const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax)); vmax = _mm_max_ps(vi8, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8))); _mm_store_ps(ab, vmax); ab += 4; _mm_store_si128((__m128i*) ib, vidx); ib += 4; } } const __m128i v1 = _mm_set1_epi32(1); const __m128i v8 = _mm_set1_epi32(8); __m128i vidx0 = _mm_add_epi32(v1, v8); size_t k = pooling_elements; for (k -= 9; k > 8; k -= 8) { const float* i0 = *input++; const float* i1 = *input++; const float* i2 = *input++; const float* i3 = *input++; const float* i4 = *input++; const float* i5 = *input++; const float* i6 = *input++; const float* i7 = *input++; i0 = (const float*) ((uintptr_t) i0 + input_offset); i1 = (const float*) ((uintptr_t) i1 + input_offset); i2 = (const float*) ((uintptr_t) i2 + input_offset); i3 = (const float*) ((uintptr_t) i3 + input_offset); i4 = (const float*) ((uintptr_t) i4 + input_offset); i5 = (const float*) ((uintptr_t) i5 + input_offset); i6 = (const float*) ((uintptr_t) i6 + input_offset); i7 = (const float*) ((uintptr_t) i7 + input_offset); float* ab = accumulation_buffer; uint32_t* ib = index_buffer; for (size_t c = 0; c < channels; c += 4) { const __m128 vi0 = _mm_loadu_ps(i0); i0 += 4; const __m128 vi1 = _mm_loadu_ps(i1); i1 += 4; const __m128 vi2 = _mm_loadu_ps(i2); i2 += 4; const __m128 vi3 = _mm_loadu_ps(i3); i3 += 4; const __m128 vi4 = _mm_loadu_ps(i4); i4 += 4; const __m128 vi5 = _mm_loadu_ps(i5); i5 += 4; const __m128 vi6 = _mm_loadu_ps(i6); i6 += 4; const __m128 vi7 = _mm_loadu_ps(i7); i7 += 4; __m128 vmax = _mm_load_ps(ab); __m128i vidx = _mm_load_si128((const __m128i*) ib); const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax)); vmax = _mm_max_ps(vi0, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0)); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); const __m128i vidx1 = _mm_add_epi32(vidx0, v1); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1)); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); const __m128i vidx2 = _mm_add_epi32(vidx1, v1); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2)); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); const __m128i vidx3 = _mm_add_epi32(vidx2, v1); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3)); const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); const __m128i vidx4 = _mm_add_epi32(vidx3, v1); vmax = _mm_max_ps(vi4, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4)); const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); const __m128i vidx5 = _mm_add_epi32(vidx4, v1); vmax = _mm_max_ps(vi5, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5)); const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); const __m128i vidx6 = _mm_add_epi32(vidx5, v1); vmax = _mm_max_ps(vi6, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6)); const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); const __m128i vidx7 = _mm_add_epi32(vidx6, v1); vmax = _mm_max_ps(vi7, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7)); _mm_store_ps(ab, vmax); ab += 4; _mm_store_si128((__m128i*) ib, vidx); ib += 4; } vidx0 = _mm_add_epi32(vidx0, v8); } float* o = output; uint32_t* i = index; { const float* i0 = input[0]; const float* i1 = input[1]; const float* i2 = input[2]; const float* i3 = input[3]; const float* i4 = input[4]; const float* i5 = input[5]; const float* i6 = input[6]; const float* i7 = input[7]; i0 = (const float*) ((uintptr_t) i0 + input_offset); i1 = (const float*) ((uintptr_t) i1 + input_offset); i2 = (const float*) ((uintptr_t) i2 + input_offset); i3 = (const float*) ((uintptr_t) i3 + input_offset); i4 = (const float*) ((uintptr_t) i4 + input_offset); i5 = (const float*) ((uintptr_t) i5 + input_offset); i6 = (const float*) ((uintptr_t) i6 + input_offset); i7 = (const float*) ((uintptr_t) i7 + input_offset); input = (const float**) ((uintptr_t) input + input_increment); if (k < 2) { i1 = i0; } if (k <= 2) { i2 = i0; } if (k < 4) { i3 = i0; } if (k <= 4) { i4 = i0; } if (k < 6) { i5 = i0; } if (k <= 6) { i6 = i0; } if (k != 8) { i7 = i0; } size_t c = channels; float* ab = accumulation_buffer; uint32_t* ib = index_buffer; for (; c >= 4; c -= 4) { const __m128 vi0 = _mm_loadu_ps(i0); i0 += 4; const __m128 vi1 = _mm_loadu_ps(i1); i1 += 4; const __m128 vi2 = _mm_loadu_ps(i2); i2 += 4; const __m128 vi3 = _mm_loadu_ps(i3); i3 += 4; const __m128 vi4 = _mm_loadu_ps(i4); i4 += 4; const __m128 vi5 = _mm_loadu_ps(i5); i5 += 4; const __m128 vi6 = _mm_loadu_ps(i6); i6 += 4; const __m128 vi7 = _mm_loadu_ps(i7); i7 += 4; __m128 vmax = _mm_load_ps(ab); ab += 4; __m128i vidx = _mm_load_si128((const __m128i*) ib); ib += 4; const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax)); vmax = _mm_max_ps(vi0, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0)); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); const __m128i vidx1 = _mm_add_epi32(vidx0, v1); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1)); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); const __m128i vidx2 = _mm_add_epi32(vidx1, v1); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2)); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); const __m128i vidx3 = _mm_add_epi32(vidx2, v1); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3)); const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); const __m128i vidx4 = _mm_add_epi32(vidx3, v1); vmax = _mm_max_ps(vi4, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4)); const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); const __m128i vidx5 = _mm_add_epi32(vidx4, v1); vmax = _mm_max_ps(vi5, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5)); const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); const __m128i vidx6 = _mm_add_epi32(vidx5, v1); vmax = _mm_max_ps(vi6, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6)); const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); const __m128i vidx7 = _mm_add_epi32(vidx6, v1); vmax = _mm_max_ps(vi7, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7)); _mm_storeu_ps(o, vmax); o += 4; _mm_storeu_si128((__m128i*) i, vidx); i += 4; } if (c != 0) { const __m128 vi0 = _mm_loadu_ps(i0); const __m128 vi1 = _mm_loadu_ps(i1); const __m128 vi2 = _mm_loadu_ps(i2); const __m128 vi3 = _mm_loadu_ps(i3); const __m128 vi4 = _mm_loadu_ps(i4); const __m128 vi5 = _mm_loadu_ps(i5); const __m128 vi6 = _mm_loadu_ps(i6); const __m128 vi7 = _mm_loadu_ps(i7); __m128 vmax = _mm_load_ps(ab); __m128i vidx = _mm_load_si128((const __m128i*) ib); const __m128i vm0 = _mm_castps_si128(_mm_cmpgt_ps(vi0, vmax)); vmax = _mm_max_ps(vi0, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm0, vidx), _mm_and_si128(vm0, vidx0)); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); const __m128i vidx1 = _mm_add_epi32(vidx0, v1); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, vidx1)); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); const __m128i vidx2 = _mm_add_epi32(vidx1, v1); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, vidx2)); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); const __m128i vidx3 = _mm_add_epi32(vidx2, v1); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, vidx3)); const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); const __m128i vidx4 = _mm_add_epi32(vidx3, v1); vmax = _mm_max_ps(vi4, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, vidx4)); const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); const __m128i vidx5 = _mm_add_epi32(vidx4, v1); vmax = _mm_max_ps(vi5, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, vidx5)); const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); const __m128i vidx6 = _mm_add_epi32(vidx5, v1); vmax = _mm_max_ps(vi6, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, vidx6)); const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); const __m128i vidx7 = _mm_add_epi32(vidx6, v1); vmax = _mm_max_ps(vi7, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, vidx7)); if (c & 2) { _mm_storel_pi((__m64*) o, vmax); _mm_storel_epi64((__m128i*) i, vidx); vmax = _mm_movehl_ps(vmax, vmax); vidx = _mm_unpackhi_epi64(vidx, vidx); o += 2; i += 2; } if (c & 1) { _mm_store_ss(o, vmax); *i = (uint32_t) _mm_cvtsi128_si32(vidx); o += 1; i += 1; } } } output = (float*) ((uintptr_t) o + output_increment); index = (uint32_t*) i; } while (--output_pixels != 0); } void xnn_f32_argmaxpool_ukernel_9x__sse2_c4( size_t output_pixels, size_t pooling_elements, size_t channels, const float** input, size_t input_offset, float* output, uint32_t* index, size_t input_increment, size_t output_increment) XNN_OOB_READS { assert(output_pixels != 0); assert(pooling_elements != 0); assert(pooling_elements <= 9); assert(channels != 0); do { const float* i0 = input[0]; const float* i1 = input[1]; const float* i2 = input[2]; const float* i3 = input[3]; const float* i4 = input[4]; const float* i5 = input[5]; const float* i6 = input[6]; const float* i7 = input[7]; const float* i8 = input[8]; i0 = (const float*) ((uintptr_t) i0 + input_offset); i1 = (const float*) ((uintptr_t) i1 + input_offset); i2 = (const float*) ((uintptr_t) i2 + input_offset); i3 = (const float*) ((uintptr_t) i3 + input_offset); i4 = (const float*) ((uintptr_t) i4 + input_offset); i5 = (const float*) ((uintptr_t) i5 + input_offset); i6 = (const float*) ((uintptr_t) i6 + input_offset); i7 = (const float*) ((uintptr_t) i7 + input_offset); i8 = (const float*) ((uintptr_t) i8 + input_offset); if (pooling_elements < 2) { i1 = i0; } if (pooling_elements <= 2) { i2 = i0; } if (pooling_elements < 4) { i3 = i0; } if (pooling_elements <= 4) { i4 = i0; } if (pooling_elements < 6) { i5 = i0; } if (pooling_elements <= 6) { i6 = i0; } if (pooling_elements < 8) { i7 = i0; } if (pooling_elements <= 8) { i8 = i0; } size_t c = channels; for (; c >= 4; c -= 4) { const __m128 vi0 = _mm_loadu_ps(i0); i0 += 4; const __m128 vi1 = _mm_loadu_ps(i1); i1 += 4; const __m128 vi2 = _mm_loadu_ps(i2); i2 += 4; const __m128 vi3 = _mm_loadu_ps(i3); i3 += 4; const __m128 vi4 = _mm_loadu_ps(i4); i4 += 4; const __m128 vi5 = _mm_loadu_ps(i5); i5 += 4; const __m128 vi6 = _mm_loadu_ps(i6); i6 += 4; const __m128 vi7 = _mm_loadu_ps(i7); i7 += 4; const __m128 vi8 = _mm_loadu_ps(i8); i8 += 4; __m128 vmax = vi0; __m128i vidx = _mm_setzero_si128(); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1))); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2))); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3))); const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); vmax = _mm_max_ps(vi4, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4))); const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); vmax = _mm_max_ps(vi5, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5))); const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); vmax = _mm_max_ps(vi6, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6))); const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); vmax = _mm_max_ps(vi7, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7))); const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax)); vmax = _mm_max_ps(vi8, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8))); _mm_storeu_ps(output, vmax); output += 4; _mm_storeu_si128((__m128i*) index, vidx); index += 4; } if (c != 0) { const __m128 vi0 = _mm_loadu_ps(i0); const __m128 vi1 = _mm_loadu_ps(i1); const __m128 vi2 = _mm_loadu_ps(i2); const __m128 vi3 = _mm_loadu_ps(i3); const __m128 vi4 = _mm_loadu_ps(i4); const __m128 vi5 = _mm_loadu_ps(i5); const __m128 vi6 = _mm_loadu_ps(i6); const __m128 vi7 = _mm_loadu_ps(i7); const __m128 vi8 = _mm_loadu_ps(i8); __m128 vmax = vi0; __m128i vidx = _mm_setzero_si128(); const __m128i vm1 = _mm_castps_si128(_mm_cmpgt_ps(vi1, vmax)); vmax = _mm_max_ps(vi1, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm1, vidx), _mm_and_si128(vm1, _mm_set1_epi32(1))); const __m128i vm2 = _mm_castps_si128(_mm_cmpgt_ps(vi2, vmax)); vmax = _mm_max_ps(vi2, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm2, vidx), _mm_and_si128(vm2, _mm_set1_epi32(2))); const __m128i vm3 = _mm_castps_si128(_mm_cmpgt_ps(vi3, vmax)); vmax = _mm_max_ps(vi3, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm3, vidx), _mm_and_si128(vm3, _mm_set1_epi32(3))); const __m128i vm4 = _mm_castps_si128(_mm_cmpgt_ps(vi4, vmax)); vmax = _mm_max_ps(vi4, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm4, vidx), _mm_and_si128(vm4, _mm_set1_epi32(4))); const __m128i vm5 = _mm_castps_si128(_mm_cmpgt_ps(vi5, vmax)); vmax = _mm_max_ps(vi5, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm5, vidx), _mm_and_si128(vm5, _mm_set1_epi32(5))); const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); vmax = _mm_max_ps(vi6, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm6, vidx), _mm_and_si128(vm6, _mm_set1_epi32(6))); const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); vmax = _mm_max_ps(vi7, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm7, vidx), _mm_and_si128(vm7, _mm_set1_epi32(7))); const __m128i vm8 = _mm_castps_si128(_mm_cmpgt_ps(vi8, vmax)); vmax = _mm_max_ps(vi8, vmax); vidx = _mm_or_si128(_mm_andnot_si128(vm8, vidx), _mm_and_si128(vm8, _mm_set1_epi32(8))); if (c & 2) { _mm_storel_pi((__m64*) output, vmax); _mm_storel_epi64((__m128i*) index, vidx); vmax = _mm_movehl_ps(vmax, vmax); vidx = _mm_unpackhi_epi64(vidx, vidx); output += 2; index += 2; } if (c & 1) { _mm_store_ss(output, vmax); *index = (uint32_t) _mm_cvtsi128_si32(vidx); output += 1; index += 1; } } input = (const float**) ((uintptr_t) input + input_increment); output = (float*) ((uintptr_t) output + output_increment); index = (uint32_t*) index; } while (--output_pixels != 0); } void xnn_f32_f16_vcvt_ukernel__sse2_x16( size_t n, const float* input, void* output, const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); assert(input != NULL); assert(output != NULL); const __m128 vnonsign_mask = _mm_load_ps((const float*) params->sse2.nonsign_mask); const __m128i vexp_bias = _mm_load_si128((const __m128i*) params->sse2.exp_bias); const __m128 vscale_to_inf = _mm_load_ps(params->sse2.scale_to_inf); const __m128i vexpw_max = _mm_load_si128((const __m128i*) params->sse2.expw_max); const __m128 vscale_to_zero = _mm_load_ps(params->sse2.scale_to_zero); const __m128i vbias_min = _mm_load_si128((const __m128i*) params->sse2.bias_min); const __m128i vmanth_mask = _mm_load_si128((const __m128i*) params->sse2.manth_mask); const __m128i vexph_mask = _mm_load_si128((const __m128i*) params->sse2.exph_mask); const __m128i vnanh = _mm_load_si128((const __m128i*) params->sse2.nanh); uint16_t* o = (uint16_t*) output; for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { const __m128 vx0 = _mm_loadu_ps(input); const __m128 vx1 = _mm_loadu_ps(input + 4); const __m128 vx2 = _mm_loadu_ps(input + 8); const __m128 vx3 = _mm_loadu_ps(input + 12); input += 16; const __m128 vabsx0 = _mm_and_ps(vx0, vnonsign_mask); const __m128 vabsx1 = _mm_and_ps(vx1, vnonsign_mask); const __m128 vabsx2 = _mm_and_ps(vx2, vnonsign_mask); const __m128 vabsx3 = _mm_and_ps(vx3, vnonsign_mask); const __m128 vsignx0 = _mm_xor_ps(vx0, vabsx0); const __m128 vsignx1 = _mm_xor_ps(vx1, vabsx1); const __m128 vsignx2 = _mm_xor_ps(vx2, vabsx2); const __m128 vsignx3 = _mm_xor_ps(vx3, vabsx3); __m128i vbias0 = _mm_add_epi32(_mm_castps_si128(vabsx0), vexp_bias); __m128i vbias1 = _mm_add_epi32(_mm_castps_si128(vabsx1), vexp_bias); __m128i vbias2 = _mm_add_epi32(_mm_castps_si128(vabsx2), vexp_bias); __m128i vbias3 = _mm_add_epi32(_mm_castps_si128(vabsx3), vexp_bias); __m128 vf0 = _mm_mul_ps(vabsx0, vscale_to_inf); __m128 vf1 = _mm_mul_ps(vabsx1, vscale_to_inf); __m128 vf2 = _mm_mul_ps(vabsx2, vscale_to_inf); __m128 vf3 = _mm_mul_ps(vabsx3, vscale_to_inf); const __m128i vnanmaskw0 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx0), vexpw_max); const __m128i vnanmaskw1 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx1), vexpw_max); const __m128i vnanmaskw2 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx2), vexpw_max); const __m128i vnanmaskw3 = _mm_cmpgt_epi32(_mm_castps_si128(vabsx3), vexpw_max); vbias0 = _mm_and_si128(vbias0, vexpw_max); vbias1 = _mm_and_si128(vbias1, vexpw_max); vbias2 = _mm_and_si128(vbias2, vexpw_max); vbias3 = _mm_and_si128(vbias3, vexpw_max); vf0 = _mm_mul_ps(vf0, vscale_to_zero); vf1 = _mm_mul_ps(vf1, vscale_to_zero); vf2 = _mm_mul_ps(vf2, vscale_to_zero); vf3 = _mm_mul_ps(vf3, vscale_to_zero); const __m128i vnanmaskh0 = _mm_packs_epi32(vnanmaskw0, vnanmaskw1); const __m128i vnanmaskh1 = _mm_packs_epi32(vnanmaskw2, vnanmaskw3); const __m128i vsignh0 = _mm_packs_epi32(_mm_castps_si128(vsignx0), _mm_castps_si128(vsignx1)); const __m128i vsignh1 = _mm_packs_epi32(_mm_castps_si128(vsignx2), _mm_castps_si128(vsignx3)); vbias0 = _mm_max_epi16(vbias0, vbias_min); vbias1 = _mm_max_epi16(vbias1, vbias_min); vbias2 = _mm_max_epi16(vbias2, vbias_min); vbias3 = _mm_max_epi16(vbias3, vbias_min); __m128i vh0 = _mm_and_si128(vnanh, vnanmaskh0); __m128i vh1 = _mm_and_si128(vnanh, vnanmaskh1); vf0 = _mm_add_ps(vf0, _mm_castsi128_ps(vbias0)); vf1 = _mm_add_ps(vf1, _mm_castsi128_ps(vbias1)); vf2 = _mm_add_ps(vf2, _mm_castsi128_ps(vbias2)); vf3 = _mm_add_ps(vf3, _mm_castsi128_ps(vbias3)); vh0 = _mm_or_si128(vh0, vsignh0); vh1 = _mm_or_si128(vh1, vsignh1); __m128i vexpw0 = _mm_srli_epi32(_mm_castps_si128(vf0), 13); __m128i vexpw1 = _mm_srli_epi32(_mm_castps_si128(vf1), 13); __m128i vexpw2 = _mm_srli_epi32(_mm_castps_si128(vf2), 13); __m128i vexpw3 = _mm_srli_epi32(_mm_castps_si128(vf3), 13); const __m128i vmantw0 = _mm_and_si128(_mm_castps_si128(vf0), vmanth_mask); const __m128i vmantw1 = _mm_and_si128(_mm_castps_si128(vf1), vmanth_mask); const __m128i vmantw2 = _mm_and_si128(_mm_castps_si128(vf2), vmanth_mask); const __m128i vmantw3 = _mm_and_si128(_mm_castps_si128(vf3), vmanth_mask); vexpw0 = _mm_and_si128(vexpw0, vexph_mask); vexpw1 = _mm_and_si128(vexpw1, vexph_mask); vexpw2 = _mm_and_si128(vexpw2, vexph_mask); vexpw3 = _mm_and_si128(vexpw3, vexph_mask); const __m128i vnonsignw0 = _mm_add_epi32(vmantw0, vexpw0); const __m128i vnonsignw1 = _mm_add_epi32(vmantw1, vexpw1); const __m128i vnonsignw2 = _mm_add_epi32(vmantw2, vexpw2); const __m128i vnonsignw3 = _mm_add_epi32(vmantw3, vexpw3); const __m128i vnonsignh0 = _mm_packs_epi32(vnonsignw0, vnonsignw1); const __m128i vnonsignh1 = _mm_packs_epi32(vnonsignw2, vnonsignw3); vh0 = _mm_or_si128(vh0, _mm_andnot_si128(vnanmaskh0, vnonsignh0)); vh1 = _mm_or_si128(vh1, _mm_andnot_si128(vnanmaskh1, vnonsignh1)); _mm_storeu_si128((__m128i*) o, vh0); _mm_storeu_si128((__m128i*) (o + 8), vh1); o += 16; } for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx_lo = _mm_loadu_ps(input); const __m128 vx_hi = _mm_loadu_ps(input + 4); input += 8; const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask); const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask); const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo); const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi); __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias); __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias); __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf); __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf); const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max); const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max); vbias_lo = _mm_and_si128(vbias_lo, vexpw_max); vbias_hi = _mm_and_si128(vbias_hi, vexpw_max); vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero); vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero); const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi); const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi)); vbias_lo = _mm_max_epi16(vbias_lo, vbias_min); vbias_hi = _mm_max_epi16(vbias_hi, vbias_min); __m128i vh = _mm_and_si128(vnanh, vnanmaskh); vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo)); vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi)); vh = _mm_or_si128(vh, vsignh); __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13); __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13); const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask); const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask); vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask); vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask); const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo); const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi); const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi); vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh)); _mm_storeu_si128((__m128i*) o, vh); o += 8; } if XNN_UNPREDICTABLE(n != 0) { const __m128 vx_lo = _mm_loadu_ps(input); const float* input_hi = (const float*) ((uintptr_t) input + (n & (4 * sizeof(float)))); const __m128 vx_hi = _mm_loadu_ps(input_hi); const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask); const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask); const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo); const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi); __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias); __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias); __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf); __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf); const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max); const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max); vbias_lo = _mm_and_si128(vbias_lo, vexpw_max); vbias_hi = _mm_and_si128(vbias_hi, vexpw_max); vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero); vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero); const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi); const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi)); vbias_lo = _mm_max_epi16(vbias_lo, vbias_min); vbias_hi = _mm_max_epi16(vbias_hi, vbias_min); __m128i vh = _mm_and_si128(vnanh, vnanmaskh); vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo)); vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi)); vh = _mm_or_si128(vh, vsignh); __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13); __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13); const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask); const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask); vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask); vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask); const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo); const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi); const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi); vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh)); if (n & (4 * sizeof(float))) { _mm_storel_epi64((__m128i*) o, vh); vh = _mm_unpackhi_epi64(vh, vh); o += 4; } if (n & (2 * sizeof(float))) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vh)); vh = _mm_srli_epi64(vh, 32); o += 2; } if (n & (1 * sizeof(float))) { *o = (uint16_t) _mm_cvtsi128_si32(vh); } } } void xnn_f32_prelu_ukernel__sse2_2x8( size_t rows, size_t channels, const float*restrict input, size_t input_stride, const float*restrict weights, float*restrict output, size_t output_stride) XNN_OOB_READS { assert(rows != 0); assert(channels != 0); assert(channels % sizeof(float) == 0); const float* i0 = input; float* o0 = output; const float* i1 = (const float*) ((uintptr_t) i0 + input_stride); float* o1 = (float*) ((uintptr_t) o0 + output_stride); const size_t input_increment = input_stride * 2 - channels; const size_t output_increment = output_stride * 2 - channels; do { if XNN_UNPREDICTABLE(rows < 2) { i1 = i0; o1 = o0; } const float* w = weights; size_t c = channels; for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) { const __m128 vw0123 = _mm_load_ps(w); const __m128 vw4567 = _mm_load_ps(w + 4); w += 8; const __m128 vi0x0123 = _mm_loadu_ps(i0); const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4); i0 += 8; const __m128 vi1x0123 = _mm_loadu_ps(i1); const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4); i1 += 8; const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); const __m128 vprod0x4567 = _mm_mul_ps(vi0x4567, vw4567); const __m128 vmask0x4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x4567))); const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123); const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123))); const __m128 vprod1x4567 = _mm_mul_ps(vi1x4567, vw4567); const __m128 vmask1x4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x4567))); const __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); const __m128 vacc0x4567 = _mm_or_ps(_mm_and_ps(vprod0x4567, vmask0x4567), _mm_andnot_ps(vmask0x4567, vi0x4567)); const __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123)); const __m128 vacc1x4567 = _mm_or_ps(_mm_and_ps(vprod1x4567, vmask1x4567), _mm_andnot_ps(vmask1x4567, vi1x4567)); _mm_storeu_ps(o0, vacc0x0123); _mm_storeu_ps(o0 + 4, vacc0x4567); o0 += 8; _mm_storeu_ps(o1, vacc1x0123); _mm_storeu_ps(o1 + 4, vacc1x4567); o1 += 8; } for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) { const __m128 vw0123 = _mm_load_ps(w); w += 4; const __m128 vi0x0123 = _mm_loadu_ps(i0); i0 += 4; const __m128 vi1x0123 = _mm_loadu_ps(i1); i1 += 4; const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123); const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123))); __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123)); _mm_storeu_ps(o0, vacc0x0123); o0 += 4; _mm_storeu_ps(o1, vacc1x0123); o1 += 4; } if XNN_UNLIKELY(c != 0) { const __m128 vw0123 = _mm_load_ps(w); w = (const float*) ((uintptr_t) w + c); const __m128 vi0x0123 = _mm_loadu_ps(i0); i0 = (const float*) ((uintptr_t) i0 + c); const __m128 vi1x0123 = _mm_loadu_ps(i1); i1 = (const float*) ((uintptr_t) i1 + c); const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123); const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123))); __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123)); if (c & (2 * sizeof(float))) { _mm_storel_pi((__m64*) o0, vacc0x0123); _mm_storel_pi((__m64*) o1, vacc1x0123); vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123); vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123); o0 += 2; o1 += 2; } if (c & (1 * sizeof(float))) { _mm_store_ss(o0, vacc0x0123); _mm_store_ss(o1, vacc1x0123); o0 += 1; o1 += 1; } } i0 = (const float*) ((uintptr_t) i0 + input_increment); o0 = (float*) ((uintptr_t) o0 + output_increment); i1 = (const float*) ((uintptr_t) i1 + input_increment); o1 = (float*) ((uintptr_t) o1 + output_increment); rows = doz(rows, 2); } while (rows != 0); } void xnn_f32_qs8_vcvt_ukernel__sse2_x32( size_t n, const float* x, int8_t* y, const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); assert(x != NULL); assert(y != NULL); const __m128 vscale = _mm_load_ps(params->sse2.scale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) { __m128 vx0123 = _mm_loadu_ps(x); __m128 vx4567 = _mm_loadu_ps(x + 4); __m128 vx89AB = _mm_loadu_ps(x + 8); __m128 vxCDEF = _mm_loadu_ps(x + 12); __m128 vxGHIJ = _mm_loadu_ps(x + 16); __m128 vxKLMN = _mm_loadu_ps(x + 20); __m128 vxOPQR = _mm_loadu_ps(x + 24); __m128 vxSTUV = _mm_loadu_ps(x + 28); x += 32; vx0123 = _mm_mul_ps(vx0123, vscale); vx4567 = _mm_mul_ps(vx4567, vscale); vx89AB = _mm_mul_ps(vx89AB, vscale); vxCDEF = _mm_mul_ps(vxCDEF, vscale); vxGHIJ = _mm_mul_ps(vxGHIJ, vscale); vxKLMN = _mm_mul_ps(vxKLMN, vscale); vxOPQR = _mm_mul_ps(vxOPQR, vscale); vxSTUV = _mm_mul_ps(vxSTUV, vscale); vx0123 = _mm_min_ps(vx0123, voutput_max_less_zero_point); vx4567 = _mm_min_ps(vx4567, voutput_max_less_zero_point); vx89AB = _mm_min_ps(vx89AB, voutput_max_less_zero_point); vxCDEF = _mm_min_ps(vxCDEF, voutput_max_less_zero_point); vxGHIJ = _mm_min_ps(vxGHIJ, voutput_max_less_zero_point); vxKLMN = _mm_min_ps(vxKLMN, voutput_max_less_zero_point); vxOPQR = _mm_min_ps(vxOPQR, voutput_max_less_zero_point); vxSTUV = _mm_min_ps(vxSTUV, voutput_max_less_zero_point); const __m128i vy0123 = _mm_cvtps_epi32(vx0123); const __m128i vy4567 = _mm_cvtps_epi32(vx4567); const __m128i vy89AB = _mm_cvtps_epi32(vx89AB); const __m128i vyCDEF = _mm_cvtps_epi32(vxCDEF); const __m128i vyGHIJ = _mm_cvtps_epi32(vxGHIJ); const __m128i vyKLMN = _mm_cvtps_epi32(vxKLMN); const __m128i vyOPQR = _mm_cvtps_epi32(vxOPQR); const __m128i vySTUV = _mm_cvtps_epi32(vxSTUV); __m128i vy01234567 = _mm_packs_epi32(vy0123, vy4567); __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF); __m128i vyGHIJKLMN = _mm_packs_epi32(vyGHIJ, vyKLMN); __m128i vyOPQRSTUV = _mm_packs_epi32(vyOPQR, vySTUV); vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point); vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point); vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point); vyOPQRSTUV = _mm_adds_epi16(vyOPQRSTUV, voutput_zero_point); vy01234567 = _mm_max_epi16(vy01234567, voutput_min); vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min); vyGHIJKLMN = _mm_max_epi16(vyGHIJKLMN, voutput_min); vyOPQRSTUV = _mm_max_epi16(vyOPQRSTUV, voutput_min); __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF); __m128i vyGHIJKLMNOPQRSTUV = _mm_packs_epi16(vyGHIJKLMN, vyOPQRSTUV); _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF); _mm_storeu_si128((__m128i*) (y + 16), vyGHIJKLMNOPQRSTUV); y += 32; } for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { __m128 vx_lo = _mm_loadu_ps(x); __m128 vx_hi = _mm_loadu_ps(x + 4); x += 8; vx_lo = _mm_mul_ps(vx_lo, vscale); vx_hi = _mm_mul_ps(vx_hi, vscale); vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point); vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point); const __m128i vy_lo = _mm_cvtps_epi32(vx_lo); const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); vy = _mm_adds_epi16(vy, voutput_zero_point); vy = _mm_max_epi16(vy, voutput_min); vy = _mm_packs_epi16(vy, vy); _mm_storel_epi64((__m128i*) y, vy); y += 8; } if XNN_UNLIKELY(n != 0) { __m128 vx_lo = _mm_loadu_ps(x); const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float)))); __m128 vx_hi = _mm_loadu_ps(x_hi); vx_lo = _mm_mul_ps(vx_lo, vscale); vx_hi = _mm_mul_ps(vx_hi, vscale); vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point); vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point); const __m128i vy_lo = _mm_cvtps_epi32(vx_lo); const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); vy = _mm_adds_epi16(vy, voutput_zero_point); vy = _mm_max_epi16(vy, voutput_min); vy = _mm_packs_epi16(vy, vy); if (n & (4 * sizeof(float))) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy)); y += 4; vy = _mm_srli_epi64(vy, 32); } { uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy); if (n & (2 * sizeof(float))) { unaligned_store_u16(y, (uint16_t) vy_lo); y += 2; vy_lo >>= 16; } if (n & (1 * sizeof(float))) { *y = (int8_t) vy_lo; } } } } void xnn_f32_qu8_vcvt_ukernel__sse2_x32( size_t n, const float* x, uint8_t* y, const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); assert(x != NULL); assert(y != NULL); const __m128 vscale = _mm_load_ps(params->sse2.scale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) { __m128 vx0123 = _mm_loadu_ps(x); __m128 vx4567 = _mm_loadu_ps(x + 4); __m128 vx89AB = _mm_loadu_ps(x + 8); __m128 vxCDEF = _mm_loadu_ps(x + 12); __m128 vxGHIJ = _mm_loadu_ps(x + 16); __m128 vxKLMN = _mm_loadu_ps(x + 20); __m128 vxOPQR = _mm_loadu_ps(x + 24); __m128 vxSTUV = _mm_loadu_ps(x + 28); x += 32; vx0123 = _mm_mul_ps(vx0123, vscale); vx4567 = _mm_mul_ps(vx4567, vscale); vx89AB = _mm_mul_ps(vx89AB, vscale); vxCDEF = _mm_mul_ps(vxCDEF, vscale); vxGHIJ = _mm_mul_ps(vxGHIJ, vscale); vxKLMN = _mm_mul_ps(vxKLMN, vscale); vxOPQR = _mm_mul_ps(vxOPQR, vscale); vxSTUV = _mm_mul_ps(vxSTUV, vscale); vx0123 = _mm_min_ps(vx0123, voutput_max_less_zero_point); vx4567 = _mm_min_ps(vx4567, voutput_max_less_zero_point); vx89AB = _mm_min_ps(vx89AB, voutput_max_less_zero_point); vxCDEF = _mm_min_ps(vxCDEF, voutput_max_less_zero_point); vxGHIJ = _mm_min_ps(vxGHIJ, voutput_max_less_zero_point); vxKLMN = _mm_min_ps(vxKLMN, voutput_max_less_zero_point); vxOPQR = _mm_min_ps(vxOPQR, voutput_max_less_zero_point); vxSTUV = _mm_min_ps(vxSTUV, voutput_max_less_zero_point); const __m128i vy0123 = _mm_cvtps_epi32(vx0123); const __m128i vy4567 = _mm_cvtps_epi32(vx4567); const __m128i vy89AB = _mm_cvtps_epi32(vx89AB); const __m128i vyCDEF = _mm_cvtps_epi32(vxCDEF); const __m128i vyGHIJ = _mm_cvtps_epi32(vxGHIJ); const __m128i vyKLMN = _mm_cvtps_epi32(vxKLMN); const __m128i vyOPQR = _mm_cvtps_epi32(vxOPQR); const __m128i vySTUV = _mm_cvtps_epi32(vxSTUV); __m128i vy01234567 = _mm_packs_epi32(vy0123, vy4567); __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF); __m128i vyGHIJKLMN = _mm_packs_epi32(vyGHIJ, vyKLMN); __m128i vyOPQRSTUV = _mm_packs_epi32(vyOPQR, vySTUV); vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point); vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point); vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point); vyOPQRSTUV = _mm_adds_epi16(vyOPQRSTUV, voutput_zero_point); __m128i vy0123456789ABCDEF = _mm_packus_epi16(vy01234567, vy89ABCDEF); __m128i vyGHIJKLMNOPQRSTUV = _mm_packus_epi16(vyGHIJKLMN, vyOPQRSTUV); vy0123456789ABCDEF = _mm_max_epu8(vy0123456789ABCDEF, voutput_min); vyGHIJKLMNOPQRSTUV = _mm_max_epu8(vyGHIJKLMNOPQRSTUV, voutput_min); _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF); _mm_storeu_si128((__m128i*) (y + 16), vyGHIJKLMNOPQRSTUV); y += 32; } for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { __m128 vx_lo = _mm_loadu_ps(x); __m128 vx_hi = _mm_loadu_ps(x + 4); x += 8; vx_lo = _mm_mul_ps(vx_lo, vscale); vx_hi = _mm_mul_ps(vx_hi, vscale); vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point); vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point); const __m128i vy_lo = _mm_cvtps_epi32(vx_lo); const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); vy = _mm_adds_epi16(vy, voutput_zero_point); vy = _mm_packus_epi16(vy, vy); vy = _mm_max_epu8(vy, voutput_min); _mm_storel_epi64((__m128i*) y, vy); y += 8; } if XNN_UNLIKELY(n != 0) { __m128 vx_lo = _mm_loadu_ps(x); const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float)))); __m128 vx_hi = _mm_loadu_ps(x_hi); vx_lo = _mm_mul_ps(vx_lo, vscale); vx_hi = _mm_mul_ps(vx_hi, vscale); vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point); vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point); const __m128i vy_lo = _mm_cvtps_epi32(vx_lo); const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); vy = _mm_adds_epi16(vy, voutput_zero_point); vy = _mm_packus_epi16(vy, vy); vy = _mm_max_epu8(vy, voutput_min); if (n & (4 * sizeof(float))) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy)); y += 4; vy = _mm_srli_epi64(vy, 32); } { uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy); if (n & (2 * sizeof(float))) { unaligned_store_u16(y, (uint16_t) vy_lo); y += 2; vy_lo >>= 16; } if (n & (1 * sizeof(float))) { *y = (uint8_t) vy_lo; } } } } void xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2( size_t elements, const float* input, const float* max, float* output, float* sum, const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(elements % sizeof(float) == 0); const __m128 vi_max = _mm_load1_ps(max); const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_p5.log2e); const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_p5.magic_bias); const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_p5.minus_ln2_hi); const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_p5.minus_ln2_lo); const __m128 vc5 = _mm_load_ps(params->sse2_rr2_p5.c5); const __m128 vc4 = _mm_load_ps(params->sse2_rr2_p5.c4); const __m128 vc3 = _mm_load_ps(params->sse2_rr2_p5.c3); const __m128 vc2 = _mm_load_ps(params->sse2_rr2_p5.c2); const __m128 vc1 = _mm_load_ps(params->sse2_rr2_p5.c1); const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_p5.denorm_cutoff); __m128 vacc0 = _mm_setzero_ps(); __m128 vacc1 = _mm_setzero_ps(); for (; elements >= 20 * sizeof(float); elements -= 20 * sizeof(float)) { // Load 20 (5x4) inputs at a time. const __m128 vi0123 = _mm_loadu_ps(input); const __m128 vi4567 = _mm_loadu_ps(input + 4); const __m128 vi89AB = _mm_loadu_ps(input + 8); const __m128 viCDEF = _mm_loadu_ps(input + 12); const __m128 viGHIJ = _mm_loadu_ps(input + 16); input += 20; // Subtract maximum input x := i - i_max. This implies x <= 0. const __m128 vx0123 = _mm_sub_ps(vi0123, vi_max); const __m128 vx4567 = _mm_sub_ps(vi4567, vi_max); const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); const __m128 vxCDEF = _mm_sub_ps(viCDEF, vi_max); const __m128 vxGHIJ = _mm_sub_ps(viGHIJ, vi_max); // Compute reduced argument elements := round(x / log(2)). __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vx0123, vlog2e), vmagic_bias); __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vx4567, vlog2e), vmagic_bias); __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vx89AB, vlog2e), vmagic_bias); __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vxCDEF, vlog2e), vmagic_bias); __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vxGHIJ, vlog2e), vmagic_bias); // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e. // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly. const __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23)); const __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23)); const __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23)); const __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23)); const __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23)); // Subtract the large number back to get final elements := round(x / log(2)). vn0123 = _mm_sub_ps(vn0123, vmagic_bias); vn4567 = _mm_sub_ps(vn4567, vmagic_bias); vn89AB = _mm_sub_ps(vn89AB, vmagic_bias); vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias); vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias); // Compute reduced argument t := x - elements * log(2). // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vx0123); __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vx4567); __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vx89AB); __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vxCDEF); __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vxGHIJ); vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF); vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ); // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc5, vt0123), vc4); __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc5, vt4567), vc4); __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc5, vt89AB), vc4); __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc5, vtCDEF), vc4); __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc5, vtGHIJ), vc4); vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3); vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3); vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3); vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3); vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3); vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2); vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2); vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2); vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2); vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2); vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc1); vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc1); vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc1); vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc1); vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc1); // Reconstruct the final f value: // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) // = s + (t * s) * p vt0123 = _mm_mul_ps(vt0123, vs0123); vt4567 = _mm_mul_ps(vt4567, vs4567); vt89AB = _mm_mul_ps(vt89AB, vs89AB); vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF); vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ); __m128 vf0123 = _mm_add_ps(_mm_mul_ps(vt0123, vp0123), vs0123); __m128 vf4567 = _mm_add_ps(_mm_mul_ps(vt4567, vp4567), vs4567); __m128 vf89AB = _mm_add_ps(_mm_mul_ps(vt89AB, vp89AB), vs89AB); __m128 vfCDEF = _mm_add_ps(_mm_mul_ps(vtCDEF, vpCDEF), vsCDEF); __m128 vfGHIJ = _mm_add_ps(_mm_mul_ps(vtGHIJ, vpGHIJ), vsGHIJ); // For inputs below zero cutoff, replace output with +0.0f. // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vx0123, vdenorm_cutoff), vf0123); vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vx4567, vdenorm_cutoff), vf4567); vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vx89AB, vdenorm_cutoff), vf89AB); vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vxCDEF, vdenorm_cutoff), vfCDEF); vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vxGHIJ, vdenorm_cutoff), vfGHIJ); // Store 20 (5x4) outputs at a time. _mm_storeu_ps(output, vf0123); _mm_storeu_ps(output + 4, vf4567); _mm_storeu_ps(output + 8, vf89AB); _mm_storeu_ps(output + 12, vfCDEF); _mm_storeu_ps(output + 16, vfGHIJ); output += 20; // Accumulate computed exponents. vacc0 = _mm_add_ps(vacc0, vf0123); vacc0 = _mm_add_ps(vacc0, vf4567); vacc0 = _mm_add_ps(vacc0, vf89AB); vacc0 = _mm_add_ps(vacc0, vfCDEF); vacc0 = _mm_add_ps(vacc0, vfGHIJ); } // Add up all accumulators to vacc0 vacc0 = _mm_add_ps(vacc0, vacc1); __m128 vacc = vacc0; for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) { // Load 4 inputs at a time. const __m128 vi = _mm_loadu_ps(input); input += 4; // Subtract maximum input x := i - i_max. This implies x <= 0. const __m128 vx = _mm_sub_ps(vi, vi_max); // Compute reduced argument elements := round(x / log(2)). __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias); // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e. // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly. const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23)); // Subtract the large number back to get final elements := round(x / log(2)). vn = _mm_sub_ps(vn, vmagic_bias); // Compute reduced argument t := x - elements * log(2). // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx); vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt); // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1); // Reconstruct the final f value: // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) // = s + (t * s) * p vt = _mm_mul_ps(vt, vs); __m128 vf = _mm_add_ps(_mm_mul_ps(vt, vp), vs); // For inputs below zero cutoff, replace output with +0.0f. // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. vf = _mm_andnot_ps(_mm_cmplt_ps(vx, vdenorm_cutoff), vf); // Store 4 outputs at a time. _mm_storeu_ps(output, vf); output += 4; // Accumulate computed exponents. vacc = _mm_add_ps(vacc, vf); } if (elements != 0) { assert(elements >= 1 * sizeof(float)); assert(elements <= 3 * sizeof(float)); // Load 4 inputs at a time. const __m128 vi = _mm_loadu_ps(input); // Subtract maximum input x := i - i_max. This implies x <= 0. const __m128 vx = _mm_sub_ps(vi, vi_max); // Compute reduced argument elements := round(x / log(2)). __m128 vn = _mm_add_ps(_mm_mul_ps(vx, vlog2e), vmagic_bias); // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e. // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly. const __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23)); // Subtract the large number back to get final elements := round(x / log(2)). vn = _mm_sub_ps(vn, vmagic_bias); // Compute reduced argument t := x - elements * log(2). // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vx); vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt); // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. __m128 vp = _mm_add_ps(_mm_mul_ps(vc5, vt), vc4); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc1); // Reconstruct the final f value: // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) // = s + (t * s) * p vt = _mm_mul_ps(vt, vs); __m128 vf = _mm_add_ps(_mm_mul_ps(vt, vp), vs); // For inputs below zero cutoff, replace output with +0.0f. // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. vf = _mm_andnot_ps(_mm_cmplt_ps(vx, vdenorm_cutoff), vf); if (elements & (2 * sizeof(float))) { // Store 2 outputs at a time. _mm_storel_pi((__m64*) output, vf); output += 2; // Accumulate 2 computed exponents. vacc = _mm_add_ps(vacc, _mm_movelh_ps(vf, _mm_setzero_ps())); vf = _mm_movehl_ps(vf, vf); } if (elements & (1 * sizeof(float))) { // Store 1 output at a time. _mm_store_ss(output, vf); // Accumulate 1 computed exponent. vacc = _mm_add_ss(vacc, vf); } } // Reduce 4 elements in the SIMD register vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc)); vacc = _mm_add_ss(vacc, _mm_shuffle_ps(vacc, vacc, _MM_SHUFFLE(2, 3, 0, 1))); _mm_store_ss(sum, vacc); } extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; void xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12( size_t n, const float* x, float* y, const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); assert(x != NULL); assert(y != NULL); const __m128 vprescale = _mm_load_ps(params->sse2_rr2_lut16_p3.prescale); const __m128 valpha = _mm_load_ps(params->sse2_rr2_lut16_p3.alpha); const __m128 vbeta = _mm_load_ps(params->sse2_rr2_lut16_p3.beta); const __m128 vsat_cutoff = _mm_load_ps(params->sse2_rr2_lut16_p3.sat_cutoff); const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut16_p3.magic_bias); const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut16_p3.log2e); const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut16_p3.index_mask); const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_hi); const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_lo); const __m128 vc3 = _mm_load_ps(params->sse2_rr2_lut16_p3.c3); const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut16_p3.c2); const __m128 vone = _mm_load_ps(params->sse2_rr2_lut16_p3.one); for (; n >= 12 * sizeof(float); n -= 12 * sizeof(float)) { __m128 vx0123 = _mm_loadu_ps(x); __m128 vx4567 = _mm_loadu_ps(x + 4); __m128 vx89AB = _mm_loadu_ps(x + 8); x += 12; const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale)); const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale)); const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale)); __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias); __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias); const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2); const __m128i ven0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 19); const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2); const __m128i ven4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 19); const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2); const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19); #if XNN_ARCH_X86_64 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123); const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123)); const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01))); const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23))); const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32)))); const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1); const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32)))); const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3); const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567); const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567)); const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45))); const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67))); const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32)))); const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5); const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32)))); const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7); const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB); const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB)); const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89))); const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB))); const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32)))); const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9); const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32)))); const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); #else // !XNN_ARCH_X86_64 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123); const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2); const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4); const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6); const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx0))); const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx2))); const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx1))); const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1); const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx3))); const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3); const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567); const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2); const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4); const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6); const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx4))); const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx6))); const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx5))); const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5); const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx7))); const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7); const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB); const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2); const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4); const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6); const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx8))); const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxA))); const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx9))); const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9); const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB))); const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB); const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); #endif // XNN_ARCH_X86_64 vn0123 = _mm_sub_ps(vn0123, vmagic_bias); __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ven0123)); vn4567 = _mm_sub_ps(vn4567, vmagic_bias); __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ven4567)); vn89AB = _mm_sub_ps(vn89AB, vmagic_bias); __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB)); __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123); __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567); __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB); vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123); vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567); vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB); __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc3, vt0123), vc2); __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc3, vt4567), vc2); __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc3, vt89AB), vc2); vp0123 = _mm_mul_ps(vp0123, vt0123); vp4567 = _mm_mul_ps(vp4567, vt4567); vp89AB = _mm_mul_ps(vp89AB, vt89AB); vt0123 = _mm_mul_ps(vt0123, vs0123); vs0123 = _mm_sub_ps(vs0123, vone); vt4567 = _mm_mul_ps(vt4567, vs4567); vs4567 = _mm_sub_ps(vs4567, vone); vt89AB = _mm_mul_ps(vt89AB, vs89AB); vs89AB = _mm_sub_ps(vs89AB, vone); vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123); vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567); vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB); const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha); const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha); const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha); const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123))); vx0123 = _mm_mul_ps(vx0123, vbeta); const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567))); vx4567 = _mm_mul_ps(vx4567, vbeta); const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB))); vx89AB = _mm_mul_ps(vx89AB, vbeta); const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123)); const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567)); const __m128 vy89AB = _mm_or_ps(_mm_and_ps(ve89AB, vm89AB), _mm_andnot_ps(vm89AB, vx89AB)); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); _mm_storeu_ps(y + 8, vy89AB); y += 12; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { __m128 vx = _mm_loadu_ps(x); x += 4; const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale)); __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias); const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19); const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2); #if XNN_ARCH_X86_64 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx); const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx)); const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #else // !XNN_ARCH_X86_64 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx)))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4)))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #endif // XNN_ARCH_X86_64 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi); __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven)); vn = _mm_sub_ps(vn, vmagic_bias); __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz); vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt); __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2); vp = _mm_mul_ps(vp, vt); vt = _mm_mul_ps(vt, vs); vs = _mm_sub_ps(vs, vone); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt); const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha); const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx))); vx = _mm_mul_ps(vx, vbeta); const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx)); _mm_storeu_ps(y, vy); y += 4; } if XNN_UNLIKELY(n != 0) { __m128 vx = _mm_loadu_ps(x); const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale)); __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias); const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19); const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2); #if XNN_ARCH_X86_64 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx); const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx)); const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #else // !XNN_ARCH_X86_64 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx)))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4)))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #endif // XNN_ARCH_X86_64 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi); __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven)); vn = _mm_sub_ps(vn, vmagic_bias); __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz); vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt); __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2); vp = _mm_mul_ps(vp, vt); vt = _mm_mul_ps(vt, vs); vs = _mm_sub_ps(vs, vone); vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt); const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha); const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx))); vx = _mm_mul_ps(vx, vbeta); __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx)); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vy); } } } void xnn_f32_vlrelu_ukernel__sse2_x8( size_t n, const float* x, float* y, const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); const __m128 vslope = _mm_load_ps(params->sse.slope); for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx0123 = _mm_loadu_ps(x); const __m128 vx4567 = _mm_loadu_ps(x + 4); x += 8; __m128 vacc0123 = _mm_mul_ps(vx0123, vslope); const __m128 vmask0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123))); __m128 vacc4567 = _mm_mul_ps(vx4567, vslope); const __m128 vmask4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567))); vacc0123 = _mm_or_ps(_mm_and_ps(vacc0123, vmask0123), _mm_andnot_ps(vmask0123, vx0123)); vacc4567 = _mm_or_ps(_mm_and_ps(vacc4567, vmask4567), _mm_andnot_ps(vmask4567, vx4567)); _mm_storeu_ps(y, vacc0123); _mm_storeu_ps(y + 4, vacc4567); y += 8; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { const __m128 vx = _mm_loadu_ps(x); x += 4; __m128 vacc = _mm_mul_ps(vx, vslope); const __m128 vmask = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx))); vacc = _mm_or_ps(_mm_and_ps(vacc, vmask), _mm_andnot_ps(vmask, vx)); _mm_storeu_ps(y, vacc); y += 4; } if XNN_UNLIKELY(n != 0) { const __m128 vx = _mm_loadu_ps(x); __m128 vacc = _mm_mul_ps(vx, vslope); const __m128 vmask = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx))); vacc = _mm_or_ps(_mm_and_ps(vacc, vmask), _mm_andnot_ps(vmask, vx)); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vacc); vacc = _mm_movehl_ps(vacc, vacc); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vacc); } } } void xnn_f32_vrndd_ukernel__sse2_x8( size_t n, const float* x, float* y, const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); const __m128i vmagic = _mm_load_si128((const __m128i*) params->sse2.sign_mask); const __m128 vone = _mm_load_ps(params->sse2.one); for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx0123 = _mm_loadu_ps(x); const __m128 vx4567 = _mm_loadu_ps(x + 4); x += 8; const __m128i vintx0123 = _mm_cvttps_epi32(vx0123); const __m128i vintx4567 = _mm_cvttps_epi32(vx4567); const __m128 vrndmask0123 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx0123, vmagic))); const __m128 vrndmask4567 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx4567, vmagic))); const __m128 vprerndx0123 = _mm_cvtepi32_ps(vintx0123); const __m128 vprerndx4567 = _mm_cvtepi32_ps(vintx4567); const __m128 vrndx0123 = _mm_or_ps(_mm_and_ps(vx0123, vrndmask0123), _mm_andnot_ps(vrndmask0123, vprerndx0123)); const __m128 vrndx4567 = _mm_or_ps(_mm_and_ps(vx4567, vrndmask4567), _mm_andnot_ps(vrndmask4567, vprerndx4567)); const __m128 vy0123 = _mm_sub_ps(vrndx0123, _mm_and_ps(_mm_cmpgt_ps(vrndx0123, vx0123), vone)); const __m128 vy4567 = _mm_sub_ps(vrndx4567, _mm_and_ps(_mm_cmpgt_ps(vrndx4567, vx4567), vone)); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); y += 8; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { const __m128 vx = _mm_loadu_ps(x); x += 4; const __m128i vintx = _mm_cvttps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vprerndx = _mm_cvtepi32_ps(vintx); const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx)); const __m128 vy = _mm_sub_ps(vrndx, _mm_and_ps(_mm_cmpgt_ps(vrndx, vx), vone)); _mm_storeu_ps(y, vy); y += 4; } if XNN_UNLIKELY(n != 0) { const __m128 vx = _mm_loadu_ps(x); const __m128i vintx = _mm_cvttps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vprerndx = _mm_cvtepi32_ps(vintx); const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx)); __m128 vy = _mm_sub_ps(vrndx, _mm_and_ps(_mm_cmpgt_ps(vrndx, vx), vone)); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vy); } } } void xnn_f32_vrndne_ukernel__sse2_x8( size_t n, const float* x, float* y, const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); const __m128i vmagic = _mm_load_si128((const __m128i*) params->sse2.sign_mask); for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx0123 = _mm_loadu_ps(x); const __m128 vx4567 = _mm_loadu_ps(x + 4); x += 8; const __m128i vintx0123 = _mm_cvtps_epi32(vx0123); const __m128i vintx4567 = _mm_cvtps_epi32(vx4567); const __m128 vrndmask0123 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx0123, vmagic))); const __m128 vrndmask4567 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx4567, vmagic))); const __m128 vrndx0123 = _mm_cvtepi32_ps(vintx0123); const __m128 vrndx4567 = _mm_cvtepi32_ps(vintx4567); const __m128 vy0123 = _mm_or_ps(_mm_and_ps(vx0123, vrndmask0123), _mm_andnot_ps(vrndmask0123, vrndx0123)); const __m128 vy4567 = _mm_or_ps(_mm_and_ps(vx4567, vrndmask4567), _mm_andnot_ps(vrndmask4567, vrndx4567)); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); y += 8; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { const __m128 vx = _mm_loadu_ps(x); x += 4; const __m128i vintx = _mm_cvtps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vrndx = _mm_cvtepi32_ps(vintx); const __m128 vy = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vrndx)); _mm_storeu_ps(y, vy); y += 4; } if XNN_UNLIKELY(n != 0) { const __m128 vx = _mm_loadu_ps(x); const __m128i vintx = _mm_cvtps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vrndx = _mm_cvtepi32_ps(vintx); __m128 vy = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vrndx)); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vy); } } } void xnn_f32_vrndu_ukernel__sse2_x8( size_t n, const float* x, float* y, const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); const __m128i vmagic = _mm_load_si128((const __m128i*) params->sse2.sign_mask); const __m128 vone = _mm_load_ps(params->sse2.one); for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx0123 = _mm_loadu_ps(x); const __m128 vx4567 = _mm_loadu_ps(x + 4); x += 8; const __m128i vintx0123 = _mm_cvttps_epi32(vx0123); const __m128i vintx4567 = _mm_cvttps_epi32(vx4567); const __m128 vrndmask0123 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx0123, vmagic))); const __m128 vrndmask4567 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx4567, vmagic))); const __m128 vprerndx0123 = _mm_cvtepi32_ps(vintx0123); const __m128 vprerndx4567 = _mm_cvtepi32_ps(vintx4567); const __m128 vrndx0123 = _mm_or_ps(_mm_and_ps(vx0123, vrndmask0123), _mm_andnot_ps(vrndmask0123, vprerndx0123)); const __m128 vrndx4567 = _mm_or_ps(_mm_and_ps(vx4567, vrndmask4567), _mm_andnot_ps(vrndmask4567, vprerndx4567)); const __m128 vadjmask0123 = _mm_or_ps(_mm_cmpge_ps(vrndx0123, vx0123), _mm_castsi128_ps(vmagic)); const __m128 vadjmask4567 = _mm_or_ps(_mm_cmpge_ps(vrndx4567, vx4567), _mm_castsi128_ps(vmagic)); const __m128 vadjrndx0123 = _mm_add_ps(vrndx0123, vone); const __m128 vadjrndx4567 = _mm_add_ps(vrndx4567, vone); const __m128 vy0123 = _mm_or_ps(_mm_and_ps(vrndx0123, vadjmask0123), _mm_andnot_ps(vadjmask0123, vadjrndx0123)); const __m128 vy4567 = _mm_or_ps(_mm_and_ps(vrndx4567, vadjmask4567), _mm_andnot_ps(vadjmask4567, vadjrndx4567)); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); y += 8; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { const __m128 vx = _mm_loadu_ps(x); x += 4; const __m128i vintx = _mm_cvttps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vprerndx = _mm_cvtepi32_ps(vintx); const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx)); const __m128 vadjmask = _mm_or_ps(_mm_cmpge_ps(vrndx, vx), _mm_castsi128_ps(vmagic)); const __m128 vadjrndx = _mm_add_ps(vrndx, vone); const __m128 vy = _mm_or_ps(_mm_and_ps(vrndx, vadjmask), _mm_andnot_ps(vadjmask, vadjrndx)); _mm_storeu_ps(y, vy); y += 4; } if XNN_UNLIKELY(n != 0) { const __m128 vx = _mm_loadu_ps(x); const __m128i vintx = _mm_cvttps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vprerndx = _mm_cvtepi32_ps(vintx); const __m128 vrndx = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vprerndx)); const __m128 vadjmask = _mm_or_ps(_mm_cmpge_ps(vrndx, vx), _mm_castsi128_ps(vmagic)); const __m128 vadjrndx = _mm_add_ps(vrndx, vone); __m128 vy = _mm_or_ps(_mm_and_ps(vrndx, vadjmask), _mm_andnot_ps(vadjmask, vadjrndx)); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vy); } } } void xnn_f32_vrndz_ukernel__sse2_x8( size_t n, const float* x, float* y, const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(float) == 0); const __m128i vmagic = _mm_load_si128((const __m128i*) params->sse2.sign_mask); for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx0123 = _mm_loadu_ps(x); const __m128 vx4567 = _mm_loadu_ps(x + 4); x += 8; const __m128i vintx0123 = _mm_cvttps_epi32(vx0123); const __m128i vintx4567 = _mm_cvttps_epi32(vx4567); const __m128 vrndmask0123 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx0123, vmagic))); const __m128 vrndmask4567 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx4567, vmagic))); const __m128 vrndx0123 = _mm_cvtepi32_ps(vintx0123); const __m128 vrndx4567 = _mm_cvtepi32_ps(vintx4567); const __m128 vy0123 = _mm_or_ps(_mm_and_ps(vx0123, vrndmask0123), _mm_andnot_ps(vrndmask0123, vrndx0123)); const __m128 vy4567 = _mm_or_ps(_mm_and_ps(vx4567, vrndmask4567), _mm_andnot_ps(vrndmask4567, vrndx4567)); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); y += 8; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { const __m128 vx = _mm_loadu_ps(x); x += 4; const __m128i vintx = _mm_cvttps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vrndx = _mm_cvtepi32_ps(vintx); const __m128 vy = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vrndx)); _mm_storeu_ps(y, vy); y += 4; } if XNN_UNLIKELY(n != 0) { const __m128 vx = _mm_loadu_ps(x); const __m128i vintx = _mm_cvttps_epi32(vx); const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); const __m128 vrndx = _mm_cvtepi32_ps(vintx); __m128 vy = _mm_or_ps(_mm_and_ps(vx, vrndmask), _mm_andnot_ps(vrndmask, vrndx)); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vy); } } } extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64]; void xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x8( size_t n, const float* x, float* y, const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n % sizeof(float) == 0); const __m128 vsign_mask = _mm_load_ps(params->sse2_rr2_lut64_p2.sign_mask); const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut64_p2.magic_bias); const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut64_p2.log2e); const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut64_p2.index_mask); const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_hi); const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_lo); const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut64_p2.c2); const __m128 vone = _mm_load_ps(params->sse2_rr2_lut64_p2.one); const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_lut64_p2.denorm_cutoff); for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { const __m128 vx0123 = _mm_loadu_ps(x); const __m128 vx4567 = _mm_loadu_ps(x + 4); x += 8; const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask); const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask); __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias); __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias); const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17); const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17); const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2); const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2); #if XNN_ARCH_X86_64 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123); const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123)); const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01))); const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23))); const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32)))); const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1); const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32)))); const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3); const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567); const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567)); const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45))); const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67))); const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32)))); const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5); const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32)))); const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7); const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); #else // !XNN_ARCH_X86_64 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123); const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2); const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4); const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6); const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0))); const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2))); const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1))); const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1); const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3))); const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3); const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567); const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2); const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4); const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6); const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4))); const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6))); const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5))); const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5); const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7))); const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7); const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); #endif // XNN_ARCH_X86_64 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123)); const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567)); vn0123 = _mm_sub_ps(vn0123, vmagic_bias); vn4567 = _mm_sub_ps(vn4567, vmagic_bias); __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi)); __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi)); vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo)); vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo)); __m128 vp0123 = _mm_mul_ps(vt0123, vc2); __m128 vp4567 = _mm_mul_ps(vt4567, vc2); vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123)); vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567)); const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123)); const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567)); __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone)); __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone)); vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123); vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567); const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123))); const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567))); vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123))); vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567))); _mm_storeu_ps(y, vf0123); _mm_storeu_ps(y + 4, vf4567); y += 8; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { const __m128 vx = _mm_loadu_ps(x); x += 4; const __m128 vz = _mm_or_ps(vx, vsign_mask); __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias); const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17); const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2); #if XNN_ARCH_X86_64 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx); const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx)); const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #else // !XNN_ARCH_X86_64 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx)))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4)))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #endif // XNN_ARCH_X86_64 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi); const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve)); vn = _mm_sub_ps(vn, vmagic_bias); __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi)); vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo)); __m128 vp = _mm_mul_ps(vt, vc2); vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt)); const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp)); __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone)); vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf); const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx))); vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf))); _mm_storeu_ps(y, vf); y += 4; } if XNN_UNLIKELY(n != 0) { const __m128 vx = _mm_loadu_ps(x); const __m128 vz = _mm_or_ps(vx, vsign_mask); __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias); const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17); const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2); #if XNN_ARCH_X86_64 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx); const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx)); const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #else // !XNN_ARCH_X86_64 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx)))); const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4)))); const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2)))); const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh); const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6)))); const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh); #endif // XNN_ARCH_X86_64 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi); const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve)); vn = _mm_sub_ps(vn, vmagic_bias); __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi)); vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo)); __m128 vp = _mm_mul_ps(vt, vc2); vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt)); const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp)); __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone)); vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf); const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx))); vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf))); if (n & (2 * sizeof(float))) { _mm_storel_pi((__m64*) y, vf); vf = _mm_movehl_ps(vf, vf); y += 2; } if (n & (1 * sizeof(float))) { _mm_store_ss(y, vf); } } } void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16( size_t channels, size_t output_width, const int8_t** input, const void* weights, int8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const int8_t* zero, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const int8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); } const int8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); } const int8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); } const int8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); } const int8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); } const int8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); } const int8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); } const int8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); } const int8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); } const int8_t* i9 = input[9]; assert(i9 != NULL); if XNN_UNPREDICTABLE(i9 != zero) { i9 = (const int8_t*) ((uintptr_t) i9 + input_offset); } const int8_t* i10 = input[10]; assert(i10 != NULL); if XNN_UNPREDICTABLE(i10 != zero) { i10 = (const int8_t*) ((uintptr_t) i10 + input_offset); } const int8_t* i11 = input[11]; assert(i11 != NULL); if XNN_UNPREDICTABLE(i11 != zero) { i11 = (const int8_t*) ((uintptr_t) i11 + input_offset); } const int8_t* i12 = input[12]; assert(i12 != NULL); if XNN_UNPREDICTABLE(i12 != zero) { i12 = (const int8_t*) ((uintptr_t) i12 + input_offset); } const int8_t* i13 = input[13]; assert(i13 != NULL); if XNN_UNPREDICTABLE(i13 != zero) { i13 = (const int8_t*) ((uintptr_t) i13 + input_offset); } const int8_t* i14 = input[14]; assert(i14 != NULL); if XNN_UNPREDICTABLE(i14 != zero) { i14 = (const int8_t*) ((uintptr_t) i14 + input_offset); } const int8_t* i15 = input[15]; assert(i15 != NULL); if XNN_UNPREDICTABLE(i15 != zero) { i15 = (const int8_t*) ((uintptr_t) i15 + input_offset); } const int8_t* i16 = input[16]; assert(i16 != NULL); if XNN_UNPREDICTABLE(i16 != zero) { i16 = (const int8_t*) ((uintptr_t) i16 + input_offset); } const int8_t* i17 = input[17]; assert(i17 != NULL); if XNN_UNPREDICTABLE(i17 != zero) { i17 = (const int8_t*) ((uintptr_t) i17 + input_offset); } const int8_t* i18 = input[18]; assert(i18 != NULL); if XNN_UNPREDICTABLE(i18 != zero) { i18 = (const int8_t*) ((uintptr_t) i18 + input_offset); } const int8_t* i19 = input[19]; assert(i19 != NULL); if XNN_UNPREDICTABLE(i19 != zero) { i19 = (const int8_t*) ((uintptr_t) i19 + input_offset); } const int8_t* i20 = input[20]; assert(i20 != NULL); if XNN_UNPREDICTABLE(i20 != zero) { i20 = (const int8_t*) ((uintptr_t) i20 + input_offset); } const int8_t* i21 = input[21]; assert(i21 != NULL); if XNN_UNPREDICTABLE(i21 != zero) { i21 = (const int8_t*) ((uintptr_t) i21 + input_offset); } const int8_t* i22 = input[22]; assert(i22 != NULL); if XNN_UNPREDICTABLE(i22 != zero) { i22 = (const int8_t*) ((uintptr_t) i22 + input_offset); } const int8_t* i23 = input[23]; assert(i23 != NULL); if XNN_UNPREDICTABLE(i23 != zero) { i23 = (const int8_t*) ((uintptr_t) i23 + input_offset); } const int8_t* i24 = input[24]; assert(i24 != NULL); if XNN_UNPREDICTABLE(i24 != zero) { i24 = (const int8_t*) ((uintptr_t) i24 + input_offset); } input = (const int8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); i3 += 8; const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); i4 += 8; const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); i5 += 8; const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); i6 += 8; const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); i7 += 8; const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); i8 += 8; const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9); const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))); i9 += 8; const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8); const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8); const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567); const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi)); const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10); const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))); i10 += 8; const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8); const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8); const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567); const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi)); const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11); const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))); i11 += 8; const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8); const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8); const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567); const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi)); const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12); const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))); i12 += 8; const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8); const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8); const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567); const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi)); const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13); const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))); i13 += 8; const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8); const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8); const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567); const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi)); const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14); const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))); i14 += 8; const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8); const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8); const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567); const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi)); const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15); const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))); i15 += 8; const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8); const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8); const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567); const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi)); const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16); const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))); i16 += 8; const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8); const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8); const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567); const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi)); const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17); const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))); i17 += 8; const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8); const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8); const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567); const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi)); const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18); const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))); i18 += 8; const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8); const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8); const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567); const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi)); const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19); const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))); i19 += 8; const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8); const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8); const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567); const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi)); const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20); const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))); i20 += 8; const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8); const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8); const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567); const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi)); const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21); const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))); i21 += 8; const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8); const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8); const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567); const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi)); const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22); const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))); i22 += 8; const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8); const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567); const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi)); const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23); const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))); i23 += 8; const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8); const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8); const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567); const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi)); const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24); const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))); i24 += 8; const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8); const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8); const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567); const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale0123 = _mm_loadu_ps((const float*) w); const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4); w = (const void*) ((const float*) w + 8); vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123); vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9); const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))); const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8); const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8); const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567); const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi)); const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10); const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))); const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8); const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8); const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567); const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi)); const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11); const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))); const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8); const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8); const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567); const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi)); const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12); const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))); const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8); const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8); const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567); const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi)); const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13); const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))); const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8); const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8); const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567); const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi)); const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14); const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))); const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8); const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8); const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567); const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi)); const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15); const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))); const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8); const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8); const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567); const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi)); const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16); const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))); const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8); const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8); const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567); const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi)); const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17); const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))); const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8); const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8); const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567); const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi)); const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18); const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))); const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8); const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8); const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567); const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi)); const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19); const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))); const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8); const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8); const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567); const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi)); const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20); const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))); const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8); const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8); const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567); const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi)); const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21); const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))); const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8); const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8); const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567); const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi)); const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22); const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))); const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8); const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567); const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi)); const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23); const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))); const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8); const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8); const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567); const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi)); const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24); const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))); const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8); const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8); const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567); const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t))); const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t) + 4 * sizeof(float))); vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123); vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__sse2_mul16( size_t channels, size_t output_width, const int8_t** input, const void* weights, int8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const int8_t* zero, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const int8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); } const int8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); } const int8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); } input = (const int8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale0123 = _mm_loadu_ps((const float*) w); const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4); w = (const void*) ((const float*) w + 8); vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123); vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t) + 4 * sizeof(float))); vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123); vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16( size_t channels, size_t output_width, const int8_t** input, const void* weights, int8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const int8_t* zero, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const int8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); } const int8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); } const int8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); } const int8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); } const int8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); } const int8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); } const int8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); } const int8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); } const int8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); } input = (const int8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); i3 += 8; const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); i4 += 8; const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); i5 += 8; const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); i6 += 8; const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); i7 += 8; const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); i8 += 8; const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale0123 = _mm_loadu_ps((const float*) w); const __m128 vscale4567 = _mm_loadu_ps((const float*) w + 4); w = (const void*) ((const float*) w + 8); vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123); vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale0123 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))); const __m128 vscale4567 = _mm_loadu_ps((const float*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t) + 4 * sizeof(float))); vscaled0123 = _mm_mul_ps(vscaled0123, vscale0123); vscaled4567 = _mm_mul_ps(vscaled4567, vscale4567); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 1); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); const int8_t* a0 = a; int8_t* c0 = c; do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); w = (const int32_t*) w + 4; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); const __m128 vscale0123 = _mm_load_ps((const float*) w); w = (const void*) ((const float*) w + 4); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); a0 = (const int8_t*) ((uintptr_t) a0 - kc); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (int8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 3); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); const int8_t* a0 = a; int8_t* c0 = c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { a1 = a0; c1 = c0; } const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { a2 = a1; c2 = c1; } do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); __m128i vacc1x0 = vacc0x0; __m128i vacc1x1 = vacc0x1; __m128i vacc1x2 = vacc0x2; __m128i vacc1x3 = vacc0x3; __m128i vacc2x0 = vacc0x0; __m128i vacc2x1 = vacc0x1; __m128i vacc2x2 = vacc0x2; __m128i vacc2x3 = vacc0x3; w = (const int32_t*) w + 4; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1); const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8); a1 += 8; const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2); const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8); a2 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0)); vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1)); vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2)); vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3)); vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2)); const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3)); const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2)); const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13)); __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); const __m128 vscale0123 = _mm_load_ps((const float*) w); w = (const void*) ((const float*) w + 4); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123); vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123); vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point); vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123); vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point); __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min); vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_si128(vout, 4); unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_si128(vout, 4); unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); a2 = (const int8_t*) ((uintptr_t) a2 - kc); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2)); c1 += 2; unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4)); c2 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (int8_t) _mm_cvtsi128_si32(vout); *c1 = (int8_t) _mm_extract_epi16(vout, 2); *c2 = (int8_t) _mm_extract_epi16(vout, 4); } nc = 0; } } while (nc != 0); } void xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const int8_t* zero, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 1); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (1 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); int8_t* c0 = c; do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); w = (const int32_t*) w + 4; size_t p = ks; do { const int8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); } a += 1; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } p -= 1 * sizeof(void*); } while (p != 0); const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); const __m128 vscale0123 = _mm_load_ps((const float*) w); w = (const void*) ((const float*) w + 4); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); a = (const int8_t**restrict) ((uintptr_t) a - ks); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (int8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const int8_t* zero, const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 3); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (3 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); int8_t* c0 = c; int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { c1 = c0; } int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { c2 = c1; } do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); __m128i vacc1x0 = vacc0x0; __m128i vacc1x1 = vacc0x1; __m128i vacc1x2 = vacc0x2; __m128i vacc1x3 = vacc0x3; __m128i vacc2x0 = vacc0x0; __m128i vacc2x1 = vacc0x1; __m128i vacc2x2 = vacc0x2; __m128i vacc2x3 = vacc0x3; w = (const int32_t*) w + 4; size_t p = ks; do { const int8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); } const int8_t* restrict a1 = a[1]; if XNN_UNPREDICTABLE(a1 != zero) { a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); } const int8_t* restrict a2 = a[2]; if XNN_UNPREDICTABLE(a2 != zero) { a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); } a += 3; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1); const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8); a1 += 8; const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2); const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8); a2 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0)); vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1)); vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2)); vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3)); vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } p -= 3 * sizeof(void*); } while (p != 0); const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2)); const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3)); const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2)); const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13)); __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); const __m128 vscale0123 = _mm_load_ps((const float*) w); w = (const void*) ((const float*) w + 4); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123); vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123); vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point); vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123); vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point); __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min); vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); if (nc >= 4) { unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2)))); c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)))); c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); a = (const int8_t**restrict) ((uintptr_t) a - ks); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4)); c2 += 2; unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2)); c1 += 2; unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c2 = (int8_t) _mm_extract_epi16(vout, 4); *c1 = (int8_t) _mm_extract_epi16(vout, 2); *c0 = (int8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16( size_t channels, size_t output_width, const int8_t** input, const void* weights, int8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const int8_t* zero, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const int8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); } const int8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); } const int8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); } const int8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); } const int8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); } const int8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); } const int8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); } const int8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); } const int8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); } const int8_t* i9 = input[9]; assert(i9 != NULL); if XNN_UNPREDICTABLE(i9 != zero) { i9 = (const int8_t*) ((uintptr_t) i9 + input_offset); } const int8_t* i10 = input[10]; assert(i10 != NULL); if XNN_UNPREDICTABLE(i10 != zero) { i10 = (const int8_t*) ((uintptr_t) i10 + input_offset); } const int8_t* i11 = input[11]; assert(i11 != NULL); if XNN_UNPREDICTABLE(i11 != zero) { i11 = (const int8_t*) ((uintptr_t) i11 + input_offset); } const int8_t* i12 = input[12]; assert(i12 != NULL); if XNN_UNPREDICTABLE(i12 != zero) { i12 = (const int8_t*) ((uintptr_t) i12 + input_offset); } const int8_t* i13 = input[13]; assert(i13 != NULL); if XNN_UNPREDICTABLE(i13 != zero) { i13 = (const int8_t*) ((uintptr_t) i13 + input_offset); } const int8_t* i14 = input[14]; assert(i14 != NULL); if XNN_UNPREDICTABLE(i14 != zero) { i14 = (const int8_t*) ((uintptr_t) i14 + input_offset); } const int8_t* i15 = input[15]; assert(i15 != NULL); if XNN_UNPREDICTABLE(i15 != zero) { i15 = (const int8_t*) ((uintptr_t) i15 + input_offset); } const int8_t* i16 = input[16]; assert(i16 != NULL); if XNN_UNPREDICTABLE(i16 != zero) { i16 = (const int8_t*) ((uintptr_t) i16 + input_offset); } const int8_t* i17 = input[17]; assert(i17 != NULL); if XNN_UNPREDICTABLE(i17 != zero) { i17 = (const int8_t*) ((uintptr_t) i17 + input_offset); } const int8_t* i18 = input[18]; assert(i18 != NULL); if XNN_UNPREDICTABLE(i18 != zero) { i18 = (const int8_t*) ((uintptr_t) i18 + input_offset); } const int8_t* i19 = input[19]; assert(i19 != NULL); if XNN_UNPREDICTABLE(i19 != zero) { i19 = (const int8_t*) ((uintptr_t) i19 + input_offset); } const int8_t* i20 = input[20]; assert(i20 != NULL); if XNN_UNPREDICTABLE(i20 != zero) { i20 = (const int8_t*) ((uintptr_t) i20 + input_offset); } const int8_t* i21 = input[21]; assert(i21 != NULL); if XNN_UNPREDICTABLE(i21 != zero) { i21 = (const int8_t*) ((uintptr_t) i21 + input_offset); } const int8_t* i22 = input[22]; assert(i22 != NULL); if XNN_UNPREDICTABLE(i22 != zero) { i22 = (const int8_t*) ((uintptr_t) i22 + input_offset); } const int8_t* i23 = input[23]; assert(i23 != NULL); if XNN_UNPREDICTABLE(i23 != zero) { i23 = (const int8_t*) ((uintptr_t) i23 + input_offset); } const int8_t* i24 = input[24]; assert(i24 != NULL); if XNN_UNPREDICTABLE(i24 != zero) { i24 = (const int8_t*) ((uintptr_t) i24 + input_offset); } input = (const int8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567)); const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); i3 += 8; const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567)); const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); i4 += 8; const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); i5 += 8; const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); i6 += 8; const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); i7 += 8; const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567)); const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); i8 += 8; const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9); const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))); i9 += 8; const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8); const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567)); const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567)); const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10); const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))); i10 += 8; const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8); const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567); const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11); const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))); i11 += 8; const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8); const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567)); const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567)); const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12); const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))); i12 += 8; const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8); const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567); const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13); const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))); i13 += 8; const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8); const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567)); const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567)); const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14); const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))); i14 += 8; const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8); const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567); const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15); const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))); i15 += 8; const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8); const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567)); const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567)); const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16); const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))); i16 += 8; const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8); const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567); const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17); const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))); i17 += 8; const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8); const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567)); const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567)); const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18); const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))); i18 += 8; const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8); const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567); const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19); const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))); i19 += 8; const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8); const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567)); const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567)); const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20); const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))); i20 += 8; const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8); const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567); const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21); const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))); i21 += 8; const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8); const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567)); const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567)); const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22); const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))); i22 += 8; const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8); const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567); const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23); const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))); i23 += 8; const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8); const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567)); const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567)); const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24); const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))); i24 += 8; const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8); const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567); const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(int8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567)); const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567)); const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567)); const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9); const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t))); const __m128i vxi9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi9x01234567, vi9x01234567), 8); const __m128i vxk9x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk9x01234567, vk9x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi9x01234567, vxk9x01234567)); const __m128i vsignprod9x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod9x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod9x01234567)); const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10); const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(int8_t))); const __m128i vxi10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi10x01234567, vi10x01234567), 8); const __m128i vxk10x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk10x01234567, vk10x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567); const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11); const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(int8_t))); const __m128i vxi11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi11x01234567, vi11x01234567), 8); const __m128i vxk11x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk11x01234567, vk11x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi11x01234567, vxk11x01234567)); const __m128i vsignprod11x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod11x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod11x01234567)); const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12); const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(int8_t))); const __m128i vxi12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi12x01234567, vi12x01234567), 8); const __m128i vxk12x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk12x01234567, vk12x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567); const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13); const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(int8_t))); const __m128i vxi13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi13x01234567, vi13x01234567), 8); const __m128i vxk13x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk13x01234567, vk13x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi13x01234567, vxk13x01234567)); const __m128i vsignprod13x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod13x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod13x01234567)); const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14); const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(int8_t))); const __m128i vxi14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi14x01234567, vi14x01234567), 8); const __m128i vxk14x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk14x01234567, vk14x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567); const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15); const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(int8_t))); const __m128i vxi15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi15x01234567, vi15x01234567), 8); const __m128i vxk15x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk15x01234567, vk15x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi15x01234567, vxk15x01234567)); const __m128i vsignprod15x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod15x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod15x01234567)); const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16); const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(int8_t))); const __m128i vxi16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi16x01234567, vi16x01234567), 8); const __m128i vxk16x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk16x01234567, vk16x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567); const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17); const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(int8_t))); const __m128i vxi17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi17x01234567, vi17x01234567), 8); const __m128i vxk17x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk17x01234567, vk17x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi17x01234567, vxk17x01234567)); const __m128i vsignprod17x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod17x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod17x01234567)); const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18); const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(int8_t))); const __m128i vxi18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi18x01234567, vi18x01234567), 8); const __m128i vxk18x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk18x01234567, vk18x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567); const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19); const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(int8_t))); const __m128i vxi19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi19x01234567, vi19x01234567), 8); const __m128i vxk19x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk19x01234567, vk19x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi19x01234567, vxk19x01234567)); const __m128i vsignprod19x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod19x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod19x01234567)); const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20); const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(int8_t))); const __m128i vxi20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi20x01234567, vi20x01234567), 8); const __m128i vxk20x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk20x01234567, vk20x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567); const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21); const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(int8_t))); const __m128i vxi21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi21x01234567, vi21x01234567), 8); const __m128i vxk21x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk21x01234567, vk21x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi21x01234567, vxk21x01234567)); const __m128i vsignprod21x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod21x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod21x01234567)); const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22); const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(int8_t))); const __m128i vxi22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi22x01234567, vi22x01234567), 8); const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567); const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23); const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(int8_t))); const __m128i vxi23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi23x01234567, vi23x01234567), 8); const __m128i vxk23x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk23x01234567, vk23x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi23x01234567, vxk23x01234567)); const __m128i vsignprod23x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod23x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod23x01234567)); const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24); const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(int8_t))); const __m128i vxi24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi24x01234567, vi24x01234567), 8); const __m128i vxk24x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk24x01234567, vk24x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567); const __m128i vsignprod24x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod24x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod24x01234567)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16( size_t channels, size_t output_width, const int8_t** input, const void* weights, int8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const int8_t* zero, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const int8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); } const int8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); } const int8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); } const int8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); } const int8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); } const int8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); } const int8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); } const int8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); } const int8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); } input = (const int8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567)); const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); i3 += 8; const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567)); const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); i4 += 8; const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); i5 += 8; const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); i6 += 8; const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); i7 += 8; const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567)); const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); i8 += 8; const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))); const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vxk0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk0x01234567, vk0x01234567), 8); __m128i vprod01234567 = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))); const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vxk1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk1x01234567, vk1x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi1x01234567, vxk1x01234567)); const __m128i vsignprod1x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod1x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod1x01234567)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vxk2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk2x01234567, vk2x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vxk3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk3x01234567, vk3x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi3x01234567, vxk3x01234567)); const __m128i vsignprod3x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod3x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod3x01234567)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vxk4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk4x01234567, vk4x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vxk5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk5x01234567, vk5x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi5x01234567, vxk5x01234567)); const __m128i vsignprod5x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod5x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod5x01234567)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); const __m128i vxk6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk6x01234567, vk6x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))); const __m128i vxi7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi7x01234567, vi7x01234567), 8); const __m128i vxk7x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk7x01234567, vk7x01234567), 8); vprod01234567 = _mm_add_epi16(vprod01234567, _mm_mullo_epi16(vxi7x01234567, vxk7x01234567)); const __m128i vsignprod7x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod7x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod7x01234567)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))); const __m128i vxi8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi8x01234567, vi8x01234567), 8); const __m128i vxk8x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk8x01234567, vk8x01234567), 8); vprod01234567 = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vsignprod8x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vprod01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod01234567, vsignprod8x01234567)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod01234567, vsignprod8x01234567)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qs8_f32_vcvt_ukernel__sse2_x32( size_t n, const int8_t* x, float* y, const union xnn_qs8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(int8_t) == 0); assert(x != NULL); assert(y != NULL); const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse2.sign_mask); const __m128i vmagic_exp = _mm_load_si128((const __m128i*) params->sse2.magic_exp); const __m128 vmagic_bias = _mm_load_ps(params->sse2.magic_bias); const __m128 vscale = _mm_load_ps(params->sse2.scale); const __m128i vzero = _mm_setzero_si128(); for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) { __m128i vx01234567 = _mm_loadl_epi64((const __m128i*) x); __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); __m128i vxGHIJKLMN = _mm_loadl_epi64((const __m128i*) (x + 16)); __m128i vxOPQRSTUV = _mm_loadl_epi64((const __m128i*) (x + 24)); x += 32; vx01234567 = _mm_xor_si128(vx01234567, vsign_mask); vx89ABCDEF = _mm_xor_si128(vx89ABCDEF, vsign_mask); vxGHIJKLMN = _mm_xor_si128(vxGHIJKLMN, vsign_mask); vxOPQRSTUV = _mm_xor_si128(vxOPQRSTUV, vsign_mask); vx01234567 = _mm_unpacklo_epi8(vx01234567, vzero); vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, vzero); vxOPQRSTUV = _mm_unpacklo_epi8(vxOPQRSTUV, vzero); __m128 vy0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(vx01234567, vmagic_exp)); __m128 vy4567 = _mm_castsi128_ps(_mm_unpackhi_epi16(vx01234567, vmagic_exp)); __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); __m128 vyGHIJ = _mm_castsi128_ps(_mm_unpacklo_epi16(vxGHIJKLMN, vmagic_exp)); __m128 vyKLMN = _mm_castsi128_ps(_mm_unpackhi_epi16(vxGHIJKLMN, vmagic_exp)); __m128 vyOPQR = _mm_castsi128_ps(_mm_unpacklo_epi16(vxOPQRSTUV, vmagic_exp)); __m128 vySTUV = _mm_castsi128_ps(_mm_unpackhi_epi16(vxOPQRSTUV, vmagic_exp)); vy0123 = _mm_sub_ps(vy0123, vmagic_bias); vy4567 = _mm_sub_ps(vy4567, vmagic_bias); vy89AB = _mm_sub_ps(vy89AB, vmagic_bias); vyCDEF = _mm_sub_ps(vyCDEF, vmagic_bias); vyGHIJ = _mm_sub_ps(vyGHIJ, vmagic_bias); vyKLMN = _mm_sub_ps(vyKLMN, vmagic_bias); vyOPQR = _mm_sub_ps(vyOPQR, vmagic_bias); vySTUV = _mm_sub_ps(vySTUV, vmagic_bias); vy0123 = _mm_mul_ps(vy0123, vscale); vy4567 = _mm_mul_ps(vy4567, vscale); vy89AB = _mm_mul_ps(vy89AB, vscale); vyCDEF = _mm_mul_ps(vyCDEF, vscale); vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); vyKLMN = _mm_mul_ps(vyKLMN, vscale); vyOPQR = _mm_mul_ps(vyOPQR, vscale); vySTUV = _mm_mul_ps(vySTUV, vscale); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); _mm_storeu_ps(y + 8, vy89AB); _mm_storeu_ps(y + 12, vyCDEF); _mm_storeu_ps(y + 16, vyGHIJ); _mm_storeu_ps(y + 20, vyKLMN); _mm_storeu_ps(y + 24, vyOPQR); _mm_storeu_ps(y + 28, vySTUV); y += 32; } for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) { __m128i vx = _mm_loadl_epi64((const __m128i*) x); vx = _mm_xor_si128(vx, vsign_mask); vx = _mm_unpacklo_epi8(vx, vzero); x += 8; __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp)); __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); vy_lo = _mm_sub_ps(vy_lo, vmagic_bias); vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); vy_lo = _mm_mul_ps(vy_lo, vscale); vy_hi = _mm_mul_ps(vy_hi, vscale); _mm_storeu_ps(y, vy_lo); _mm_storeu_ps(y + 4, vy_hi); y += 8; } if XNN_UNLIKELY(n != 0) { assert(n >= 1 * sizeof(int8_t)); assert(n <= 7 * sizeof(int8_t)); __m128i vx = _mm_loadl_epi64((const __m128i*) x); vx = _mm_xor_si128(vx, vsign_mask); vx = _mm_unpacklo_epi8(vx, vzero); __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp)); vy = _mm_sub_ps(vy, vmagic_bias); vy = _mm_mul_ps(vy, vscale); if (n & (4 * sizeof(int8_t))) { _mm_storeu_ps(y, vy); vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); vy = _mm_sub_ps(vy, vmagic_bias); vy = _mm_mul_ps(vy, vscale); y += 4; } if (n & (2 * sizeof(int8_t))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(int8_t))) { _mm_store_ss(y, vy); } } } void xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c8( size_t rows, size_t channels, const int8_t* input, size_t input_stride, const int8_t* zero, int32_t* buffer, int8_t* output, const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(rows > 7); assert(channels != 0); const int8_t* i0 = input; const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride); const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride); const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride); const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride); const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride); const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride); const size_t input_increment = 7 * input_stride - round_up_po2(channels, 8) * sizeof(int8_t); const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias); int32_t* b = buffer; size_t c = channels; for (; c != 0; c = doz(c, 8)) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567); vacc0123 = _mm_add_epi32(vacc0123, vinit_bias); vacc4567 = _mm_add_epi32(vacc4567, vinit_bias); _mm_store_si128((__m128i*) b, vacc0123); _mm_store_si128((__m128i*) (b + 4), vacc4567); b += 8; } for (rows -= 7; rows > 7; rows -= 7) { i0 = (const int8_t*) ((uintptr_t) i0 + input_increment); i1 = (const int8_t*) ((uintptr_t) i1 + input_increment); i2 = (const int8_t*) ((uintptr_t) i2 + input_increment); i3 = (const int8_t*) ((uintptr_t) i3 + input_increment); i4 = (const int8_t*) ((uintptr_t) i4 + input_increment); i5 = (const int8_t*) ((uintptr_t) i5 + input_increment); i6 = (const int8_t*) ((uintptr_t) i6 + input_increment); int32_t* b = buffer; size_t c = channels; for (; c != 0; c = doz(c, 8)) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) b)); vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (b + 4))); _mm_store_si128((__m128i*) b, vacc0123); _mm_store_si128((__m128i*) (b + 4), vacc4567); b += 8; } } i0 = (const int8_t*) ((uintptr_t) i0 + input_increment); i1 = (const int8_t*) ((uintptr_t) i1 + input_increment); if XNN_UNPREDICTABLE(rows < 2) { i1 = zero; } i2 = (const int8_t*) ((uintptr_t) i2 + input_increment); if XNN_UNPREDICTABLE(rows <= 2) { i2 = zero; } i3 = (const int8_t*) ((uintptr_t) i3 + input_increment); if XNN_UNPREDICTABLE(rows < 4) { i3 = zero; } i4 = (const int8_t*) ((uintptr_t) i4 + input_increment); if XNN_UNPREDICTABLE(rows <= 4) { i4 = zero; } i5 = (const int8_t*) ((uintptr_t) i5 + input_increment); if XNN_UNPREDICTABLE(rows < 6) { i5 = zero; } i6 = (const int8_t*) ((uintptr_t) i6 + input_increment); if XNN_UNPREDICTABLE(rows <= 6) { i6 = zero; } const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); for (; channels >= 8; channels -= 8) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer)); vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4))); buffer += 8; __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(channels != 0) { { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer)); vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4))); buffer += 8; __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (channels & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); if (channels & 2) { unaligned_store_u16(output, (uint16_t) vout0123); vout0123 >>= 16; output += 2; } if (channels & 1) { *output = (int8_t) vout0123; } } } } void xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8( size_t rows, size_t channels, const int8_t* input, size_t input_stride, const int8_t* zero, int8_t* output, const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(rows != 0); assert(rows <= 7); assert(channels != 0); const int8_t* i0 = input; const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride); if XNN_UNPREDICTABLE(rows < 2) { i1 = zero; } const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride); if XNN_UNPREDICTABLE(rows <= 2) { i2 = zero; } const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride); if XNN_UNPREDICTABLE(rows < 4) { i3 = zero; } const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride); if XNN_UNPREDICTABLE(rows <= 4) { i4 = zero; } const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride); if XNN_UNPREDICTABLE(rows < 6) { i5 = zero; } const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride); if XNN_UNPREDICTABLE(rows <= 6) { i6 = zero; } const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); for (; channels >= 8; channels -= 8) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567); vacc0123 = _mm_add_epi32(vacc0123, vinit_bias); vacc4567 = _mm_add_epi32(vacc4567, vinit_bias); __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(channels != 0) { { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi0x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi0x01234567, vi0x01234567), 8); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vxi1x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi1x01234567, vi1x01234567), 8); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi2x01234567, vi2x01234567), 8); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi3x01234567, vi3x01234567), 8); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi4x01234567, vi4x01234567), 8); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi5x01234567, vi5x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vi6x01234567, vi6x01234567), 8); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); const __m128i vsgnacc01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vsgnacc01234567); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vsgnacc01234567); vacc0123 = _mm_add_epi32(vacc0123, vinit_bias); vacc4567 = _mm_add_epi32(vacc4567, vinit_bias); __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (channels & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); if (channels & 2) { unaligned_store_u16(output, (uint16_t) vout0123); vout0123 >>= 16; output += 2; } if (channels & 1) { *output = (int8_t) vout0123; } } } } void xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 1); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); const int8_t* a0 = a; int8_t* c0 = c; do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); w = (const int32_t*) w + 4; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); a0 = (const int8_t*) ((uintptr_t) a0 - kc); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (int8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 3); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); const int8_t* a0 = a; int8_t* c0 = c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { a1 = a0; c1 = c0; } const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { a2 = a1; c2 = c1; } do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); __m128i vacc1x0 = vacc0x0; __m128i vacc1x1 = vacc0x1; __m128i vacc1x2 = vacc0x2; __m128i vacc1x3 = vacc0x3; __m128i vacc2x0 = vacc0x0; __m128i vacc2x1 = vacc0x1; __m128i vacc2x2 = vacc0x2; __m128i vacc2x3 = vacc0x3; w = (const int32_t*) w + 4; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1); const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8); a1 += 8; const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2); const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8); a2 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0)); vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1)); vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2)); vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3)); vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2)); const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3)); const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2)); const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13)); __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale); vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point); vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123); vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point); __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min); vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_si128(vout, 4); unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_si128(vout, 4); unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); a2 = (const int8_t*) ((uintptr_t) a2 - kc); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2)); c1 += 2; unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4)); c2 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (int8_t) _mm_cvtsi128_si32(vout); *c1 = (int8_t) _mm_extract_epi16(vout, 2); *c2 = (int8_t) _mm_extract_epi16(vout, 4); } nc = 0; } } while (nc != 0); } void xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const int8_t* zero, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 1); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (1 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); int8_t* c0 = c; do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); w = (const int32_t*) w + 4; size_t p = ks; do { const int8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); } a += 1; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } p -= 1 * sizeof(void*); } while (p != 0); const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); a = (const int8_t**restrict) ((uintptr_t) a - ks); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (int8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, int8_t* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const int8_t* zero, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 3); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (3 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); int8_t* c0 = c; int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { c1 = c0; } int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { c2 = c1; } do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); __m128i vacc1x0 = vacc0x0; __m128i vacc1x1 = vacc0x1; __m128i vacc1x2 = vacc0x2; __m128i vacc1x3 = vacc0x3; __m128i vacc2x0 = vacc0x0; __m128i vacc2x1 = vacc0x1; __m128i vacc2x2 = vacc0x2; __m128i vacc2x3 = vacc0x3; w = (const int32_t*) w + 4; size_t p = ks; do { const int8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); } const int8_t* restrict a1 = a[1]; if XNN_UNPREDICTABLE(a1 != zero) { a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); } const int8_t* restrict a2 = a[2]; if XNN_UNPREDICTABLE(a2 != zero) { a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); } a += 3; size_t k = 0; while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1); const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8); a1 += 8; const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2); const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8); a2 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0)); vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1)); vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2)); vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3)); vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3)); w = (const void*) ((const int8_t*) w + 32); k += 8 * sizeof(int8_t); } p -= 3 * sizeof(void*); } while (p != 0); const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2)); const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3)); const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2)); const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13)); __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale); vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point); vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123); vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point); __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min); vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min); __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123); if (nc >= 4) { unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2)))); c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)))); c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); a = (const int8_t**restrict) ((uintptr_t) a - ks); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4)); c2 += 2; unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2)); c1 += 2; unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c2 = (int8_t) _mm_extract_epi16(vout, 4); *c1 = (int8_t) _mm_extract_epi16(vout, 2); *c0 = (int8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x8( size_t n, const int8_t* input_a, const int8_t* input_b, int8_t* output, const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias); const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo); const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi); const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo); const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_hi); const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); input_a += 8; input_b += 8; va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi)); vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo)); vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi)); vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo)); vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (n & (4 * sizeof(int8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(int8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(int8_t))) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8( size_t n, const int8_t* input_a, const int8_t* input_b, int8_t* output, const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i vbias = _mm_add_epi32( _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)), _mm_load_si128((const __m128i*) params->sse2.bias)); const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo); const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi); const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); input_a += 8; va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (n & (4 * sizeof(int8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(int8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(int8_t))) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qs8_vcvt_ukernel__sse2_x32( size_t n, const int8_t* x, int8_t* y, const union xnn_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(int8_t) == 0); assert(x != NULL); assert(y != NULL); const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier); const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias); for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) { const __m128i vx0 = _mm_loadu_si128((const __m128i*) x); const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16)); x += 32; const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0); const __m128i vextx0 = _mm_unpacklo_epi8(vx0, vm0); const __m128i vextx1 = _mm_unpackhi_epi8(vx0, vm0); const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1); const __m128i vextx2 = _mm_unpacklo_epi8(vx1, vm1); const __m128i vextx3 = _mm_unpackhi_epi8(vx1, vm1); const __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier); const __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier); const __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier); const __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier); const __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier); const __m128i vprodhi2 = _mm_mulhi_epi16(vextx2, vmultiplier); const __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier); const __m128i vprodhi3 = _mm_mulhi_epi16(vextx3, vmultiplier); __m128i vacc0 = _mm_unpacklo_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_unpackhi_epi16(vprodlo0, vprodhi0); __m128i vacc2 = _mm_unpacklo_epi16(vprodlo1, vprodhi1); __m128i vacc3 = _mm_unpackhi_epi16(vprodlo1, vprodhi1); __m128i vacc4 = _mm_unpacklo_epi16(vprodlo2, vprodhi2); __m128i vacc5 = _mm_unpackhi_epi16(vprodlo2, vprodhi2); __m128i vacc6 = _mm_unpacklo_epi16(vprodlo3, vprodhi3); __m128i vacc7 = _mm_unpackhi_epi16(vprodlo3, vprodhi3); vacc0 = _mm_sub_epi32(vbias, vacc0); vacc1 = _mm_sub_epi32(vbias, vacc1); vacc2 = _mm_sub_epi32(vbias, vacc2); vacc3 = _mm_sub_epi32(vbias, vacc3); vacc4 = _mm_sub_epi32(vbias, vacc4); vacc5 = _mm_sub_epi32(vbias, vacc5); vacc6 = _mm_sub_epi32(vbias, vacc6); vacc7 = _mm_sub_epi32(vbias, vacc7); vacc0 = _mm_srai_epi32(vacc0, 8); vacc1 = _mm_srai_epi32(vacc1, 8); vacc2 = _mm_srai_epi32(vacc2, 8); vacc3 = _mm_srai_epi32(vacc3, 8); vacc4 = _mm_srai_epi32(vacc4, 8); vacc5 = _mm_srai_epi32(vacc5, 8); vacc6 = _mm_srai_epi32(vacc6, 8); vacc7 = _mm_srai_epi32(vacc7, 8); vacc0 = _mm_packs_epi32(vacc0, vacc1); vacc1 = _mm_packs_epi32(vacc2, vacc3); vacc2 = _mm_packs_epi32(vacc4, vacc5); vacc3 = _mm_packs_epi32(vacc6, vacc7); const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1); const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3); _mm_storeu_si128((__m128i*) y, vy0); _mm_storeu_si128((__m128i*) (y + 16), vy1); y += 32; } for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx); const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vm); const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vm); const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier); const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier); const __m128i vprodhi_lo = _mm_mulhi_epi16(vextx_lo, vmultiplier); const __m128i vprodhi_hi = _mm_mulhi_epi16(vextx_hi, vmultiplier); __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi); __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi); vacc_ll = _mm_sub_epi32(vbias, vacc_ll); vacc_lh = _mm_sub_epi32(vbias, vacc_lh); vacc_hl = _mm_sub_epi32(vbias, vacc_hl); vacc_hh = _mm_sub_epi32(vbias, vacc_hh); vacc_ll = _mm_srai_epi32(vacc_ll, 8); vacc_lh = _mm_srai_epi32(vacc_lh, 8); vacc_hl = _mm_srai_epi32(vacc_hl, 8); vacc_hh = _mm_srai_epi32(vacc_hh, 8); const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh); const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh); const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi); _mm_storeu_si128((__m128i*) y, vy); y += 16; } if XNN_UNLIKELY(n != 0) { assert(n >= 1 * sizeof(int8_t)); assert(n <= 15 * sizeof(int8_t)); const __m128i vx = _mm_loadu_si128((const __m128i*) x); const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx); const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vm); const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vm); const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier); const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier); const __m128i vprodhi_lo = _mm_mulhi_epi16(vextx_lo, vmultiplier); const __m128i vprodhi_hi = _mm_mulhi_epi16(vextx_hi, vmultiplier); __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi); __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi); vacc_ll = _mm_sub_epi32(vbias, vacc_ll); vacc_lh = _mm_sub_epi32(vbias, vacc_lh); vacc_hl = _mm_sub_epi32(vbias, vacc_hl); vacc_hh = _mm_sub_epi32(vbias, vacc_hh); vacc_ll = _mm_srai_epi32(vacc_ll, 8); vacc_lh = _mm_srai_epi32(vacc_lh, 8); vacc_hl = _mm_srai_epi32(vacc_hl, 8); vacc_hh = _mm_srai_epi32(vacc_hh, 8); const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh); const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh); __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi); if (n & (8 * sizeof(int8_t))) { _mm_storel_epi64((__m128i*) y, vy); vy = _mm_unpackhi_epi64(vy, vy); y += 8; } if (n & (4 * sizeof(int8_t))) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy)); vy = _mm_srli_epi64(vy, 32); y += 4; } uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy); if (n & (2 * sizeof(int8_t))) { unaligned_store_u16(y, (uint16_t) vy_lo); vy_lo >>= 16; y += 2; } if (n & (1 * sizeof(int8_t))) { *y = (int8_t) vy_lo; } } } void xnn_qs8_vlrelu_ukernel__sse2_x32( size_t n, const int8_t* x, int8_t* y, const union xnn_qs8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(int8_t) == 0); assert(x != NULL); assert(y != NULL); const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point); const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff); const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i vzero = _mm_setzero_si128(); for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) { const __m128i vx0 = _mm_loadu_si128((const __m128i*) x); const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16)); x += 32; const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0); __m128i vextx0 = _mm_unpacklo_epi8(vx0, vm0); __m128i vextx1 = _mm_unpackhi_epi8(vx0, vm0); const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1); __m128i vextx2 = _mm_unpacklo_epi8(vx1, vm1); __m128i vextx3 = _mm_unpackhi_epi8(vx1, vm1); __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point); vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0); __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point); vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1); __m128i vmultiplier2 = _mm_cmpgt_epi16(vextx2, vinput_zero_point); vextx2 = _mm_sub_epi16(vinput_zero_point, vextx2); __m128i vmultiplier3 = _mm_cmpgt_epi16(vextx3, vinput_zero_point); vextx3 = _mm_sub_epi16(vinput_zero_point, vextx3); vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff); vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff); vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff); vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff); vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base); vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base); vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base); vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base); __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0); __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1); __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier2); __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier3); vprodlo0 = _mm_srli_epi16(vprodlo0, 7); __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0); vprodlo1 = _mm_srli_epi16(vprodlo1, 7); __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1); vprodlo2 = _mm_srli_epi16(vprodlo2, 7); __m128i vprodhi2 = _mm_mulhi_epi16(vextx2, vmultiplier2); vprodlo3 = _mm_srli_epi16(vprodlo3, 7); __m128i vprodhi3 = _mm_mulhi_epi16(vextx3, vmultiplier3); vprodhi0 = _mm_slli_epi16(vprodhi0, 8); vprodlo0 = _mm_avg_epu16(vprodlo0, vzero); vprodhi1 = _mm_slli_epi16(vprodhi1, 8); vprodlo1 = _mm_avg_epu16(vprodlo1, vzero); vprodhi2 = _mm_slli_epi16(vprodhi2, 8); vprodlo2 = _mm_avg_epu16(vprodlo2, vzero); vprodhi3 = _mm_slli_epi16(vprodhi3, 8); vprodlo3 = _mm_avg_epu16(vprodlo3, vzero); __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1); __m128i vacc2 = _mm_add_epi16(vprodlo2, vprodhi2); __m128i vacc3 = _mm_add_epi16(vprodlo3, vprodhi3); vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point); vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point); const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1); const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3); _mm_storeu_si128((__m128i*) y, vy0); _mm_storeu_si128((__m128i*) (y + 16), vy1); y += 32; } for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx); __m128i vextx0 = _mm_unpacklo_epi8(vx, vm); __m128i vextx1 = _mm_unpackhi_epi8(vx, vm); __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point); __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point); vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0); vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1); vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff); vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff); vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base); vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base); __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0); __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1); vprodlo0 = _mm_srli_epi16(vprodlo0, 7); vprodlo1 = _mm_srli_epi16(vprodlo1, 7); __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0); __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1); vprodhi0 = _mm_slli_epi16(vprodhi0, 8); vprodhi1 = _mm_slli_epi16(vprodhi1, 8); vprodlo0 = _mm_avg_epu16(vprodlo0, vzero); vprodlo1 = _mm_avg_epu16(vprodlo1, vzero); __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1); vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); const __m128i vy = _mm_packs_epi16(vacc0, vacc1); _mm_storeu_si128((__m128i*) y, vy); y += 16; } if XNN_UNLIKELY(n != 0) { assert(n >= 1 * sizeof(int8_t)); assert(n <= 15 * sizeof(int8_t)); const __m128i vx = _mm_loadu_si128((const __m128i*) x); const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx); __m128i vextx0 = _mm_unpacklo_epi8(vx, vm); __m128i vextx1 = _mm_unpackhi_epi8(vx, vm); __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point); __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point); vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0); vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1); vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff); vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff); vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base); vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base); __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0); __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1); vprodlo0 = _mm_srli_epi16(vprodlo0, 7); vprodlo1 = _mm_srli_epi16(vprodlo1, 7); __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0); __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1); vprodhi0 = _mm_slli_epi16(vprodhi0, 8); vprodhi1 = _mm_slli_epi16(vprodhi1, 8); vprodlo0 = _mm_avg_epu16(vprodlo0, vzero); vprodlo1 = _mm_avg_epu16(vprodlo1, vzero); __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1); vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); __m128i vy = _mm_packs_epi16(vacc0, vacc1); if (n & (8 * sizeof(int8_t))) { _mm_storel_epi64((__m128i*) y, vy); vy = _mm_unpackhi_epi64(vy, vy); y += 8; } if (n & (4 * sizeof(int8_t))) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy)); vy = _mm_srli_epi64(vy, 32); y += 4; } uint32_t vy0 = (uint32_t) _mm_cvtsi128_si32(vy); if (n & (2 * sizeof(int8_t))) { unaligned_store_u16(y, (uint16_t) vy0); vy0 >>= 16; y += 2; } if (n & (1 * sizeof(int8_t))) { *y = (int8_t) vy0; } } } void xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8( size_t n, const int8_t* input_a, const int8_t* input_b, int8_t* output, const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point); const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max); for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); input_a += 8; input_b += 8; va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (n & (4 * sizeof(int8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(int8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(int8_t))) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8( size_t n, const int8_t* input_a, const int8_t* input_b, int8_t* output, const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max); __m128i vxb = _mm_sub_epi16( _mm_shuffle_epi32(_mm_cvtsi32_si128(UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b), 0), _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point)); for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); input_a += 8; va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); vout01234567 = _mm_max_epi16(vout01234567, voutput_min); vout01234567 = _mm_min_epi16(vout01234567, voutput_max); __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); if (n & (4 * sizeof(int8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(int8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(int8_t))) { *output = (int8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8( size_t output_pixels, size_t kernel_elements, size_t channels, const uint8_t** input, size_t input_offset, const uint8_t* zero, int32_t* buffer, uint8_t* output, size_t input_increment, size_t output_increment, const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(output_pixels != 0); assert(kernel_elements > 9); assert(channels != 0); const __m128i vbias = _mm_load_si128((const __m128i*) ¶ms->sse2.bias); const __m128i vzero = _mm_setzero_si128(); const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier); const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding); const __m128i vright_shift = _mm_loadl_epi64((const __m128i*) params->sse2.right_shift); do { { const uint8_t* i0 = *input++; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); } const uint8_t* i1 = *input++; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); } const uint8_t* i2 = *input++; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); } const uint8_t* i3 = *input++; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); } const uint8_t* i4 = *input++; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); } const uint8_t* i5 = *input++; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); } const uint8_t* i6 = *input++; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); } const uint8_t* i7 = *input++; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); } const uint8_t* i8 = *input++; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset); } int32_t* b = buffer; for (size_t c = 0; c < channels; c += 8) { const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8; const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8; const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero); const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8); const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5); const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45); const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67); const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678); const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero)); const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero)); _mm_store_si128((__m128i*) b, vacc_lo); _mm_store_si128((__m128i*) b + 1, vacc_hi); b += 8; } } size_t k = kernel_elements; for (k -= 9; k > 8; k -= 8) { const uint8_t* i0 = *input++; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); } const uint8_t* i1 = *input++; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); } const uint8_t* i2 = *input++; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); } const uint8_t* i3 = *input++; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); } const uint8_t* i4 = *input++; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); } const uint8_t* i5 = *input++; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); } const uint8_t* i6 = *input++; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); } const uint8_t* i7 = *input++; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); } int32_t* b = buffer; for (size_t c = 0; c < channels; c += 8) { const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8; __m128i vacc_lo = _mm_load_si128((const __m128i*) b); __m128i vacc_hi = _mm_load_si128((const __m128i*) b + 1); const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5); const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); const __m128i vsum0123 = _mm_add_epi16(vsum01, vsum23); const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67); const __m128i vsum = _mm_add_epi16(vsum0123, vsum4567); vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero)); vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero)); _mm_store_si128((__m128i*) b, vacc_lo); _mm_store_si128((__m128i*) b + 1, vacc_hi); b += 8; } } { const uint8_t* i0 = input[0]; assert(i0 != NULL); const uint8_t* i1 = input[1]; const uint8_t* i2 = input[2]; const uint8_t* i3 = input[3]; const uint8_t* i4 = input[4]; const uint8_t* i5 = input[5]; const uint8_t* i6 = input[6]; const uint8_t* i7 = input[7]; input = (const uint8_t**) ((uintptr_t) input + input_increment); if (k < 2) { i1 = zero; } assert(i1 != NULL); if (k <= 2) { i2 = zero; } assert(i2 != NULL); if (k < 4) { i3 = zero; } assert(i3 != NULL); if (k <= 4) { i4 = zero; } assert(i4 != NULL); if (k < 6) { i5 = zero; } assert(i5 != NULL); if (k <= 6) { i6 = zero; } assert(i6 != NULL); if (k < 8) { i7 = zero; } assert(i7 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); } if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); } if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); } if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); } if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); } if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); } if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); } if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); } size_t c = channels; int32_t* b = buffer; while (c >= 8) { const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8; __m128i vacc_lo = _mm_load_si128((const __m128i*) b); __m128i vacc_hi = _mm_load_si128((const __m128i*) b + 1); b += 8; const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5); const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); const __m128i vsum0123 = _mm_add_epi16(vsum01, vsum23); const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67); const __m128i vsum = _mm_add_epi16(vsum0123, vsum4567); vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero)); vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero)); const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo); const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi); const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi); const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier); const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier); const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier); const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier); const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift); const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift); const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift); const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift); const __m128i vabs_scaled_lo0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_hi0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi); __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi); vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_zero_point)); vout = _mm_packus_epi16(vout, vout); vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_max)); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_min)); _mm_storel_epi64((__m128i*) output, vout); output += 8; c -= 8; } if (c != 0) { const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); __m128i vacc_lo = _mm_load_si128((const __m128i*) b); __m128i vacc_hi = _mm_load_si128((const __m128i*) b + 1); const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5); const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); const __m128i vsum0123 = _mm_add_epi16(vsum01, vsum23); const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67); const __m128i vsum = _mm_add_epi16(vsum0123, vsum4567); vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vsum, vzero)); vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vsum, vzero)); const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo); const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi); const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi); const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier); const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier); const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier); const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier); const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift); const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift); const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift); const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift); const __m128i vabs_scaled_lo0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_hi0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi); __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi); vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_zero_point)); vout = _mm_packus_epi16(vout, vout); vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_max)); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_min)); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout)); output += 4; vout = _mm_srli_epi64(vout, 32); } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout, 0)); output += 2; vout = _mm_srli_epi32(vout, 16); } if (c & 1) { *output = (uint8_t) _mm_cvtsi128_si32(vout); output += 1; } } } output = (uint8_t*) ((uintptr_t) output + output_increment); } while (--output_pixels != 0); } void xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8( size_t output_pixels, size_t kernel_elements, size_t channels, const uint8_t** input, size_t input_offset, const uint8_t* zero, uint8_t* output, size_t input_increment, size_t output_increment, const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(output_pixels != 0); assert(kernel_elements != 0); assert(kernel_elements <= 9); assert(channels != 0); const __m128i vbias = _mm_load_si128((const __m128i*) ¶ms->sse2.bias); const __m128i vzero = _mm_setzero_si128(); const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier); const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding); const __m128i vright_shift = _mm_loadl_epi64((const __m128i*) params->sse2.right_shift); do { const uint8_t* i0 = input[0]; assert(i0 != NULL); const uint8_t* i1 = input[1]; const uint8_t* i2 = input[2]; const uint8_t* i3 = input[3]; const uint8_t* i4 = input[4]; const uint8_t* i5 = input[5]; const uint8_t* i6 = input[6]; const uint8_t* i7 = input[7]; const uint8_t* i8 = input[8]; input = (const uint8_t**) ((uintptr_t) input + input_increment); if (kernel_elements < 2) { i1 = zero; } assert(i1 != NULL); if (kernel_elements <= 2) { i2 = zero; } assert(i2 != NULL); if (kernel_elements < 4) { i3 = zero; } assert(i3 != NULL); if (kernel_elements <= 4) { i4 = zero; } assert(i4 != NULL); if (kernel_elements < 6) { i5 = zero; } assert(i5 != NULL); if (kernel_elements <= 6) { i6 = zero; } assert(i6 != NULL); if (kernel_elements < 8) { i7 = zero; } assert(i7 != NULL); if (kernel_elements <= 8) { i8 = zero; } assert(i8 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); } if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); } if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); } if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); } if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); } if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); } if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); } if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); } if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset); } size_t c = channels; while (c >= 8) { const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8; const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8; const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero); const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8); const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5); const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45); const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67); const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678); const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero)); const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero)); const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo); const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi); const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi); const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier); const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier); const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier); const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier); const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift); const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift); const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift); const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift); const __m128i vabs_scaled_lo0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_hi0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi); __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi); vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_zero_point)); vout = _mm_packus_epi16(vout, vout); vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_max)); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_min)); _mm_storel_epi64((__m128i*) output, vout); output += 8; c -= 8; } if (c != 0) { const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero); const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8); const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5); const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45); const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67); const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678); const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero)); const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero)); const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo); const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi); const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi); const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier); const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier); const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier); const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier); const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift); const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift); const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift); const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift); const __m128i vabs_scaled_lo0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_hi0213 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo); const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi); __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi); vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_zero_point)); vout = _mm_packus_epi16(vout, vout); vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_max)); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) ¶ms->sse2.output_min)); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout)); output += 4; vout = _mm_srli_epi64(vout, 32); } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout, 0)); output += 2; vout = _mm_srli_epi32(vout, 16); } if (c & 1) { *output = (uint8_t) _mm_cvtsi128_si32(vout); output += 1; } } output = (uint8_t*) ((uintptr_t) output + output_increment); } while (--output_pixels != 0); } void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16( size_t channels, size_t output_width, const uint8_t** input, const void* weights, uint8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const uint8_t* zero, const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const uint8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); } const uint8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); } const uint8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); } const uint8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); } const uint8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); } const uint8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); } const uint8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); } const uint8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); } const uint8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset); } const uint8_t* i9 = input[9]; assert(i9 != NULL); if XNN_UNPREDICTABLE(i9 != zero) { i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset); } const uint8_t* i10 = input[10]; assert(i10 != NULL); if XNN_UNPREDICTABLE(i10 != zero) { i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset); } const uint8_t* i11 = input[11]; assert(i11 != NULL); if XNN_UNPREDICTABLE(i11 != zero) { i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset); } const uint8_t* i12 = input[12]; assert(i12 != NULL); if XNN_UNPREDICTABLE(i12 != zero) { i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset); } const uint8_t* i13 = input[13]; assert(i13 != NULL); if XNN_UNPREDICTABLE(i13 != zero) { i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset); } const uint8_t* i14 = input[14]; assert(i14 != NULL); if XNN_UNPREDICTABLE(i14 != zero) { i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset); } const uint8_t* i15 = input[15]; assert(i15 != NULL); if XNN_UNPREDICTABLE(i15 != zero) { i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset); } const uint8_t* i16 = input[16]; assert(i16 != NULL); if XNN_UNPREDICTABLE(i16 != zero) { i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset); } const uint8_t* i17 = input[17]; assert(i17 != NULL); if XNN_UNPREDICTABLE(i17 != zero) { i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset); } const uint8_t* i18 = input[18]; assert(i18 != NULL); if XNN_UNPREDICTABLE(i18 != zero) { i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset); } const uint8_t* i19 = input[19]; assert(i19 != NULL); if XNN_UNPREDICTABLE(i19 != zero) { i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset); } const uint8_t* i20 = input[20]; assert(i20 != NULL); if XNN_UNPREDICTABLE(i20 != zero) { i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset); } const uint8_t* i21 = input[21]; assert(i21 != NULL); if XNN_UNPREDICTABLE(i21 != zero) { i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset); } const uint8_t* i22 = input[22]; assert(i22 != NULL); if XNN_UNPREDICTABLE(i22 != zero) { i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset); } const uint8_t* i23 = input[23]; assert(i23 != NULL); if XNN_UNPREDICTABLE(i23 != zero) { i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset); } const uint8_t* i24 = input[24]; assert(i24 != NULL); if XNN_UNPREDICTABLE(i24 != zero) { i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset); } input = (const uint8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point); for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t))); i0 += 8; const __m128i vzero = _mm_setzero_si128(); const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t))); i3 += 8; const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t))); i4 += 8; const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t))); i5 += 8; const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t))); i6 += 8; const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t))); i7 += 8; const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero); const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t))); i8 += 8; const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero); const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9); const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t))); i9 += 8; const __m128i vxi9x01234567 = _mm_unpacklo_epi8(vi9x01234567, vzero); const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk9x01234567, vzero), vk_zero_point); const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567); const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi)); const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10); const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t))); i10 += 8; const __m128i vxi10x01234567 = _mm_unpacklo_epi8(vi10x01234567, vzero); const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk10x01234567, vzero), vk_zero_point); const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567); const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi)); const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11); const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t))); i11 += 8; const __m128i vxi11x01234567 = _mm_unpacklo_epi8(vi11x01234567, vzero); const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk11x01234567, vzero), vk_zero_point); const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567); const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi)); const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12); const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t))); i12 += 8; const __m128i vxi12x01234567 = _mm_unpacklo_epi8(vi12x01234567, vzero); const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk12x01234567, vzero), vk_zero_point); const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567); const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi)); const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13); const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t))); i13 += 8; const __m128i vxi13x01234567 = _mm_unpacklo_epi8(vi13x01234567, vzero); const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk13x01234567, vzero), vk_zero_point); const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567); const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi)); const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14); const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t))); i14 += 8; const __m128i vxi14x01234567 = _mm_unpacklo_epi8(vi14x01234567, vzero); const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk14x01234567, vzero), vk_zero_point); const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567); const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi)); const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15); const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t))); i15 += 8; const __m128i vxi15x01234567 = _mm_unpacklo_epi8(vi15x01234567, vzero); const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk15x01234567, vzero), vk_zero_point); const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567); const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi)); const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16); const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t))); i16 += 8; const __m128i vxi16x01234567 = _mm_unpacklo_epi8(vi16x01234567, vzero); const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk16x01234567, vzero), vk_zero_point); const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567); const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi)); const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17); const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t))); i17 += 8; const __m128i vxi17x01234567 = _mm_unpacklo_epi8(vi17x01234567, vzero); const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk17x01234567, vzero), vk_zero_point); const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567); const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi)); const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18); const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t))); i18 += 8; const __m128i vxi18x01234567 = _mm_unpacklo_epi8(vi18x01234567, vzero); const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk18x01234567, vzero), vk_zero_point); const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567); const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi)); const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19); const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t))); i19 += 8; const __m128i vxi19x01234567 = _mm_unpacklo_epi8(vi19x01234567, vzero); const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk19x01234567, vzero), vk_zero_point); const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567); const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi)); const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20); const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t))); i20 += 8; const __m128i vxi20x01234567 = _mm_unpacklo_epi8(vi20x01234567, vzero); const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk20x01234567, vzero), vk_zero_point); const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567); const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi)); const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21); const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t))); i21 += 8; const __m128i vxi21x01234567 = _mm_unpacklo_epi8(vi21x01234567, vzero); const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk21x01234567, vzero), vk_zero_point); const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567); const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi)); const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22); const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t))); i22 += 8; const __m128i vxi22x01234567 = _mm_unpacklo_epi8(vi22x01234567, vzero); const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk22x01234567, vzero), vk_zero_point); const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567); const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi)); const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23); const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t))); i23 += 8; const __m128i vxi23x01234567 = _mm_unpacklo_epi8(vi23x01234567, vzero); const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk23x01234567, vzero), vk_zero_point); const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567); const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi)); const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24); const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t))); i24 += 8; const __m128i vxi24x01234567 = _mm_unpacklo_epi8(vi24x01234567, vzero); const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk24x01234567, vzero), vk_zero_point); const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567); const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 200 * sizeof(uint8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t))); const __m128i vzero = _mm_setzero_si128(); const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t))); const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t))); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t))); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t))); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t))); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t))); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t))); const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero); const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t))); const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero); const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); const __m128i vi9x01234567 = _mm_loadl_epi64((const __m128i*) i9); const __m128i vk9x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t))); const __m128i vxi9x01234567 = _mm_unpacklo_epi8(vi9x01234567, vzero); const __m128i vxk9x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk9x01234567, vzero), vk_zero_point); const __m128i vprod9x01234567lo = _mm_mullo_epi16(vxi9x01234567, vxk9x01234567); const __m128i vprod9x01234567hi = _mm_mulhi_epi16(vxi9x01234567, vxk9x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod9x01234567lo, vprod9x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod9x01234567lo, vprod9x01234567hi)); const __m128i vi10x01234567 = _mm_loadl_epi64((const __m128i*) i10); const __m128i vk10x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 80 * sizeof(uint8_t))); const __m128i vxi10x01234567 = _mm_unpacklo_epi8(vi10x01234567, vzero); const __m128i vxk10x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk10x01234567, vzero), vk_zero_point); const __m128i vprod10x01234567lo = _mm_mullo_epi16(vxi10x01234567, vxk10x01234567); const __m128i vprod10x01234567hi = _mm_mulhi_epi16(vxi10x01234567, vxk10x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod10x01234567lo, vprod10x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod10x01234567lo, vprod10x01234567hi)); const __m128i vi11x01234567 = _mm_loadl_epi64((const __m128i*) i11); const __m128i vk11x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 88 * sizeof(uint8_t))); const __m128i vxi11x01234567 = _mm_unpacklo_epi8(vi11x01234567, vzero); const __m128i vxk11x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk11x01234567, vzero), vk_zero_point); const __m128i vprod11x01234567lo = _mm_mullo_epi16(vxi11x01234567, vxk11x01234567); const __m128i vprod11x01234567hi = _mm_mulhi_epi16(vxi11x01234567, vxk11x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod11x01234567lo, vprod11x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod11x01234567lo, vprod11x01234567hi)); const __m128i vi12x01234567 = _mm_loadl_epi64((const __m128i*) i12); const __m128i vk12x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 96 * sizeof(uint8_t))); const __m128i vxi12x01234567 = _mm_unpacklo_epi8(vi12x01234567, vzero); const __m128i vxk12x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk12x01234567, vzero), vk_zero_point); const __m128i vprod12x01234567lo = _mm_mullo_epi16(vxi12x01234567, vxk12x01234567); const __m128i vprod12x01234567hi = _mm_mulhi_epi16(vxi12x01234567, vxk12x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod12x01234567lo, vprod12x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod12x01234567lo, vprod12x01234567hi)); const __m128i vi13x01234567 = _mm_loadl_epi64((const __m128i*) i13); const __m128i vk13x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 104 * sizeof(uint8_t))); const __m128i vxi13x01234567 = _mm_unpacklo_epi8(vi13x01234567, vzero); const __m128i vxk13x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk13x01234567, vzero), vk_zero_point); const __m128i vprod13x01234567lo = _mm_mullo_epi16(vxi13x01234567, vxk13x01234567); const __m128i vprod13x01234567hi = _mm_mulhi_epi16(vxi13x01234567, vxk13x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod13x01234567lo, vprod13x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod13x01234567lo, vprod13x01234567hi)); const __m128i vi14x01234567 = _mm_loadl_epi64((const __m128i*) i14); const __m128i vk14x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 112 * sizeof(uint8_t))); const __m128i vxi14x01234567 = _mm_unpacklo_epi8(vi14x01234567, vzero); const __m128i vxk14x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk14x01234567, vzero), vk_zero_point); const __m128i vprod14x01234567lo = _mm_mullo_epi16(vxi14x01234567, vxk14x01234567); const __m128i vprod14x01234567hi = _mm_mulhi_epi16(vxi14x01234567, vxk14x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod14x01234567lo, vprod14x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod14x01234567lo, vprod14x01234567hi)); const __m128i vi15x01234567 = _mm_loadl_epi64((const __m128i*) i15); const __m128i vk15x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 120 * sizeof(uint8_t))); const __m128i vxi15x01234567 = _mm_unpacklo_epi8(vi15x01234567, vzero); const __m128i vxk15x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk15x01234567, vzero), vk_zero_point); const __m128i vprod15x01234567lo = _mm_mullo_epi16(vxi15x01234567, vxk15x01234567); const __m128i vprod15x01234567hi = _mm_mulhi_epi16(vxi15x01234567, vxk15x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod15x01234567lo, vprod15x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod15x01234567lo, vprod15x01234567hi)); const __m128i vi16x01234567 = _mm_loadl_epi64((const __m128i*) i16); const __m128i vk16x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 128 * sizeof(uint8_t))); const __m128i vxi16x01234567 = _mm_unpacklo_epi8(vi16x01234567, vzero); const __m128i vxk16x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk16x01234567, vzero), vk_zero_point); const __m128i vprod16x01234567lo = _mm_mullo_epi16(vxi16x01234567, vxk16x01234567); const __m128i vprod16x01234567hi = _mm_mulhi_epi16(vxi16x01234567, vxk16x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod16x01234567lo, vprod16x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod16x01234567lo, vprod16x01234567hi)); const __m128i vi17x01234567 = _mm_loadl_epi64((const __m128i*) i17); const __m128i vk17x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 136 * sizeof(uint8_t))); const __m128i vxi17x01234567 = _mm_unpacklo_epi8(vi17x01234567, vzero); const __m128i vxk17x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk17x01234567, vzero), vk_zero_point); const __m128i vprod17x01234567lo = _mm_mullo_epi16(vxi17x01234567, vxk17x01234567); const __m128i vprod17x01234567hi = _mm_mulhi_epi16(vxi17x01234567, vxk17x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod17x01234567lo, vprod17x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod17x01234567lo, vprod17x01234567hi)); const __m128i vi18x01234567 = _mm_loadl_epi64((const __m128i*) i18); const __m128i vk18x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 144 * sizeof(uint8_t))); const __m128i vxi18x01234567 = _mm_unpacklo_epi8(vi18x01234567, vzero); const __m128i vxk18x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk18x01234567, vzero), vk_zero_point); const __m128i vprod18x01234567lo = _mm_mullo_epi16(vxi18x01234567, vxk18x01234567); const __m128i vprod18x01234567hi = _mm_mulhi_epi16(vxi18x01234567, vxk18x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod18x01234567lo, vprod18x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod18x01234567lo, vprod18x01234567hi)); const __m128i vi19x01234567 = _mm_loadl_epi64((const __m128i*) i19); const __m128i vk19x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 152 * sizeof(uint8_t))); const __m128i vxi19x01234567 = _mm_unpacklo_epi8(vi19x01234567, vzero); const __m128i vxk19x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk19x01234567, vzero), vk_zero_point); const __m128i vprod19x01234567lo = _mm_mullo_epi16(vxi19x01234567, vxk19x01234567); const __m128i vprod19x01234567hi = _mm_mulhi_epi16(vxi19x01234567, vxk19x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod19x01234567lo, vprod19x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod19x01234567lo, vprod19x01234567hi)); const __m128i vi20x01234567 = _mm_loadl_epi64((const __m128i*) i20); const __m128i vk20x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 160 * sizeof(uint8_t))); const __m128i vxi20x01234567 = _mm_unpacklo_epi8(vi20x01234567, vzero); const __m128i vxk20x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk20x01234567, vzero), vk_zero_point); const __m128i vprod20x01234567lo = _mm_mullo_epi16(vxi20x01234567, vxk20x01234567); const __m128i vprod20x01234567hi = _mm_mulhi_epi16(vxi20x01234567, vxk20x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod20x01234567lo, vprod20x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod20x01234567lo, vprod20x01234567hi)); const __m128i vi21x01234567 = _mm_loadl_epi64((const __m128i*) i21); const __m128i vk21x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 168 * sizeof(uint8_t))); const __m128i vxi21x01234567 = _mm_unpacklo_epi8(vi21x01234567, vzero); const __m128i vxk21x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk21x01234567, vzero), vk_zero_point); const __m128i vprod21x01234567lo = _mm_mullo_epi16(vxi21x01234567, vxk21x01234567); const __m128i vprod21x01234567hi = _mm_mulhi_epi16(vxi21x01234567, vxk21x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod21x01234567lo, vprod21x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod21x01234567lo, vprod21x01234567hi)); const __m128i vi22x01234567 = _mm_loadl_epi64((const __m128i*) i22); const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 176 * sizeof(uint8_t))); const __m128i vxi22x01234567 = _mm_unpacklo_epi8(vi22x01234567, vzero); const __m128i vxk22x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk22x01234567, vzero), vk_zero_point); const __m128i vprod22x01234567lo = _mm_mullo_epi16(vxi22x01234567, vxk22x01234567); const __m128i vprod22x01234567hi = _mm_mulhi_epi16(vxi22x01234567, vxk22x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod22x01234567lo, vprod22x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod22x01234567lo, vprod22x01234567hi)); const __m128i vi23x01234567 = _mm_loadl_epi64((const __m128i*) i23); const __m128i vk23x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 184 * sizeof(uint8_t))); const __m128i vxi23x01234567 = _mm_unpacklo_epi8(vi23x01234567, vzero); const __m128i vxk23x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk23x01234567, vzero), vk_zero_point); const __m128i vprod23x01234567lo = _mm_mullo_epi16(vxi23x01234567, vxk23x01234567); const __m128i vprod23x01234567hi = _mm_mulhi_epi16(vxi23x01234567, vxk23x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod23x01234567lo, vprod23x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod23x01234567lo, vprod23x01234567hi)); const __m128i vi24x01234567 = _mm_loadl_epi64((const __m128i*) i24); const __m128i vk24x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 192 * sizeof(uint8_t))); const __m128i vxi24x01234567 = _mm_unpacklo_epi8(vi24x01234567, vzero); const __m128i vxk24x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk24x01234567, vzero), vk_zero_point); const __m128i vprod24x01234567lo = _mm_mullo_epi16(vxi24x01234567, vxk24x01234567); const __m128i vprod24x01234567hi = _mm_mulhi_epi16(vxi24x01234567, vxk24x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod24x01234567lo, vprod24x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod24x01234567lo, vprod24x01234567hi)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (uint8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16( size_t channels, size_t output_width, const uint8_t** input, const void* weights, uint8_t* output, size_t input_stride, size_t output_increment, size_t input_offset, const uint8_t* zero, const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); do { const uint8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); } const uint8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); } const uint8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); } const uint8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); } const uint8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); } const uint8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); } const uint8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); } const uint8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); } const uint8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset); } input = (const uint8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; const __m128i vk_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point); for (; c >= 8; c -= 8) { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t))); i0 += 8; const __m128i vzero = _mm_setzero_si128(); const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t))); i1 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t))); i2 += 8; const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t))); i3 += 8; const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t))); i4 += 8; const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t))); i5 += 8; const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t))); i6 += 8; const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t))); i7 += 8; const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero); const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t))); i8 += 8; const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero); const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(uint8_t)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(c != 0) { { __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w); __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4)); const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(uint8_t))); const __m128i vzero = _mm_setzero_si128(); const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vxk0x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0x01234567, vzero), vk_zero_point); const __m128i vprod0x01234567lo = _mm_mullo_epi16(vxi0x01234567, vxk0x01234567); const __m128i vprod0x01234567hi = _mm_mulhi_epi16(vxi0x01234567, vxk0x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod0x01234567lo, vprod0x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod0x01234567lo, vprod0x01234567hi)); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(uint8_t))); const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vxk1x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1x01234567, vzero), vk_zero_point); const __m128i vprod1x01234567lo = _mm_mullo_epi16(vxi1x01234567, vxk1x01234567); const __m128i vprod1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod1x01234567lo, vprod1x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod1x01234567lo, vprod1x01234567hi)); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(uint8_t))); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vxk2x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2x01234567, vzero), vk_zero_point); const __m128i vprod2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); const __m128i vprod2x01234567hi = _mm_mulhi_epi16(vxi2x01234567, vxk2x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod2x01234567lo, vprod2x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod2x01234567lo, vprod2x01234567hi)); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(uint8_t))); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vxk3x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3x01234567, vzero), vk_zero_point); const __m128i vprod3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); const __m128i vprod3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod3x01234567lo, vprod3x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod3x01234567lo, vprod3x01234567hi)); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(uint8_t))); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vxk4x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4x01234567, vzero), vk_zero_point); const __m128i vprod4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); const __m128i vprod4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod4x01234567lo, vprod4x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod4x01234567lo, vprod4x01234567hi)); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); const __m128i vk5x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(uint8_t))); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vxk5x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk5x01234567, vzero), vk_zero_point); const __m128i vprod5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); const __m128i vprod5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod5x01234567lo, vprod5x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod5x01234567lo, vprod5x01234567hi)); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); const __m128i vk6x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(uint8_t))); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); const __m128i vxk6x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk6x01234567, vzero), vk_zero_point); const __m128i vprod6x01234567lo = _mm_mullo_epi16(vxi6x01234567, vxk6x01234567); const __m128i vprod6x01234567hi = _mm_mulhi_epi16(vxi6x01234567, vxk6x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod6x01234567lo, vprod6x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod6x01234567lo, vprod6x01234567hi)); const __m128i vi7x01234567 = _mm_loadl_epi64((const __m128i*) i7); const __m128i vk7x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(uint8_t))); const __m128i vxi7x01234567 = _mm_unpacklo_epi8(vi7x01234567, vzero); const __m128i vxk7x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk7x01234567, vzero), vk_zero_point); const __m128i vprod7x01234567lo = _mm_mullo_epi16(vxi7x01234567, vxk7x01234567); const __m128i vprod7x01234567hi = _mm_mulhi_epi16(vxi7x01234567, vxk7x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod7x01234567lo, vprod7x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod7x01234567lo, vprod7x01234567hi)); const __m128i vi8x01234567 = _mm_loadl_epi64((const __m128i*) i8); const __m128i vk8x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(uint8_t))); const __m128i vxi8x01234567 = _mm_unpacklo_epi8(vi8x01234567, vzero); const __m128i vxk8x01234567 = _mm_sub_epi16(_mm_unpacklo_epi8(vk8x01234567, vzero), vk_zero_point); const __m128i vprod8x01234567lo = _mm_mullo_epi16(vxi8x01234567, vxk8x01234567); const __m128i vprod8x01234567hi = _mm_mulhi_epi16(vxi8x01234567, vxk8x01234567); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vprod8x01234567lo, vprod8x01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vprod8x01234567lo, vprod8x01234567hi)); __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123); __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0123 = _mm_mul_ps(vscaled0123, vscale); vscaled4567 = _mm_mul_ps(vscaled4567, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point); vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vscaled0123); vacc4567 = _mm_cvtps_epi32(vscaled4567); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); if (c & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (c & 2) { unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (c & 1) { *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567); output += 1; } } } output = (uint8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); } void xnn_qu8_f32_vcvt_ukernel__sse2_x32( size_t n, const uint8_t* x, float* y, const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(uint8_t) == 0); assert(x != NULL); assert(y != NULL); const __m128i vmagic_exp = _mm_load_si128((const __m128i*) params->sse2.magic_exp); const __m128 vmagic_bias = _mm_load_ps(params->sse2.magic_bias); const __m128 vscale = _mm_load_ps(params->sse2.scale); const __m128i vzero = _mm_setzero_si128(); for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) { __m128i vx01234567 = _mm_loadl_epi64((const __m128i*) x); __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); __m128i vxGHIJKLMN = _mm_loadl_epi64((const __m128i*) (x + 16)); __m128i vxOPQRSTUV = _mm_loadl_epi64((const __m128i*) (x + 24)); x += 32; vx01234567 = _mm_unpacklo_epi8(vx01234567, vzero); vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, vzero); vxOPQRSTUV = _mm_unpacklo_epi8(vxOPQRSTUV, vzero); __m128 vy0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(vx01234567, vmagic_exp)); __m128 vy4567 = _mm_castsi128_ps(_mm_unpackhi_epi16(vx01234567, vmagic_exp)); __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); __m128 vyGHIJ = _mm_castsi128_ps(_mm_unpacklo_epi16(vxGHIJKLMN, vmagic_exp)); __m128 vyKLMN = _mm_castsi128_ps(_mm_unpackhi_epi16(vxGHIJKLMN, vmagic_exp)); __m128 vyOPQR = _mm_castsi128_ps(_mm_unpacklo_epi16(vxOPQRSTUV, vmagic_exp)); __m128 vySTUV = _mm_castsi128_ps(_mm_unpackhi_epi16(vxOPQRSTUV, vmagic_exp)); vy0123 = _mm_sub_ps(vy0123, vmagic_bias); vy4567 = _mm_sub_ps(vy4567, vmagic_bias); vy89AB = _mm_sub_ps(vy89AB, vmagic_bias); vyCDEF = _mm_sub_ps(vyCDEF, vmagic_bias); vyGHIJ = _mm_sub_ps(vyGHIJ, vmagic_bias); vyKLMN = _mm_sub_ps(vyKLMN, vmagic_bias); vyOPQR = _mm_sub_ps(vyOPQR, vmagic_bias); vySTUV = _mm_sub_ps(vySTUV, vmagic_bias); vy0123 = _mm_mul_ps(vy0123, vscale); vy4567 = _mm_mul_ps(vy4567, vscale); vy89AB = _mm_mul_ps(vy89AB, vscale); vyCDEF = _mm_mul_ps(vyCDEF, vscale); vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); vyKLMN = _mm_mul_ps(vyKLMN, vscale); vyOPQR = _mm_mul_ps(vyOPQR, vscale); vySTUV = _mm_mul_ps(vySTUV, vscale); _mm_storeu_ps(y, vy0123); _mm_storeu_ps(y + 4, vy4567); _mm_storeu_ps(y + 8, vy89AB); _mm_storeu_ps(y + 12, vyCDEF); _mm_storeu_ps(y + 16, vyGHIJ); _mm_storeu_ps(y + 20, vyKLMN); _mm_storeu_ps(y + 24, vyOPQR); _mm_storeu_ps(y + 28, vySTUV); y += 32; } for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) { __m128i vx = _mm_loadl_epi64((const __m128i*) x); vx = _mm_unpacklo_epi8(vx, vzero); x += 8; __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp)); __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); vy_lo = _mm_sub_ps(vy_lo, vmagic_bias); vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); vy_lo = _mm_mul_ps(vy_lo, vscale); vy_hi = _mm_mul_ps(vy_hi, vscale); _mm_storeu_ps(y, vy_lo); _mm_storeu_ps(y + 4, vy_hi); y += 8; } if XNN_UNLIKELY(n != 0) { assert(n >= 1 * sizeof(uint8_t)); assert(n <= 7 * sizeof(uint8_t)); __m128i vx = _mm_loadl_epi64((const __m128i*) x); vx = _mm_unpacklo_epi8(vx, vzero); __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp)); vy = _mm_sub_ps(vy, vmagic_bias); vy = _mm_mul_ps(vy, vscale); if (n & (4 * sizeof(uint8_t))) { _mm_storeu_ps(y, vy); vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); vy = _mm_sub_ps(vy, vmagic_bias); vy = _mm_mul_ps(vy, vscale); y += 4; } if (n & (2 * sizeof(uint8_t))) { _mm_storel_pi((__m64*) y, vy); vy = _mm_movehl_ps(vy, vy); y += 2; } if (n & (1 * sizeof(uint8_t))) { _mm_store_ss(y, vy); } } } void xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c8( size_t rows, size_t channels, const uint8_t* input, size_t input_stride, const uint8_t* zero, int32_t* buffer, uint8_t* output, const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(rows > 7); assert(channels != 0); const uint8_t* i0 = input; const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride); const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride); const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride); const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride); const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride); const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride); const size_t input_increment = 7 * input_stride - round_up_po2(channels, 8) * sizeof(uint8_t); const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias); const __m128i vzero = _mm_setzero_si128(); int32_t* b = buffer; size_t c = channels; for (; c != 0; c = doz(c, 8)) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero); vacc0123 = _mm_add_epi32(vacc0123, vinit_bias); vacc4567 = _mm_add_epi32(vacc4567, vinit_bias); _mm_store_si128((__m128i*) b, vacc0123); _mm_store_si128((__m128i*) (b + 4), vacc4567); b += 8; } for (rows -= 7; rows > 7; rows -= 7) { i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment); i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment); i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment); i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment); i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment); i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment); i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment); int32_t* b = buffer; size_t c = channels; for (; c != 0; c = doz(c, 8)) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero); vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) b)); vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (b + 4))); _mm_store_si128((__m128i*) b, vacc0123); _mm_store_si128((__m128i*) (b + 4), vacc4567); b += 8; } } i0 = (const uint8_t*) ((uintptr_t) i0 + input_increment); i1 = (const uint8_t*) ((uintptr_t) i1 + input_increment); if XNN_UNPREDICTABLE(rows < 2) { i1 = zero; } i2 = (const uint8_t*) ((uintptr_t) i2 + input_increment); if XNN_UNPREDICTABLE(rows <= 2) { i2 = zero; } i3 = (const uint8_t*) ((uintptr_t) i3 + input_increment); if XNN_UNPREDICTABLE(rows < 4) { i3 = zero; } i4 = (const uint8_t*) ((uintptr_t) i4 + input_increment); if XNN_UNPREDICTABLE(rows <= 4) { i4 = zero; } i5 = (const uint8_t*) ((uintptr_t) i5 + input_increment); if XNN_UNPREDICTABLE(rows < 6) { i5 = zero; } i6 = (const uint8_t*) ((uintptr_t) i6 + input_increment); if XNN_UNPREDICTABLE(rows <= 6) { i6 = zero; } const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); for (; channels >= 8; channels -= 8) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero); vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer)); vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4))); buffer += 8; __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(channels != 0) { { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero); vacc0123 = _mm_add_epi32(vacc0123, _mm_load_si128((const __m128i*) buffer)); vacc4567 = _mm_add_epi32(vacc4567, _mm_load_si128((const __m128i*) (buffer + 4))); buffer += 8; __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); if (channels & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); if (channels & 2) { unaligned_store_u16(output, (uint16_t) vout0123); vout0123 >>= 16; output += 2; } if (channels & 1) { *output = (uint8_t) vout0123; } } } } void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8( size_t rows, size_t channels, const uint8_t* input, size_t input_stride, const uint8_t* zero, uint8_t* output, const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(rows != 0); assert(rows <= 7); assert(channels != 0); const uint8_t* i0 = input; const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride); if XNN_UNPREDICTABLE(rows < 2) { i1 = zero; } const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride); if XNN_UNPREDICTABLE(rows <= 2) { i2 = zero; } const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride); if XNN_UNPREDICTABLE(rows < 4) { i3 = zero; } const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride); if XNN_UNPREDICTABLE(rows <= 4) { i4 = zero; } const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride); if XNN_UNPREDICTABLE(rows < 6) { i5 = zero; } const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride); if XNN_UNPREDICTABLE(rows <= 6) { i6 = zero; } const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse2.init_bias); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); const __m128i vzero = _mm_setzero_si128(); for (; channels >= 8; channels -= 8) { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero); vacc0123 = _mm_add_epi32(vacc0123, vinit_bias); vacc4567 = _mm_add_epi32(vacc4567, vinit_bias); __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(channels != 0) { { const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, vzero); const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, vzero); const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, vzero); const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567); const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, vzero); const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567); const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, vzero); const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567); const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567); const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, vzero); vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567); __m128i vacc0123 = _mm_unpacklo_epi16(vacc01234567, vzero); __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero); vacc0123 = _mm_add_epi32(vacc0123, vinit_bias); vacc4567 = _mm_add_epi32(vacc4567, vinit_bias); __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point); vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); vacc0123 = _mm_cvtps_epi32(vfpacc0123); vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); if (channels & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } uint32_t vout0123 = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); if (channels & 2) { unaligned_store_u16(output, (uint16_t) vout0123); vout0123 >>= 16; output += 2; } if (channels & 1) { *output = (uint8_t) vout0123; } } } } void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, const uint8_t* restrict a, size_t a_stride, const void* restrict w, uint8_t* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 1); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(uint8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); const uint8_t* a0 = a; uint8_t* c0 = c; do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); w = (const int32_t*) w + 4; size_t k = 0; const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point); const __m128i vzero = _mm_setzero_si128(); while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero); a0 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8)); const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16)); const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24)); const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const void*) ((const uint8_t*) w + 32); k += 8 * sizeof(uint8_t); } const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point); __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride); a0 = (const uint8_t*) ((uintptr_t) a0 - kc); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (uint8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, const uint8_t* restrict a, size_t a_stride, const void* restrict w, uint8_t* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 3); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(uint8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); const uint8_t* a0 = a; uint8_t* c0 = c; const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride); uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { a1 = a0; c1 = c0; } const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride); uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { a2 = a1; c2 = c1; } do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); __m128i vacc1x0 = vacc0x0; __m128i vacc1x1 = vacc0x1; __m128i vacc1x2 = vacc0x2; __m128i vacc1x3 = vacc0x3; __m128i vacc2x0 = vacc0x0; __m128i vacc2x1 = vacc0x1; __m128i vacc2x2 = vacc0x2; __m128i vacc2x3 = vacc0x3; w = (const int32_t*) w + 4; size_t k = 0; const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point); const __m128i vzero = _mm_setzero_si128(); while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero); a0 += 8; const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1); const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero); a1 += 8; const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2); const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero); a2 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0)); vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8)); const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1)); vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16)); const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2)); vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24)); const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3)); vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3)); w = (const void*) ((const uint8_t*) w + 32); k += 8 * sizeof(uint8_t); } const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2)); const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3)); const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2)); const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13)); __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale); vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point); vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123); vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point); __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_si128(vout, 4); unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_si128(vout, 4); unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride); c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride); c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride); a0 = (const uint8_t*) ((uintptr_t) a0 - kc); a1 = (const uint8_t*) ((uintptr_t) a1 - kc); a2 = (const uint8_t*) ((uintptr_t) a2 - kc); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2)); c1 += 2; unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4)); c2 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (uint8_t) _mm_cvtsi128_si32(vout); *c1 = (uint8_t) _mm_extract_epi16(vout, 2); *c2 = (uint8_t) _mm_extract_epi16(vout, 4); } nc = 0; } } while (nc != 0); } void xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, size_t ks, const uint8_t** restrict a, const void* restrict w, uint8_t* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const uint8_t* zero, const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 1); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (1 * sizeof(void*)) == 0); assert(a_offset % sizeof(uint8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); uint8_t* c0 = c; do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); w = (const int32_t*) w + 4; size_t p = ks; do { const uint8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset); } a += 1; size_t k = 0; const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point); const __m128i vzero = _mm_setzero_si128(); while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero); a0 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8)); const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16)); const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24)); const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const void*) ((const uint8_t*) w + 32); k += 8 * sizeof(uint8_t); } p -= 1 * sizeof(void*); } while (p != 0); const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point); __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); if (nc >= 4) { unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride); a = (const uint8_t**restrict) ((uintptr_t) a - ks); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c0 = (uint8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64( size_t mr, size_t nc, size_t kc, size_t ks, const uint8_t** restrict a, const void* restrict w, uint8_t* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const uint8_t* zero, const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 3); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (3 * sizeof(void*)) == 0); assert(a_offset % sizeof(uint8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8); uint8_t* c0 = c; uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { c1 = c0; } uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { c2 = c1; } do { __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]); __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]); __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]); __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]); __m128i vacc1x0 = vacc0x0; __m128i vacc1x1 = vacc0x1; __m128i vacc1x2 = vacc0x2; __m128i vacc1x3 = vacc0x3; __m128i vacc2x0 = vacc0x0; __m128i vacc2x1 = vacc0x1; __m128i vacc2x2 = vacc0x2; __m128i vacc2x3 = vacc0x3; w = (const int32_t*) w + 4; size_t p = ks; do { const uint8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset); } const uint8_t* restrict a1 = a[1]; if XNN_UNPREDICTABLE(a1 != zero) { a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset); } const uint8_t* restrict a2 = a[2]; if XNN_UNPREDICTABLE(a2 != zero) { a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset); } a += 3; size_t k = 0; const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point); const __m128i vzero = _mm_setzero_si128(); while (k < kc) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero); a0 += 8; const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1); const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero); a1 += 8; const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2); const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero); a2 += 8; const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0)); vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0)); const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8)); const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1)); vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1)); const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16)); const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2)); vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2)); const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24)); const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3)); vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3)); w = (const void*) ((const uint8_t*) w + 32); k += 8 * sizeof(uint8_t); } p -= 3 * sizeof(void*); } while (p != 0); const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2)); const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3)); const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2)); const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13)); __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13)); __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale); vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale); vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale); const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point); vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point); vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point); vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point); vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123); vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123); vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point); __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point); __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123); vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min)); if (nc >= 4) { unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2)))); c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride); unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)))); c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride); unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout)); c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride); a = (const uint8_t**restrict) ((uintptr_t) a - ks); nc -= 4; } else { if (nc & 2) { unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4)); c2 += 2; unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2)); c1 += 2; unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0)); c0 += 2; vout = _mm_srli_epi32(vout, 16); } if (nc & 1) { *c2 = (uint8_t) _mm_extract_epi16(vout, 4); *c1 = (uint8_t) _mm_extract_epi16(vout, 2); *c0 = (uint8_t) _mm_cvtsi128_si32(vout); } nc = 0; } } while (nc != 0); } void xnn_qu8_vadd_minmax_ukernel__sse2_mul16_ld64_x8( size_t n, const uint8_t* input_a, const uint8_t* input_b, uint8_t* output, const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias); const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo); const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi); const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo); const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_hi); const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); input_a += 8; input_b += 8; const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi)); vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); if (n & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(uint8_t))) { *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8( size_t n, const uint8_t* input_a, const uint8_t* input_b, uint8_t* output, const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i vbias = _mm_add_epi32( _mm_shuffle_epi32(_mm_cvtsi32_si128(params->sse2.b_multiplier * (int32_t) *input_b), _MM_SHUFFLE(0, 0, 0, 0)), _mm_load_si128((const __m128i*) params->sse2.bias)); const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo); const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi); const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max); for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); input_a += 8; const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); va01234567 = _mm_unpacklo_epi8(va01234567, _mm_setzero_si128()); __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo); const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo); vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi)); __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi)); __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi)); vacc0123 = _mm_sra_epi32(vacc0123, vshift); vacc4567 = _mm_sra_epi32(vacc4567, vshift); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); if (n & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(uint8_t))) { *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qu8_vcvt_ukernel__sse2_x32( size_t n, const uint8_t* x, uint8_t* y, const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(uint8_t) == 0); assert(x != NULL); assert(y != NULL); const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier); const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias); const __m128i vzero = _mm_setzero_si128(); for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) { const __m128i vx0 = _mm_loadu_si128((const __m128i*) x); const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16)); x += 32; const __m128i vextx0 = _mm_unpacklo_epi8(vx0, vzero); const __m128i vextx1 = _mm_unpackhi_epi8(vx0, vzero); const __m128i vextx2 = _mm_unpacklo_epi8(vx1, vzero); const __m128i vextx3 = _mm_unpackhi_epi8(vx1, vzero); const __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier); const __m128i vprodhi0 = _mm_mulhi_epu16(vextx0, vmultiplier); const __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier); const __m128i vprodhi1 = _mm_mulhi_epu16(vextx1, vmultiplier); const __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier); const __m128i vprodhi2 = _mm_mulhi_epu16(vextx2, vmultiplier); const __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier); const __m128i vprodhi3 = _mm_mulhi_epu16(vextx3, vmultiplier); __m128i vacc0 = _mm_unpacklo_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_unpackhi_epi16(vprodlo0, vprodhi0); __m128i vacc2 = _mm_unpacklo_epi16(vprodlo1, vprodhi1); __m128i vacc3 = _mm_unpackhi_epi16(vprodlo1, vprodhi1); __m128i vacc4 = _mm_unpacklo_epi16(vprodlo2, vprodhi2); __m128i vacc5 = _mm_unpackhi_epi16(vprodlo2, vprodhi2); __m128i vacc6 = _mm_unpacklo_epi16(vprodlo3, vprodhi3); __m128i vacc7 = _mm_unpackhi_epi16(vprodlo3, vprodhi3); vacc0 = _mm_add_epi32(vacc0, vbias); vacc1 = _mm_add_epi32(vacc1, vbias); vacc2 = _mm_add_epi32(vacc2, vbias); vacc3 = _mm_add_epi32(vacc3, vbias); vacc4 = _mm_add_epi32(vacc4, vbias); vacc5 = _mm_add_epi32(vacc5, vbias); vacc6 = _mm_add_epi32(vacc6, vbias); vacc7 = _mm_add_epi32(vacc7, vbias); vacc0 = _mm_srai_epi32(vacc0, 8); vacc1 = _mm_srai_epi32(vacc1, 8); vacc2 = _mm_srai_epi32(vacc2, 8); vacc3 = _mm_srai_epi32(vacc3, 8); vacc4 = _mm_srai_epi32(vacc4, 8); vacc5 = _mm_srai_epi32(vacc5, 8); vacc6 = _mm_srai_epi32(vacc6, 8); vacc7 = _mm_srai_epi32(vacc7, 8); vacc0 = _mm_packs_epi32(vacc0, vacc1); vacc1 = _mm_packs_epi32(vacc2, vacc3); vacc2 = _mm_packs_epi32(vacc4, vacc5); vacc3 = _mm_packs_epi32(vacc6, vacc7); const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1); const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3); _mm_storeu_si128((__m128i*) y, vy0); _mm_storeu_si128((__m128i*) (y + 16), vy1); y += 32; } for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vzero); const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vzero); const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier); const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier); const __m128i vprodhi_lo = _mm_mulhi_epu16(vextx_lo, vmultiplier); const __m128i vprodhi_hi = _mm_mulhi_epu16(vextx_hi, vmultiplier); __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi); __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi); vacc_ll = _mm_add_epi32(vacc_ll, vbias); vacc_lh = _mm_add_epi32(vacc_lh, vbias); vacc_hl = _mm_add_epi32(vacc_hl, vbias); vacc_hh = _mm_add_epi32(vacc_hh, vbias); vacc_ll = _mm_srai_epi32(vacc_ll, 8); vacc_lh = _mm_srai_epi32(vacc_lh, 8); vacc_hl = _mm_srai_epi32(vacc_hl, 8); vacc_hh = _mm_srai_epi32(vacc_hh, 8); const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh); const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh); const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi); _mm_storeu_si128((__m128i*) y, vy); y += 16; } if XNN_UNLIKELY(n != 0) { assert(n >= 1 * sizeof(uint8_t)); assert(n <= 15 * sizeof(uint8_t)); const __m128i vx = _mm_loadu_si128((const __m128i*) x); const __m128i vextx_lo = _mm_unpacklo_epi8(vx, vzero); const __m128i vextx_hi = _mm_unpackhi_epi8(vx, vzero); const __m128i vprodlo_lo = _mm_mullo_epi16(vextx_lo, vmultiplier); const __m128i vprodlo_hi = _mm_mullo_epi16(vextx_hi, vmultiplier); const __m128i vprodhi_lo = _mm_mulhi_epu16(vextx_lo, vmultiplier); const __m128i vprodhi_hi = _mm_mulhi_epu16(vextx_hi, vmultiplier); __m128i vacc_ll = _mm_unpacklo_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_lh = _mm_unpackhi_epi16(vprodlo_lo, vprodhi_lo); __m128i vacc_hl = _mm_unpacklo_epi16(vprodlo_hi, vprodhi_hi); __m128i vacc_hh = _mm_unpackhi_epi16(vprodlo_hi, vprodhi_hi); vacc_ll = _mm_add_epi32(vacc_ll, vbias); vacc_lh = _mm_add_epi32(vacc_lh, vbias); vacc_hl = _mm_add_epi32(vacc_hl, vbias); vacc_hh = _mm_add_epi32(vacc_hh, vbias); vacc_ll = _mm_srai_epi32(vacc_ll, 8); vacc_lh = _mm_srai_epi32(vacc_lh, 8); vacc_hl = _mm_srai_epi32(vacc_hl, 8); vacc_hh = _mm_srai_epi32(vacc_hh, 8); const __m128i vacc_lo = _mm_packs_epi32(vacc_ll, vacc_lh); const __m128i vacc_hi = _mm_packs_epi32(vacc_hl, vacc_hh); __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi); if (n & (8 * sizeof(uint8_t))) { _mm_storel_epi64((__m128i*) y, vy); vy = _mm_unpackhi_epi64(vy, vy); y += 8; } if (n & (4 * sizeof(uint8_t))) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy)); vy = _mm_srli_epi64(vy, 32); y += 4; } uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy); if (n & (2 * sizeof(uint8_t))) { unaligned_store_u16(y, (uint16_t) vy_lo); vy_lo >>= 16; y += 2; } if (n & (1 * sizeof(uint8_t))) { *y = (uint8_t) vy_lo; } } } void xnn_qu8_vlrelu_ukernel__sse2_x32( size_t n, const uint8_t* x, uint8_t* y, const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); assert(n % sizeof(uint8_t) == 0); assert(x != NULL); assert(y != NULL); const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point); const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff); const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); const __m128i vzero = _mm_setzero_si128(); for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) { const __m128i vx0 = _mm_loadu_si128((const __m128i*) x); const __m128i vx1 = _mm_loadu_si128((const __m128i*) (x + 16)); x += 32; __m128i vextx0 = _mm_unpacklo_epi8(vx0, vzero); __m128i vextx1 = _mm_unpackhi_epi8(vx0, vzero); __m128i vextx2 = _mm_unpacklo_epi8(vx1, vzero); __m128i vextx3 = _mm_unpackhi_epi8(vx1, vzero); __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point); vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0); __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point); vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1); __m128i vmultiplier2 = _mm_cmpgt_epi16(vextx2, vinput_zero_point); vextx2 = _mm_sub_epi16(vinput_zero_point, vextx2); __m128i vmultiplier3 = _mm_cmpgt_epi16(vextx3, vinput_zero_point); vextx3 = _mm_sub_epi16(vinput_zero_point, vextx3); vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff); vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff); vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff); vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff); vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base); vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base); vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base); vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base); __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0); __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1); __m128i vprodlo2 = _mm_mullo_epi16(vextx2, vmultiplier2); __m128i vprodlo3 = _mm_mullo_epi16(vextx3, vmultiplier3); vprodlo0 = _mm_srli_epi16(vprodlo0, 7); __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0); vprodlo1 = _mm_srli_epi16(vprodlo1, 7); __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1); vprodlo2 = _mm_srli_epi16(vprodlo2, 7); __m128i vprodhi2 = _mm_mulhi_epi16(vextx2, vmultiplier2); vprodlo3 = _mm_srli_epi16(vprodlo3, 7); __m128i vprodhi3 = _mm_mulhi_epi16(vextx3, vmultiplier3); vprodhi0 = _mm_slli_epi16(vprodhi0, 8); vprodlo0 = _mm_avg_epu16(vprodlo0, vzero); vprodhi1 = _mm_slli_epi16(vprodhi1, 8); vprodlo1 = _mm_avg_epu16(vprodlo1, vzero); vprodhi2 = _mm_slli_epi16(vprodhi2, 8); vprodlo2 = _mm_avg_epu16(vprodlo2, vzero); vprodhi3 = _mm_slli_epi16(vprodhi3, 8); vprodlo3 = _mm_avg_epu16(vprodlo3, vzero); __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1); __m128i vacc2 = _mm_add_epi16(vprodlo2, vprodhi2); __m128i vacc3 = _mm_add_epi16(vprodlo3, vprodhi3); vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point); vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point); const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1); const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3); _mm_storeu_si128((__m128i*) y, vy0); _mm_storeu_si128((__m128i*) (y + 16), vy1); y += 32; } for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; __m128i vextx0 = _mm_unpacklo_epi8(vx, vzero); __m128i vextx1 = _mm_unpackhi_epi8(vx, vzero); __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point); __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point); vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0); vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1); vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff); vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff); vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base); vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base); __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0); __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1); vprodlo0 = _mm_srli_epi16(vprodlo0, 7); vprodlo1 = _mm_srli_epi16(vprodlo1, 7); __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0); __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1); vprodhi0 = _mm_slli_epi16(vprodhi0, 8); vprodhi1 = _mm_slli_epi16(vprodhi1, 8); vprodlo0 = _mm_avg_epu16(vprodlo0, vzero); vprodlo1 = _mm_avg_epu16(vprodlo1, vzero); __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1); vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); const __m128i vy = _mm_packus_epi16(vacc0, vacc1); _mm_storeu_si128((__m128i*) y, vy); y += 16; } if XNN_UNLIKELY(n != 0) { assert(n >= 1 * sizeof(uint8_t)); assert(n <= 15 * sizeof(uint8_t)); const __m128i vx = _mm_loadu_si128((const __m128i*) x); __m128i vextx0 = _mm_unpacklo_epi8(vx, vzero); __m128i vextx1 = _mm_unpackhi_epi8(vx, vzero); __m128i vmultiplier0 = _mm_cmpgt_epi16(vextx0, vinput_zero_point); __m128i vmultiplier1 = _mm_cmpgt_epi16(vextx1, vinput_zero_point); vextx0 = _mm_sub_epi16(vinput_zero_point, vextx0); vextx1 = _mm_sub_epi16(vinput_zero_point, vextx1); vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff); vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff); vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base); vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base); __m128i vprodlo0 = _mm_mullo_epi16(vextx0, vmultiplier0); __m128i vprodlo1 = _mm_mullo_epi16(vextx1, vmultiplier1); vprodlo0 = _mm_srli_epi16(vprodlo0, 7); vprodlo1 = _mm_srli_epi16(vprodlo1, 7); __m128i vprodhi0 = _mm_mulhi_epi16(vextx0, vmultiplier0); __m128i vprodhi1 = _mm_mulhi_epi16(vextx1, vmultiplier1); vprodhi0 = _mm_slli_epi16(vprodhi0, 8); vprodhi1 = _mm_slli_epi16(vprodhi1, 8); vprodlo0 = _mm_avg_epu16(vprodlo0, vzero); vprodlo1 = _mm_avg_epu16(vprodlo1, vzero); __m128i vacc0 = _mm_add_epi16(vprodlo0, vprodhi0); __m128i vacc1 = _mm_add_epi16(vprodlo1, vprodhi1); vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); __m128i vy = _mm_packus_epi16(vacc0, vacc1); if (n & (8 * sizeof(uint8_t))) { _mm_storel_epi64((__m128i*) y, vy); vy = _mm_unpackhi_epi64(vy, vy); y += 8; } if (n & (4 * sizeof(uint8_t))) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy)); vy = _mm_srli_epi64(vy, 32); y += 4; } uint32_t vy0 = (uint32_t) _mm_cvtsi128_si32(vy); if (n & (2 * sizeof(uint8_t))) { unaligned_store_u16(y, (uint16_t) vy0); vy0 >>= 16; y += 2; } if (n & (1 * sizeof(uint8_t))) { *y = (uint8_t) vy0; } } } void xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8( size_t n, const uint8_t* input_a, const uint8_t* input_b, uint8_t* output, const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point); const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max); for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); input_a += 8; input_b += 8; const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b); const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vxb01234567 = _mm_sub_epi16(vb01234567, vb_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb01234567); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb01234567); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); if (n & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(uint8_t))) { *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8( size_t n, const uint8_t* input_a, const uint8_t* input_b, uint8_t* output, const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point); const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale); const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max); __m128i vxb = _mm_sub_epi16( _mm_shuffle_epi32(_mm_cvtsi32_si128(UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b), 0), _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point)); for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); input_a += 8; const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); _mm_storel_epi64((__m128i*) output, vout0123456701234567); output += 8; } if XNN_UNLIKELY(n != 0) { { __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a); const __m128i vzero = _mm_setzero_si128(); va01234567 = _mm_unpacklo_epi8(va01234567, vzero); const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point); const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb); const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb); const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi); const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi); __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123); __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale); vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123); const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); if (n & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); output += 4; } if (n & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567)); vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); output += 2; } if (n & (1 * sizeof(uint8_t))) { *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567); } } } } void xnn_s8_ibilinear_ukernel__sse2_c8( size_t output_pixels, size_t channels, const int8_t**restrict input, size_t input_offset, const int16_t*restrict weights, int8_t*restrict output, size_t output_increment) XNN_OOB_READS { assert(output_pixels != 0); assert(channels != 0); do { const int8_t* i0 = (const int8_t*) ((uintptr_t) input[0] + input_offset); const int8_t* i1 = (const int8_t*) ((uintptr_t) input[1] + input_offset); const int8_t* i2 = (const int8_t*) ((uintptr_t) input[2] + input_offset); const int8_t* i3 = (const int8_t*) ((uintptr_t) input[3] + input_offset); input += 4; const __m128i valpha = _mm_cvtsi32_si128(*((const int*) weights)); weights += 2; __m128i valphah = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(0, 0, 0, 0)); valphah = _mm_unpacklo_epi64(valphah, valphah); __m128i valphav = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(1, 1, 1, 1)); valphav = _mm_unpacklo_epi64(valphav, valphav); valphah = _mm_xor_si128(valphah, _mm_set1_epi32(0xFFFF0000)); valphah = _mm_add_epi16(valphah, _mm_set1_epi32(0x08010000)); const __m128i vrounding = _mm_set1_epi32(0x00200000); size_t c = channels; for (; c >= 8 * sizeof(int8_t); c -= 8 * sizeof(int8_t)) { __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; vtl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtl01234567, vtl01234567), 8); vtr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtr01234567, vtr01234567), 8); vbl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbl01234567, vbl01234567), 8); vbr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbr01234567, vbr01234567), 8); const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567); const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567); const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah); const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah); const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah); __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16); __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16); vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123); vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567); vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123); vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567); vacc0123 = _mm_srai_epi32(_mm_add_epi16(vacc0123, vrounding), 22); vacc4567 = _mm_srai_epi32(_mm_add_epi16(vacc4567, vrounding), 22); const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567); const __m128i vo01234567 = _mm_packs_epi16(vacc01234567, vacc01234567); _mm_storel_epi64((__m128i*) output, vo01234567); output += 8; } if XNN_UNLIKELY(c != 0) { __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0); __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1); __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2); __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3); vtl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtl01234567, vtl01234567), 8); vtr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vtr01234567, vtr01234567), 8); vbl01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbl01234567, vbl01234567), 8); vbr01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vbr01234567, vbr01234567), 8); const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567); const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567); const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah); const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah); const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah); __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16); __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16); vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123); vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567); vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123); vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567); vacc0123 = _mm_srai_epi32(_mm_add_epi16(vacc0123, vrounding), 22); vacc4567 = _mm_srai_epi32(_mm_add_epi16(vacc4567, vrounding), 22); const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567); __m128i vo01234567 = _mm_packs_epi16(vacc01234567, vacc01234567); if (c & (4 * sizeof(int8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vo01234567)); output += 4; vo01234567 = _mm_srli_epi64(vo01234567, 32); } uint32_t vo0123 = (uint32_t) _mm_cvtsi128_si32(vo01234567); if (c & (2 * sizeof(int8_t))) { unaligned_store_u16(output, (uint16_t) vo0123); output += 2; vo0123 >>= 16; } if (c & (1 * sizeof(int8_t))) { *output++ = (uint8_t) vo0123; } } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_pixels != 0); } void xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16( size_t output_pixels, size_t kernel_elements, size_t channels, const int8_t** input, size_t input_offset, int8_t* output, size_t input_increment, size_t output_increment, const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(output_pixels != 0); assert(kernel_elements != 0); assert(channels != 0); const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias); const __m128i voutput_max_with_bias = _mm_load_si128((const __m128i*) params->sse2.max_with_bias); const __m128i voutput_min_with_bias = _mm_load_si128((const __m128i*) params->sse2.min_with_bias); do { int8_t* o = output; { const int8_t* i0 = *input++; const int8_t* i1 = *input++; const int8_t* i2 = *input++; const int8_t* i3 = *input++; const int8_t* i4 = *input++; const int8_t* i5 = *input++; const int8_t* i6 = *input++; const int8_t* i7 = *input++; const int8_t* i8 = *input++; i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); if (kernel_elements < 2) { i1 = i0; } if (kernel_elements <= 2) { i2 = i0; } if (kernel_elements < 4) { i3 = i0; } if (kernel_elements <= 4) { i4 = i0; } if (kernel_elements < 6) { i5 = i0; } if (kernel_elements <= 6) { i6 = i0; } if (kernel_elements < 8) { i7 = i0; } if (kernel_elements <= 8) { i8 = i0; } size_t c = channels; for (; c >= 16; c -= 16) { const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias); i0 += 16; const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias); i1 += 16; const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); i2 += 16; const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); i3 += 16; const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias); i4 += 16; const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias); i5 += 16; const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias); i6 += 16; const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias); i7 += 16; const __m128i vi8 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i8), vbias); i8 += 16; const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax01678); vout = _mm_max_epu8(vout, voutput_min_with_bias); vout = _mm_min_epu8(vout, voutput_max_with_bias); vout = _mm_xor_si128(vout, vbias); _mm_storeu_si128((__m128i*) o, vout); o += 16; } if (c != 0) { const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias); const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias); const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias); const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias); const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias); const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias); const __m128i vi8 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i8), vbias); const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax01678); vout = _mm_max_epu8(vout, voutput_min_with_bias); vout = _mm_min_epu8(vout, voutput_max_with_bias); vout = _mm_xor_si128(vout, vbias); if (c & 8) { _mm_storel_epi64((__m128i*) o, vout); vout = _mm_unpackhi_epi64(vout, vout); o += 8; } if (c & 4) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_epi64(vout, 32); o += 4; } if (c & 2) { unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0)); vout = _mm_srli_epi32(vout, 16); o += 2; } if (c & 1) { *((int8_t*) o) = (int8_t) _mm_cvtsi128_si32(vout); o += 1; } } } for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) { const int8_t* i0 = *input++; const int8_t* i1 = *input++; const int8_t* i2 = *input++; const int8_t* i3 = *input++; const int8_t* i4 = *input++; const int8_t* i5 = *input++; const int8_t* i6 = *input++; const int8_t* i7 = *input++; i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); if (k < 2) { i1 = i0; } if (k <= 2) { i2 = i0; } if (k < 4) { i3 = i0; } if (k <= 4) { i4 = i0; } if (k < 6) { i5 = i0; } if (k <= 6) { i6 = i0; } if (k < 8) { i7 = i0; } o = output; size_t c = channels; for (; c >= 16; c -= 16) { const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias); i0 += 16; const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias); i1 += 16; const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); i2 += 16; const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); i3 += 16; const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias); i4 += 16; const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias); i5 += 16; const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias); i6 += 16; const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias); i7 += 16; const __m128i vo = _mm_xor_si128(_mm_loadu_si128((const __m128i*) o), vbias); const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax0167); vout = _mm_max_epu8(vout, voutput_min_with_bias); vout = _mm_min_epu8(vout, voutput_max_with_bias); vout = _mm_xor_si128(vout, vbias); _mm_storeu_si128((__m128i*) o, vout); o += 16; } if (c != 0) { const __m128i vi0 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i0), vbias); const __m128i vi1 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i1), vbias); const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); const __m128i vi4 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i4), vbias); const __m128i vi5 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i5), vbias); const __m128i vi6 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i6), vbias); const __m128i vi7 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i7), vbias); const __m128i vo = _mm_xor_si128(_mm_loadu_si128((const __m128i*) o), vbias); const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax0167); vout = _mm_max_epu8(vout, voutput_min_with_bias); vout = _mm_min_epu8(vout, voutput_max_with_bias); vout = _mm_xor_si128(vout, vbias); if (c & 8) { _mm_storel_epi64((__m128i*) o, vout); vout = _mm_unpackhi_epi64(vout, vout); o += 8; } if (c & 4) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_epi64(vout, 32); o += 4; } if (c & 2) { unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0)); vout = _mm_srli_epi32(vout, 16); o += 2; } if (c & 1) { *o = (int8_t) _mm_cvtsi128_si32(vout); o += 1; } } } input = (const int8_t**) ((uintptr_t) input + input_increment); output = (int8_t*) ((uintptr_t) o + output_increment); } while (--output_pixels != 0); } void xnn_s8_vclamp_ukernel__sse2_x64( size_t n, const int8_t* x, int8_t* y, const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias); const __m128i voutput_max_with_bias = _mm_load_si128((const __m128i*) params->sse2.max_with_bias); const __m128i voutput_min_with_bias = _mm_load_si128((const __m128i*) params->sse2.min_with_bias); for (; n >= 64; n -= 64) { __m128i vacc0 = _mm_loadu_si128((const __m128i*) x); __m128i vacc1 = _mm_loadu_si128((const __m128i*) x + 1); __m128i vacc2 = _mm_loadu_si128((const __m128i*) x + 2); __m128i vacc3 = _mm_loadu_si128((const __m128i*) x + 3); x += 64; vacc0 = _mm_xor_si128(vacc0, vbias); vacc1 = _mm_xor_si128(vacc1, vbias); vacc2 = _mm_xor_si128(vacc2, vbias); vacc3 = _mm_xor_si128(vacc3, vbias); vacc0 = _mm_max_epu8(vacc0, voutput_min_with_bias); vacc1 = _mm_max_epu8(vacc1, voutput_min_with_bias); vacc2 = _mm_max_epu8(vacc2, voutput_min_with_bias); vacc3 = _mm_max_epu8(vacc3, voutput_min_with_bias); vacc0 = _mm_min_epu8(vacc0, voutput_max_with_bias); vacc1 = _mm_min_epu8(vacc1, voutput_max_with_bias); vacc2 = _mm_min_epu8(vacc2, voutput_max_with_bias); vacc3 = _mm_min_epu8(vacc3, voutput_max_with_bias); vacc0 = _mm_xor_si128(vacc0, vbias); vacc1 = _mm_xor_si128(vacc1, vbias); vacc2 = _mm_xor_si128(vacc2, vbias); vacc3 = _mm_xor_si128(vacc3, vbias); _mm_storeu_si128((__m128i*) y, vacc0); _mm_storeu_si128((__m128i*) y + 1, vacc1); _mm_storeu_si128((__m128i*) y + 2, vacc2); _mm_storeu_si128((__m128i*) y + 3, vacc3); y += 64; } for (; n >= 16; n -= 16) { __m128i vacc = _mm_loadu_si128((const __m128i*) x); x += 16; vacc = _mm_xor_si128(vacc, vbias); vacc = _mm_min_epu8(vacc, voutput_max_with_bias); vacc = _mm_max_epu8(vacc, voutput_min_with_bias); vacc = _mm_xor_si128(vacc, vbias); _mm_storeu_si128((__m128i*) y, vacc); y += 16; } if XNN_UNLIKELY(n != 0) { __m128i vacc = _mm_loadu_si128((const __m128i*) x); vacc = _mm_xor_si128(vacc, vbias); vacc = _mm_min_epu8(vacc, voutput_max_with_bias); vacc = _mm_max_epu8(vacc, voutput_min_with_bias); vacc = _mm_xor_si128(vacc, vbias); if (n & 8) { _mm_storel_epi64((__m128i*) y, vacc); y += 8; vacc = _mm_unpackhi_epi64(vacc, vacc); } if (n & 4) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vacc)); y += 4; vacc = _mm_srli_epi64(vacc, 32); } if (n & 2) { unaligned_store_u16(y, (uint16_t) _mm_cvtsi128_si32(vacc)); y += 2; vacc = _mm_srli_epi32(vacc, 16); } if (n & 1) { *y = (int8_t) _mm_cvtsi128_si32(vacc); } } } void xnn_u8_ibilinear_ukernel__sse2_c8( size_t output_pixels, size_t channels, const uint8_t**restrict input, size_t input_offset, const int16_t*restrict weights, uint8_t*restrict output, size_t output_increment) XNN_OOB_READS { assert(output_pixels != 0); assert(channels != 0); do { const uint8_t* i0 = (const uint8_t*) ((uintptr_t) input[0] + input_offset); const uint8_t* i1 = (const uint8_t*) ((uintptr_t) input[1] + input_offset); const uint8_t* i2 = (const uint8_t*) ((uintptr_t) input[2] + input_offset); const uint8_t* i3 = (const uint8_t*) ((uintptr_t) input[3] + input_offset); input += 4; const __m128i valpha = _mm_cvtsi32_si128(*((const int*) weights)); weights += 2; __m128i valphah = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(0, 0, 0, 0)); valphah = _mm_unpacklo_epi64(valphah, valphah); __m128i valphav = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(1, 1, 1, 1)); valphav = _mm_unpacklo_epi64(valphav, valphav); valphah = _mm_xor_si128(valphah, _mm_set1_epi32(0xFFFF0000)); valphah = _mm_add_epi16(valphah, _mm_set1_epi32(0x08010000)); const __m128i vrounding = _mm_set1_epi32(0x00200000); size_t c = channels; for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) { __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; __m128i vzero = _mm_setzero_si128(); vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero); vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero); vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero); vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero); const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567); const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567); const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah); const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah); const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah); __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16); __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16); vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123); vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567); vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123); vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567); vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22); vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22); const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567); const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567); _mm_storel_epi64((__m128i*) output, vo01234567); output += 8; } if XNN_UNLIKELY(c != 0) { __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0); __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1); __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2); __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3); __m128i vzero = _mm_setzero_si128(); vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero); vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero); vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero); vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero); const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567); const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah); const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567); const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah); const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah); const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah); __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16); __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16); vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123); vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567); vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123); vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567); vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22); vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22); const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567); __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567); if (c & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vo01234567)); output += 4; vo01234567 = _mm_srli_epi64(vo01234567, 32); } uint32_t vo0123 = (uint32_t) _mm_cvtsi128_si32(vo01234567); if (c & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) vo0123); output += 2; vo0123 >>= 16; } if (c & (1 * sizeof(uint8_t))) { *output++ = (uint8_t) vo0123; } } output = (uint8_t*) ((uintptr_t) output + output_increment); } while (--output_pixels != 0); } void xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16( size_t output_pixels, size_t kernel_elements, size_t channels, const uint8_t** input, size_t input_offset, uint8_t* output, size_t input_increment, size_t output_increment, const union xnn_u8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(output_pixels != 0); assert(kernel_elements != 0); assert(channels != 0); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.max); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.min); do { uint8_t* o = output; { const uint8_t* i0 = *input++; const uint8_t* i1 = *input++; const uint8_t* i2 = *input++; const uint8_t* i3 = *input++; const uint8_t* i4 = *input++; const uint8_t* i5 = *input++; const uint8_t* i6 = *input++; const uint8_t* i7 = *input++; const uint8_t* i8 = *input++; i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset); if (kernel_elements < 2) { i1 = i0; } if (kernel_elements <= 2) { i2 = i0; } if (kernel_elements < 4) { i3 = i0; } if (kernel_elements <= 4) { i4 = i0; } if (kernel_elements < 6) { i5 = i0; } if (kernel_elements <= 6) { i6 = i0; } if (kernel_elements < 8) { i7 = i0; } if (kernel_elements <= 8) { i8 = i0; } size_t c = channels; for (; c >= 16; c -= 16) { const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16; const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16; const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16; const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16; const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16; const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16; const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16; const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16; const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8); i8 += 16; const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax01678); vout = _mm_max_epu8(vout, voutput_min); vout = _mm_min_epu8(vout, voutput_max); _mm_storeu_si128((__m128i*) o, vout); o += 16; } if (c != 0) { const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); const __m128i vi8 = _mm_loadu_si128((const __m128i*) i8); const __m128i vmax018 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vi8); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax01678 = _mm_max_epu8(vmax018, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax01678); vout = _mm_max_epu8(vout, voutput_min); vout = _mm_min_epu8(vout, voutput_max); if (c & 8) { _mm_storel_epi64((__m128i*) o, vout); vout = _mm_unpackhi_epi64(vout, vout); o += 8; } if (c & 4) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_epi64(vout, 32); o += 4; } if (c & 2) { unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0)); vout = _mm_srli_epi32(vout, 16); o += 2; } if (c & 1) { *o = (uint8_t) _mm_cvtsi128_si32(vout); o += 1; } } } for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) { const uint8_t* i0 = *input++; const uint8_t* i1 = *input++; const uint8_t* i2 = *input++; const uint8_t* i3 = *input++; const uint8_t* i4 = *input++; const uint8_t* i5 = *input++; const uint8_t* i6 = *input++; const uint8_t* i7 = *input++; i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset); i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset); i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset); i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset); i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset); i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset); i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset); i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset); if (k < 2) { i1 = i0; } if (k <= 2) { i2 = i0; } if (k < 4) { i3 = i0; } if (k <= 4) { i4 = i0; } if (k < 6) { i5 = i0; } if (k <= 6) { i6 = i0; } if (k < 8) { i7 = i0; } o = output; size_t c = channels; for (; c >= 16; c -= 16) { const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); i0 += 16; const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); i1 += 16; const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16; const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16; const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); i4 += 16; const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); i5 += 16; const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 16; const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16; const __m128i vo = _mm_loadu_si128((const __m128i*) o); const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax0167); vout = _mm_max_epu8(vout, voutput_min); vout = _mm_min_epu8(vout, voutput_max); _mm_storeu_si128((__m128i*) o, vout); o += 16; } if (c != 0) { const __m128i vi0 = _mm_loadu_si128((const __m128i*) i0); const __m128i vi1 = _mm_loadu_si128((const __m128i*) i1); const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); const __m128i vi4 = _mm_loadu_si128((const __m128i*) i4); const __m128i vi5 = _mm_loadu_si128((const __m128i*) i5); const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); const __m128i vo = _mm_loadu_si128((const __m128i*) o); const __m128i vmax01 = _mm_max_epu8(_mm_max_epu8(vi0, vi1), vo); const __m128i vmax23 = _mm_max_epu8(vi2, vi3); const __m128i vmax45 = _mm_max_epu8(vi4, vi5); const __m128i vmax67 = _mm_max_epu8(vi6, vi7); const __m128i vmax2345 = _mm_max_epu8(vmax23, vmax45); const __m128i vmax0167 = _mm_max_epu8(vmax01, vmax67); __m128i vout = _mm_max_epu8(vmax2345, vmax0167); vout = _mm_max_epu8(vout, voutput_min); vout = _mm_min_epu8(vout, voutput_max); if (c & 8) { _mm_storel_epi64((__m128i*) o, vout); vout = _mm_unpackhi_epi64(vout, vout); o += 8; } if (c & 4) { unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vout)); vout = _mm_srli_epi64(vout, 32); o += 4; } if (c & 2) { unaligned_store_u16(o, (uint16_t) _mm_extract_epi16(vout, 0)); vout = _mm_srli_epi32(vout, 16); o += 2; } if (c & 1) { *o = (uint8_t) _mm_cvtsi128_si32(vout); o += 1; } } } input = (const uint8_t**) ((uintptr_t) input + input_increment); output = (uint8_t*) ((uintptr_t) o + output_increment); } while (--output_pixels != 0); } void xnn_u8_rmax_ukernel__sse2( size_t n, const uint8_t* x, uint8_t* y) { assert(n != 0); if XNN_LIKELY(n >= 16) { __m128i vmax = _mm_setzero_si128(); do { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; vmax = _mm_max_epu8(vmax, vx); n -= 16; } while (n >= 16); if (n != 0) { const size_t x_increment = n - 16; x = (const uint8_t*) ((uintptr_t) x + x_increment); const __m128i vx = _mm_loadu_si128((const __m128i*) x); vmax = _mm_max_epu8(vmax, vx); } vmax = _mm_max_epu8(vmax, _mm_unpackhi_epi64(vmax, vmax)); vmax = _mm_max_epu8(vmax, _mm_srli_epi64(vmax, 32)); vmax = _mm_max_epu8(vmax, _mm_srli_epi32(vmax, 16)); vmax = _mm_max_epu8(vmax, _mm_srli_epi16(vmax, 8)); *y = (uint8_t) _mm_cvtsi128_si32(vmax); } else { uint8_t vmax = 0; do { const uint8_t vx = *x++; vmax = vx > vmax ? vx : vmax; } while (--n != 0); *y = vmax; } } void xnn_u8_vclamp_ukernel__sse2_x64( size_t n, const uint8_t* x, uint8_t* y, const union xnn_u8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(n != 0); const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.max); const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.min); for (; n >= 64; n -= 64) { __m128i vacc0 = _mm_loadu_si128((const __m128i*) x); __m128i vacc1 = _mm_loadu_si128((const __m128i*) x + 1); __m128i vacc2 = _mm_loadu_si128((const __m128i*) x + 2); __m128i vacc3 = _mm_loadu_si128((const __m128i*) x + 3); x += 64; vacc0 = _mm_max_epu8(vacc0, voutput_min); vacc1 = _mm_max_epu8(vacc1, voutput_min); vacc2 = _mm_max_epu8(vacc2, voutput_min); vacc3 = _mm_max_epu8(vacc3, voutput_min); vacc0 = _mm_min_epu8(vacc0, voutput_max); vacc1 = _mm_min_epu8(vacc1, voutput_max); vacc2 = _mm_min_epu8(vacc2, voutput_max); vacc3 = _mm_min_epu8(vacc3, voutput_max); _mm_storeu_si128((__m128i*) y, vacc0); _mm_storeu_si128((__m128i*) y + 1, vacc1); _mm_storeu_si128((__m128i*) y + 2, vacc2); _mm_storeu_si128((__m128i*) y + 3, vacc3); y += 64; } for (; n >= 16; n -= 16) { __m128i vacc = _mm_loadu_si128((const __m128i*) x); x += 16; vacc = _mm_min_epu8(vacc, voutput_max); vacc = _mm_max_epu8(vacc, voutput_min); _mm_storeu_si128((__m128i*) y, vacc); y += 16; } if XNN_UNLIKELY(n != 0) { __m128i vacc = _mm_loadu_si128((const __m128i*) x); vacc = _mm_min_epu8(vacc, voutput_max); vacc = _mm_max_epu8(vacc, voutput_min); if (n & 8) { _mm_storel_epi64((__m128i*) y, vacc); y += 8; vacc = _mm_unpackhi_epi64(vacc, vacc); } if (n & 4) { unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vacc)); y += 4; vacc = _mm_srli_epi64(vacc, 32); } if (n & 2) { unaligned_store_u16(y, (uint16_t) _mm_cvtsi128_si32(vacc)); y += 2; vacc = _mm_srli_epi32(vacc, 16); } if (n & 1) { *y = (uint8_t) _mm_cvtsi128_si32(vacc); } } } void xnn_x16_transposec_ukernel__8x8_reuse_multi_sse2( const uint16_t* input, uint16_t* output, size_t input_stride, size_t output_stride, size_t block_width, size_t block_height) XNN_OOB_READS { assert(output_stride >= block_height * sizeof(uint16_t)); assert(input_stride >= block_width * sizeof(uint16_t)); const size_t tile_height = 8; const size_t tile_width = 8; const size_t tile_hbytes = tile_height * sizeof(uint16_t); const size_t tile_wbytes = tile_width * sizeof(uint16_t); const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride; const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint16_t); const uint16_t* i0 = input; uint16_t* o0 = (uint16_t*) output; uint16_t* o1 = (uint16_t*) ((uintptr_t) o0 + output_stride); uint16_t* o2 = (uint16_t*) ((uintptr_t) o1 + output_stride); uint16_t* o3 = (uint16_t*) ((uintptr_t) o2 + output_stride); uint16_t* o4 = (uint16_t*) ((uintptr_t) o3 + output_stride); uint16_t* o5 = (uint16_t*) ((uintptr_t) o4 + output_stride); uint16_t* o6 = (uint16_t*) ((uintptr_t) o5 + output_stride); uint16_t* o7 = (uint16_t*) ((uintptr_t) o6 + output_stride); do { if XNN_UNPREDICTABLE(block_width < 2) { o1 = o0; } if XNN_UNPREDICTABLE(block_width <= 2) { o2 = o0; } if XNN_UNPREDICTABLE(block_width < 4) { o3 = o0; } if XNN_UNPREDICTABLE(block_width <= 4) { o4 = o0; } if XNN_UNPREDICTABLE(block_width < 6) { o5 = o0; } if XNN_UNPREDICTABLE(block_width <= 6) { o6 = o0; } if XNN_UNPREDICTABLE(block_width < 8) { o7 = o0; } size_t bh = block_height; for (; bh >= 8; bh -= 8) { const __m128i v3_0 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_1 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_2 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_4 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_5 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_6 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_7 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint16_t*) ((uintptr_t) i0 + input_stride); const __m128i v2_0 = _mm_unpacklo_epi16(v3_0, v3_1); const __m128i v2_1 = _mm_unpackhi_epi16(v3_0, v3_1); const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); const __m128i v2_3 = _mm_unpackhi_epi16(v3_2, v3_3); const __m128i v2_4 = _mm_unpacklo_epi16(v3_4, v3_5); const __m128i v2_5 = _mm_unpackhi_epi16(v3_4, v3_5); const __m128i v2_6 = _mm_unpacklo_epi16(v3_6, v3_7); const __m128i v2_7 = _mm_unpackhi_epi16(v3_6, v3_7); const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_2); const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_2); const __m128i v1_2 = _mm_unpacklo_epi32(v2_1, v2_3); const __m128i v1_3 = _mm_unpackhi_epi32(v2_1, v2_3); const __m128i v1_4 = _mm_unpacklo_epi32(v2_4, v2_6); const __m128i v1_5 = _mm_unpackhi_epi32(v2_4, v2_6); const __m128i v1_6 = _mm_unpacklo_epi32(v2_5, v2_7); const __m128i v1_7 = _mm_unpackhi_epi32(v2_5, v2_7); const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_4); const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_4); const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_5); const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_5); const __m128i v0_4 = _mm_unpacklo_epi64(v1_2, v1_6); const __m128i v0_5 = _mm_unpackhi_epi64(v1_2, v1_6); const __m128i v0_6 = _mm_unpacklo_epi64(v1_3, v1_7); const __m128i v0_7 = _mm_unpackhi_epi64(v1_3, v1_7); _mm_storeu_si128((__m128i*) o7, v0_7); o7 = (uint16_t*) ((uintptr_t) o7 + tile_hbytes); _mm_storeu_si128((__m128i*) o6, v0_6); o6 = (uint16_t*) ((uintptr_t) o6 + tile_hbytes); _mm_storeu_si128((__m128i*) o5, v0_5); o5 = (uint16_t*) ((uintptr_t) o5 + tile_hbytes); _mm_storeu_si128((__m128i*) o4, v0_4); o4 = (uint16_t*) ((uintptr_t) o4 + tile_hbytes); _mm_storeu_si128((__m128i*) o3, v0_3); o3 = (uint16_t*) ((uintptr_t) o3 + tile_hbytes); _mm_storeu_si128((__m128i*) o2, v0_2); o2 = (uint16_t*) ((uintptr_t) o2 + tile_hbytes); _mm_storeu_si128((__m128i*) o1, v0_1); o1 = (uint16_t*) ((uintptr_t) o1 + tile_hbytes); _mm_storeu_si128((__m128i*) o0, v0_0); o0 = (uint16_t*) ((uintptr_t) o0 + tile_hbytes); } if (bh != 0) { const __m128i v3_0 = _mm_loadu_si128((const __m128i*) i0); const uint16_t *i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride); if XNN_UNPREDICTABLE(bh < 2) { i1 = i0; } const __m128i v3_1 = _mm_loadu_si128((const __m128i*) i1); const uint16_t *i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride); if XNN_UNPREDICTABLE(bh <= 2) { i2 = i1; } const __m128i v3_2 = _mm_loadu_si128((const __m128i*) i2); const uint16_t *i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride); if XNN_UNPREDICTABLE(bh < 4) { i3 = i2; } const __m128i v3_3 = _mm_loadu_si128((const __m128i*) i3); const uint16_t *i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride); if XNN_UNPREDICTABLE(bh <= 4) { i4 = i3; } const __m128i v3_4 = _mm_loadu_si128((const __m128i*) i4); const uint16_t *i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride); if XNN_UNPREDICTABLE(bh < 6) { i5 = i4; } const __m128i v3_5 = _mm_loadu_si128((const __m128i*) i5); const uint16_t *i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride); if XNN_UNPREDICTABLE(bh <= 6) { i6 = i5; } const __m128i v3_6 = _mm_loadu_si128((const __m128i*) i6); const __m128i v3_7 = _mm_undefined_si128(); const __m128i v2_0 = _mm_unpacklo_epi16(v3_0, v3_1); const __m128i v2_1 = _mm_unpackhi_epi16(v3_0, v3_1); const __m128i v2_2 = _mm_unpacklo_epi16(v3_2, v3_3); const __m128i v2_3 = _mm_unpackhi_epi16(v3_2, v3_3); const __m128i v2_4 = _mm_unpacklo_epi16(v3_4, v3_5); const __m128i v2_5 = _mm_unpackhi_epi16(v3_4, v3_5); const __m128i v2_6 = _mm_unpacklo_epi16(v3_6, v3_7); const __m128i v2_7 = _mm_unpackhi_epi16(v3_6, v3_7); const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_2); const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_2); const __m128i v1_2 = _mm_unpacklo_epi32(v2_1, v2_3); const __m128i v1_3 = _mm_unpackhi_epi32(v2_1, v2_3); const __m128i v1_4 = _mm_unpacklo_epi32(v2_4, v2_6); const __m128i v1_5 = _mm_unpackhi_epi32(v2_4, v2_6); const __m128i v1_6 = _mm_unpacklo_epi32(v2_5, v2_7); const __m128i v1_7 = _mm_unpackhi_epi32(v2_5, v2_7); __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_4); __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_4); __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_5); __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_5); __m128i v0_4 = _mm_unpacklo_epi64(v1_2, v1_6); __m128i v0_5 = _mm_unpackhi_epi64(v1_2, v1_6); __m128i v0_6 = _mm_unpacklo_epi64(v1_3, v1_7); __m128i v0_7 = _mm_unpackhi_epi64(v1_3, v1_7); if (bh & 4) { _mm_storel_epi64((__m128i*) o7, v0_7); o7 += 4; _mm_storel_epi64((__m128i*) o6, v0_6); o6 += 4; _mm_storel_epi64((__m128i*) o5, v0_5); o5 += 4; _mm_storel_epi64((__m128i*) o4, v0_4); o4 += 4; _mm_storel_epi64((__m128i*) o3, v0_3); o3 += 4; _mm_storel_epi64((__m128i*) o2, v0_2); o2 += 4; _mm_storel_epi64((__m128i*) o1, v0_1); o1 += 4; _mm_storel_epi64((__m128i*) o0, v0_0); o0 += 4; v0_0 = _mm_unpackhi_epi64(v0_0, v0_0); v0_1 = _mm_unpackhi_epi64(v0_1, v0_1); v0_2 = _mm_unpackhi_epi64(v0_2, v0_2); v0_3 = _mm_unpackhi_epi64(v0_3, v0_3); v0_4 = _mm_unpackhi_epi64(v0_4, v0_4); v0_5 = _mm_unpackhi_epi64(v0_5, v0_5); v0_6 = _mm_unpackhi_epi64(v0_6, v0_6); v0_7 = _mm_unpackhi_epi64(v0_7, v0_7); } if (bh & 2) { unaligned_store_u32(o7, (uint32_t) _mm_cvtsi128_si32(v0_7)); o7 += 2; unaligned_store_u32(o6, (uint32_t) _mm_cvtsi128_si32(v0_6)); o6 += 2; unaligned_store_u32(o5, (uint32_t) _mm_cvtsi128_si32(v0_5)); o5 += 2; unaligned_store_u32(o4, (uint32_t) _mm_cvtsi128_si32(v0_4)); o4 += 2; unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v0_3)); o3 += 2; unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v0_2)); o2 += 2; unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v0_1)); o1 += 2; unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v0_0)); o0 += 2; v0_0 = _mm_srli_epi64(v0_0, 32); v0_1 = _mm_srli_epi64(v0_1, 32); v0_2 = _mm_srli_epi64(v0_2, 32); v0_3 = _mm_srli_epi64(v0_3, 32); v0_4 = _mm_srli_epi64(v0_4, 32); v0_5 = _mm_srli_epi64(v0_5, 32); v0_6 = _mm_srli_epi64(v0_6, 32); v0_7 = _mm_srli_epi64(v0_7, 32); } if (bh & 1) { unaligned_store_u16(o7, (uint16_t) _mm_cvtsi128_si32(v0_7)); unaligned_store_u16(o6, (uint16_t) _mm_cvtsi128_si32(v0_6)); unaligned_store_u16(o5, (uint16_t) _mm_cvtsi128_si32(v0_5)); unaligned_store_u16(o4, (uint16_t) _mm_cvtsi128_si32(v0_4)); unaligned_store_u16(o3, (uint16_t) _mm_cvtsi128_si32(v0_3)); unaligned_store_u16(o2, (uint16_t) _mm_cvtsi128_si32(v0_2)); unaligned_store_u16(o1, (uint16_t) _mm_cvtsi128_si32(v0_1)); unaligned_store_u16(o0, (uint16_t) _mm_cvtsi128_si32(v0_0)); } } i0 = (const uint16_t*) ((uintptr_t) i0 + input_reset); o0 = (uint16_t*) ((uintptr_t) o0 + output_reset); o1 = (uint16_t*) ((uintptr_t) o1 + output_reset); o2 = (uint16_t*) ((uintptr_t) o2 + output_reset); o3 = (uint16_t*) ((uintptr_t) o3 + output_reset); o4 = (uint16_t*) ((uintptr_t) o4 + output_reset); o5 = (uint16_t*) ((uintptr_t) o5 + output_reset); o6 = (uint16_t*) ((uintptr_t) o6 + output_reset); o7 = (uint16_t*) ((uintptr_t) o7 + output_reset); block_width = doz(block_width, tile_width); } while (block_width != 0); } void xnn_x32_unpool_ukernel__sse2( size_t kernel_elements, size_t channels, uint32_t fill, const uint32_t* input, const uint32_t* index, uint32_t** output) { // Pre-initialize outputs with constant. const __m128i vfill = _mm_set1_epi32((int) fill); uint32_t** os = output; do { uint32_t* o = *os++; size_t c = channels; for (; c >= 4; c -= 4) { _mm_storeu_si128((__m128i*) o, vfill); o += 4; } if (c != 0) { if (c & 2) { _mm_storel_epi64((__m128i*) o, vfill); o += 2; } if (c & 1) { *((int*) o) = _mm_cvtsi128_si32(vfill); } } } while (--kernel_elements != 0); // Copy indexed elements to output. size_t offset = 0; do { const uint32_t i = *index++; *((uint32_t*) ((uintptr_t) output[i] + offset)) = *input++; offset += sizeof(uint32_t); } while (--channels != 0); } void xnn_x32_zip_x2_ukernel__sse2( size_t n, const uint32_t* input, uint32_t* output) { assert(n != 0); assert(n % 4 == 0); const uint32_t* x = input; const uint32_t* y = (const uint32_t*) ((uintptr_t) x + n); uint32_t* o = output; while (n >= 16) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 4; const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 4; const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi32(vx, vy); _mm_storeu_si128((__m128i*) o, vxy_lo); _mm_storeu_si128((__m128i*) (o + 4), vxy_hi); o += 8; n -= 16; } if XNN_UNLIKELY(n != 0) { if (n & 8) { const __m128i vx = _mm_loadl_epi64((const __m128i*) x); x += 2; const __m128i vy = _mm_loadl_epi64((const __m128i*) y); y += 2; const __m128i vxy = _mm_unpacklo_epi32(vx, vy); _mm_storeu_si128((__m128i*) o, vxy); o += 4; } if (n & 4) { const uint32_t vx = *x; const uint32_t vy = *y; o[0] = vx; o[1] = vy; } } } void xnn_x32_zip_x3_ukernel__sse2( size_t n, const uint32_t* input, uint32_t* output) { assert(n != 0); assert(n % 4 == 0); const float* x = (const float*) input; const float* y = (const float*) ((uintptr_t) x + n); const float* z = (const float*) ((uintptr_t) y + n); float* o = (float*) output; while (n >= 16) { // vx = ( x3, x2, x1, x0 ) const __m128 vx = _mm_loadu_ps(x); x += 4; // vy = ( y3, y2, y1, y0 ) const __m128 vy = _mm_loadu_ps(y); y += 4; // vz = ( z3, z2, z1, z0 ) const __m128 vz = _mm_loadu_ps(z); z += 4; // vxy = ( y2, y0, x2, x0 ) const __m128 vxy = _mm_shuffle_ps(vx, vy, _MM_SHUFFLE(2, 0, 2, 0)); // vyz = ( z3, z1, y3, y1 ) const __m128 vyz = _mm_shuffle_ps(vy, vz, _MM_SHUFFLE(3, 1, 3, 1)); // vzx = ( x3, x1, z2, z0 ) const __m128 vzx = _mm_shuffle_ps(vz, vx, _MM_SHUFFLE(3, 1, 2, 0)); // vxyz0 = ( x1, z0, y0, x0 ) const __m128 vxyz0 = _mm_shuffle_ps(vxy, vzx, _MM_SHUFFLE(2, 0, 2, 0)); // vxyz1 = ( y2, x2, z1, y1 ) const __m128 vxyz1 = _mm_shuffle_ps(vyz, vxy, _MM_SHUFFLE(3, 1, 2, 0)); // vxyz2 = ( z3, y3, x3, z2 ) const __m128 vxyz2 = _mm_shuffle_ps(vzx, vyz, _MM_SHUFFLE(3, 1, 3, 1)); _mm_storeu_ps(o, vxyz0); _mm_storeu_ps(o + 4, vxyz1); _mm_storeu_ps(o + 8, vxyz2); o += 12; n -= 16; } if XNN_UNLIKELY(n != 0) { if (n & 8) { // vx = ( -, -, x1, x0 ) const __m128 vx = _mm_castpd_ps(_mm_load_sd((const double*) x)); x += 2; // vy = ( -, -, y1, y0 ) const __m128 vy = _mm_castpd_ps(_mm_load_sd((const double*) y)); y += 2; // vz = ( -, -, z1, z0 ) const __m128 vz = _mm_castpd_ps(_mm_load_sd((const double*) z)); z += 2; // vxy = ( y1, x1, y0, x0 ) const __m128 vxy = _mm_unpacklo_ps(vx, vy); // vzx = ( x1, z1, x0, z0 ) const __m128 vzx = _mm_unpacklo_ps(vz, vx); // vyz = ( z1, y1, z0, y0 ) const __m128 vyz = _mm_unpacklo_ps(vy, vz); _mm_storeu_ps(o, _mm_shuffle_ps(vxy, vzx, _MM_SHUFFLE(3, 0, 1, 0))); _mm_storeh_pi((__m64*) (o + 4), vyz); o += 6; } if (n & 4) { const __m128 vx = _mm_load_ss(x); const __m128 vy = _mm_load_ss(y); const __m128 vz = _mm_load_ss(z); _mm_store_ss(o, vx); _mm_store_ss(o + 1, vy); _mm_store_ss(o + 2, vz); } } } void xnn_x32_zip_x4_ukernel__sse2( size_t n, const uint32_t* input, uint32_t* output) { assert(n != 0); assert(n % 4 == 0); const uint32_t* x = input; const uint32_t* y = (const uint32_t*) ((uintptr_t) x + n); const uint32_t* z = (const uint32_t*) ((uintptr_t) y + n); const uint32_t* w = (const uint32_t*) ((uintptr_t) z + n); uint32_t* o = output; while (n >= 16) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 4; const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 4; const __m128i vz = _mm_loadu_si128((const __m128i*) z); z += 4; const __m128i vw = _mm_loadu_si128((const __m128i*) w); w += 4; const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi32(vx, vy); const __m128i vzw_lo = _mm_unpacklo_epi32(vz, vw); const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw); const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo); const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi); const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); _mm_storeu_si128((__m128i*) o, vxyzw0); _mm_storeu_si128((__m128i*) (o + 4), vxyzw1); _mm_storeu_si128((__m128i*) (o + 8), vxyzw2); _mm_storeu_si128((__m128i*) (o + 12), vxyzw3); o += 16; n -= 16; } if XNN_UNLIKELY(n != 0) { if (n & 8) { const __m128i vx = _mm_loadl_epi64((const __m128i*) x); x += 2; const __m128i vy = _mm_loadl_epi64((const __m128i*) y); y += 2; const __m128i vz = _mm_loadl_epi64((const __m128i*) z); z += 2; const __m128i vw = _mm_loadl_epi64((const __m128i*) w); w += 2; const __m128i vxy = _mm_unpacklo_epi32(vx, vy); const __m128i vzw = _mm_unpacklo_epi32(vz, vw); const __m128i vxyzw_lo = _mm_unpacklo_epi64(vxy, vzw); const __m128i vxyzw_hi = _mm_unpackhi_epi64(vxy, vzw); _mm_storeu_si128((__m128i*) o, vxyzw_lo); _mm_storeu_si128((__m128i*) (o + 4), vxyzw_hi); o += 8; } if (n & 4) { const uint32_t vx = *x; const uint32_t vy = *y; const uint32_t vz = *z; const uint32_t vw = *w; o[0] = vx; o[1] = vy; o[2] = vz; o[3] = vw; } } } void xnn_x32_zip_xm_ukernel__sse2( size_t n, size_t m, const uint32_t* input, uint32_t* output) { assert(n != 0); assert(n % 4 == 0); assert(m >= 4); const uint32_t* w = input; const size_t group_increment = m * 4; const size_t input_increment = n * 3; const size_t output_increment = 16 - m * n; const uint32_t* last_input = (const uint32_t*) ((uintptr_t) input + n * (m - 1)); uint32_t* last_output = (uint32_t*) ((uintptr_t) output + (m * 4 - 16)); for (size_t i = 0; i < m; i += 4) { w = (const uint32_t*) ((uintptr_t) w + input_increment); if (w >= last_input) { w = last_input; } const uint32_t* z = (const uint32_t*) ((uintptr_t) w - n); const uint32_t* y = (const uint32_t*) ((uintptr_t) z - n); const uint32_t* x = (const uint32_t*) ((uintptr_t) y - n); size_t k = n; while (k >= 16) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 4; const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 4; const __m128i vz = _mm_loadu_si128((const __m128i*) z); z += 4; const __m128i vw = _mm_loadu_si128((const __m128i*) w); w += 4; const __m128i vxy_lo = _mm_unpacklo_epi32(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi32(vx, vy); const __m128i vzw_lo = _mm_unpacklo_epi32(vz, vw); const __m128i vzw_hi = _mm_unpackhi_epi32(vz, vw); const __m128i vxyzw0 = _mm_unpacklo_epi64(vxy_lo, vzw_lo); const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); const __m128i vxyzw2 = _mm_unpacklo_epi64(vxy_hi, vzw_hi); const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); _mm_storeu_si128((__m128i*) output, vxyzw0); output = (uint32_t*) ((uintptr_t) output + group_increment); _mm_storeu_si128((__m128i*) output, vxyzw1); output = (uint32_t*) ((uintptr_t) output + group_increment); _mm_storeu_si128((__m128i*) output, vxyzw2); output = (uint32_t*) ((uintptr_t) output + group_increment); _mm_storeu_si128((__m128i*) output, vxyzw3); output = (uint32_t*) ((uintptr_t) output + group_increment); k -= 16; } if XNN_UNLIKELY(k != 0) { if (k & 8) { const __m128i vx = _mm_loadl_epi64((const __m128i*) x); x += 2; const __m128i vy = _mm_loadl_epi64((const __m128i*) y); y += 2; const __m128i vz = _mm_loadl_epi64((const __m128i*) z); z += 2; const __m128i vw = _mm_loadl_epi64((const __m128i*) w); w += 2; const __m128i vxy = _mm_unpacklo_epi32(vx, vy); const __m128i vzw = _mm_unpacklo_epi32(vz, vw); const __m128i vxyzw_lo = _mm_unpacklo_epi64(vxy, vzw); const __m128i vxyzw_hi = _mm_unpackhi_epi64(vxy, vzw); _mm_storeu_si128((__m128i*) output, vxyzw_lo); output = (uint32_t*) ((uintptr_t) output + group_increment); _mm_storeu_si128((__m128i*) output, vxyzw_hi); output = (uint32_t*) ((uintptr_t) output + group_increment); } if (k & 4) { const uint32_t vx = *x; const uint32_t vy = *y; const uint32_t vz = *z; const uint32_t vw = *w++; output[0] = vx; output[1] = vy; output[2] = vz; output[3] = vw; output = (uint32_t*) ((uintptr_t) output + group_increment); } } output = (uint32_t*) ((uintptr_t) output + output_increment); if (output > last_output) { output = last_output; } } } void xnn_x8_transposec_ukernel__16x16_reuse_mov_sse2( const uint8_t* input, uint8_t* output, size_t input_stride, size_t output_stride, size_t block_width, size_t block_height) XNN_OOB_READS { assert(output_stride >= block_height * sizeof(uint8_t)); assert(input_stride >= block_width * sizeof(uint8_t)); const size_t tile_height = 16; const size_t tile_width = 16; const size_t tile_hbytes = tile_height * sizeof(uint8_t); const size_t tile_wbytes = tile_width * sizeof(uint8_t); const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride; const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint8_t) - tile_hbytes; const uint8_t* i0 = input; uint8_t* o = (uint8_t*) ((uintptr_t) output - tile_hbytes); const size_t minus_output_stride = -output_stride; do { const size_t rem = min(block_width - 1, 15); const size_t oN_stride = rem * output_stride; const size_t oN_offset = oN_stride + tile_hbytes; size_t bh = block_height; for (; bh >= 16; bh -= 16) { const __m128i v4_0 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_1 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_2 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_3 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_4 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_5 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_6 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_7 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_8 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_9 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_10 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_11 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_12 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_13 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_14 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v4_15 = _mm_loadu_si128((const __m128i*) i0); i0 = (uint8_t*) ((uintptr_t) i0 + input_stride); const __m128i v3_0 = _mm_unpacklo_epi8(v4_0, v4_1); const __m128i v3_1 = _mm_unpackhi_epi8(v4_0, v4_1); const __m128i v3_2 = _mm_unpacklo_epi8(v4_2, v4_3); const __m128i v3_3 = _mm_unpackhi_epi8(v4_2, v4_3); const __m128i v3_4 = _mm_unpacklo_epi8(v4_4, v4_5); const __m128i v3_5 = _mm_unpackhi_epi8(v4_4, v4_5); const __m128i v3_6 = _mm_unpacklo_epi8(v4_6, v4_7); const __m128i v3_7 = _mm_unpackhi_epi8(v4_6, v4_7); const __m128i v3_8 = _mm_unpacklo_epi8(v4_8, v4_9); const __m128i v3_9 = _mm_unpackhi_epi8(v4_8, v4_9); const __m128i v3_10 = _mm_unpacklo_epi8(v4_10, v4_11); const __m128i v3_11 = _mm_unpackhi_epi8(v4_10, v4_11); const __m128i v3_12 = _mm_unpacklo_epi8(v4_12, v4_13); const __m128i v3_13 = _mm_unpackhi_epi8(v4_12, v4_13); const __m128i v3_14 = _mm_unpacklo_epi8(v4_14, v4_15); const __m128i v3_15 = _mm_unpackhi_epi8(v4_14, v4_15); const __m128i v2_0 = _mm_unpacklo_epi16(v3_0, v3_2); const __m128i v2_1 = _mm_unpackhi_epi16(v3_0, v3_2); const __m128i v2_2 = _mm_unpacklo_epi16(v3_1, v3_3); const __m128i v2_3 = _mm_unpackhi_epi16(v3_1, v3_3); const __m128i v2_4 = _mm_unpacklo_epi16(v3_4, v3_6); const __m128i v2_5 = _mm_unpackhi_epi16(v3_4, v3_6); const __m128i v2_6 = _mm_unpacklo_epi16(v3_5, v3_7); const __m128i v2_7 = _mm_unpackhi_epi16(v3_5, v3_7); const __m128i v2_8 = _mm_unpacklo_epi16(v3_8, v3_10); const __m128i v2_9 = _mm_unpackhi_epi16(v3_8, v3_10); const __m128i v2_10 = _mm_unpacklo_epi16(v3_9, v3_11); const __m128i v2_11 = _mm_unpackhi_epi16(v3_9, v3_11); const __m128i v2_12 = _mm_unpacklo_epi16(v3_12, v3_14); const __m128i v2_13 = _mm_unpackhi_epi16(v3_12, v3_14); const __m128i v2_14 = _mm_unpacklo_epi16(v3_13, v3_15); const __m128i v2_15 = _mm_unpackhi_epi16(v3_13, v3_15); const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_4); const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_4); const __m128i v1_2 = _mm_unpacklo_epi32(v2_1, v2_5); const __m128i v1_3 = _mm_unpackhi_epi32(v2_1, v2_5); const __m128i v1_4 = _mm_unpacklo_epi32(v2_2, v2_6); const __m128i v1_5 = _mm_unpackhi_epi32(v2_2, v2_6); const __m128i v1_6 = _mm_unpacklo_epi32(v2_3, v2_7); const __m128i v1_7 = _mm_unpackhi_epi32(v2_3, v2_7); const __m128i v1_8 = _mm_unpacklo_epi32(v2_8, v2_12); const __m128i v1_9 = _mm_unpackhi_epi32(v2_8, v2_12); const __m128i v1_10 = _mm_unpacklo_epi32(v2_9, v2_13); const __m128i v1_11 = _mm_unpackhi_epi32(v2_9, v2_13); const __m128i v1_12 = _mm_unpacklo_epi32(v2_10, v2_14); const __m128i v1_13 = _mm_unpackhi_epi32(v2_10, v2_14); const __m128i v1_14 = _mm_unpacklo_epi32(v2_11, v2_15); const __m128i v1_15 = _mm_unpackhi_epi32(v2_11, v2_15); const __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_8); const __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_8); const __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_9); const __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_9); const __m128i v0_4 = _mm_unpacklo_epi64(v1_2, v1_10); const __m128i v0_5 = _mm_unpackhi_epi64(v1_2, v1_10); const __m128i v0_6 = _mm_unpacklo_epi64(v1_3, v1_11); const __m128i v0_7 = _mm_unpackhi_epi64(v1_3, v1_11); const __m128i v0_8 = _mm_unpacklo_epi64(v1_4, v1_12); const __m128i v0_9 = _mm_unpackhi_epi64(v1_4, v1_12); const __m128i v0_10 = _mm_unpacklo_epi64(v1_5, v1_13); const __m128i v0_11 = _mm_unpackhi_epi64(v1_5, v1_13); const __m128i v0_12 = _mm_unpacklo_epi64(v1_6, v1_14); const __m128i v0_13 = _mm_unpackhi_epi64(v1_6, v1_14); const __m128i v0_14 = _mm_unpacklo_epi64(v1_7, v1_15); const __m128i v0_15 = _mm_unpackhi_epi64(v1_7, v1_15); o = (uint8_t*) ((uintptr_t) o + oN_offset); _mm_storeu_si128((__m128i*) o, v0_15); uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 15) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_14); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 15) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_13); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 13) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_12); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 13) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_11); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 11) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_10); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 11) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_9); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 9) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_8); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 9) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_7); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 7) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_6); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 7) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_5); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 5) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_4); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 5) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_3); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 3) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_2); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 3) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_1); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 1) { o = oN; } _mm_storeu_si128((__m128i*) o, v0_0); } o = (uint8_t*) ((uintptr_t) o + tile_hbytes); if (bh != 0) { const __m128i v4_0 = _mm_loadu_si128((const __m128i*) i0); const uint8_t *i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride); if XNN_UNPREDICTABLE(bh < 2) { i1 = i0; } const __m128i v4_1 = _mm_loadu_si128((const __m128i*) i1); const uint8_t *i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride); if XNN_UNPREDICTABLE(bh <= 2) { i2 = i1; } const __m128i v4_2 = _mm_loadu_si128((const __m128i*) i2); const uint8_t *i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride); if XNN_UNPREDICTABLE(bh < 4) { i3 = i2; } const __m128i v4_3 = _mm_loadu_si128((const __m128i*) i3); const uint8_t *i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride); if XNN_UNPREDICTABLE(bh <= 4) { i4 = i3; } const __m128i v4_4 = _mm_loadu_si128((const __m128i*) i4); const uint8_t *i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride); if XNN_UNPREDICTABLE(bh < 6) { i5 = i4; } const __m128i v4_5 = _mm_loadu_si128((const __m128i*) i5); const uint8_t *i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride); if XNN_UNPREDICTABLE(bh <= 6) { i6 = i5; } const __m128i v4_6 = _mm_loadu_si128((const __m128i*) i6); const uint8_t *i7 = (const uint8_t*) ((uintptr_t) i6 + input_stride); if XNN_UNPREDICTABLE(bh < 8) { i7 = i6; } const __m128i v4_7 = _mm_loadu_si128((const __m128i*) i7); const uint8_t *i8 = (const uint8_t*) ((uintptr_t) i7 + input_stride); if XNN_UNPREDICTABLE(bh <= 8) { i8 = i7; } const __m128i v4_8 = _mm_loadu_si128((const __m128i*) i8); const uint8_t *i9 = (const uint8_t*) ((uintptr_t) i8 + input_stride); if XNN_UNPREDICTABLE(bh < 10) { i9 = i8; } const __m128i v4_9 = _mm_loadu_si128((const __m128i*) i9); const uint8_t *i10 = (const uint8_t*) ((uintptr_t) i9 + input_stride); if XNN_UNPREDICTABLE(bh <= 10) { i10 = i9; } const __m128i v4_10 = _mm_loadu_si128((const __m128i*) i10); const uint8_t *i11 = (const uint8_t*) ((uintptr_t) i10 + input_stride); if XNN_UNPREDICTABLE(bh < 12) { i11 = i10; } const __m128i v4_11 = _mm_loadu_si128((const __m128i*) i11); const uint8_t *i12 = (const uint8_t*) ((uintptr_t) i11 + input_stride); if XNN_UNPREDICTABLE(bh <= 12) { i12 = i11; } const __m128i v4_12 = _mm_loadu_si128((const __m128i*) i12); const uint8_t *i13 = (const uint8_t*) ((uintptr_t) i12 + input_stride); if XNN_UNPREDICTABLE(bh < 14) { i13 = i12; } const __m128i v4_13 = _mm_loadu_si128((const __m128i*) i13); const uint8_t *i14 = (const uint8_t*) ((uintptr_t) i13 + input_stride); if XNN_UNPREDICTABLE(bh <= 14) { i14 = i13; } const __m128i v4_14 = _mm_loadu_si128((const __m128i*) i14); const __m128i v4_15 = _mm_undefined_si128(); const __m128i v3_0 = _mm_unpacklo_epi8(v4_0, v4_1); const __m128i v3_1 = _mm_unpackhi_epi8(v4_0, v4_1); const __m128i v3_2 = _mm_unpacklo_epi8(v4_2, v4_3); const __m128i v3_3 = _mm_unpackhi_epi8(v4_2, v4_3); const __m128i v3_4 = _mm_unpacklo_epi8(v4_4, v4_5); const __m128i v3_5 = _mm_unpackhi_epi8(v4_4, v4_5); const __m128i v3_6 = _mm_unpacklo_epi8(v4_6, v4_7); const __m128i v3_7 = _mm_unpackhi_epi8(v4_6, v4_7); const __m128i v3_8 = _mm_unpacklo_epi8(v4_8, v4_9); const __m128i v3_9 = _mm_unpackhi_epi8(v4_8, v4_9); const __m128i v3_10 = _mm_unpacklo_epi8(v4_10, v4_11); const __m128i v3_11 = _mm_unpackhi_epi8(v4_10, v4_11); const __m128i v3_12 = _mm_unpacklo_epi8(v4_12, v4_13); const __m128i v3_13 = _mm_unpackhi_epi8(v4_12, v4_13); const __m128i v3_14 = _mm_unpacklo_epi8(v4_14, v4_15); const __m128i v3_15 = _mm_unpackhi_epi8(v4_14, v4_15); const __m128i v2_0 = _mm_unpacklo_epi16(v3_0, v3_2); const __m128i v2_1 = _mm_unpackhi_epi16(v3_0, v3_2); const __m128i v2_2 = _mm_unpacklo_epi16(v3_1, v3_3); const __m128i v2_3 = _mm_unpackhi_epi16(v3_1, v3_3); const __m128i v2_4 = _mm_unpacklo_epi16(v3_4, v3_6); const __m128i v2_5 = _mm_unpackhi_epi16(v3_4, v3_6); const __m128i v2_6 = _mm_unpacklo_epi16(v3_5, v3_7); const __m128i v2_7 = _mm_unpackhi_epi16(v3_5, v3_7); const __m128i v2_8 = _mm_unpacklo_epi16(v3_8, v3_10); const __m128i v2_9 = _mm_unpackhi_epi16(v3_8, v3_10); const __m128i v2_10 = _mm_unpacklo_epi16(v3_9, v3_11); const __m128i v2_11 = _mm_unpackhi_epi16(v3_9, v3_11); const __m128i v2_12 = _mm_unpacklo_epi16(v3_12, v3_14); const __m128i v2_13 = _mm_unpackhi_epi16(v3_12, v3_14); const __m128i v2_14 = _mm_unpacklo_epi16(v3_13, v3_15); const __m128i v2_15 = _mm_unpackhi_epi16(v3_13, v3_15); const __m128i v1_0 = _mm_unpacklo_epi32(v2_0, v2_4); const __m128i v1_1 = _mm_unpackhi_epi32(v2_0, v2_4); const __m128i v1_2 = _mm_unpacklo_epi32(v2_1, v2_5); const __m128i v1_3 = _mm_unpackhi_epi32(v2_1, v2_5); const __m128i v1_4 = _mm_unpacklo_epi32(v2_2, v2_6); const __m128i v1_5 = _mm_unpackhi_epi32(v2_2, v2_6); const __m128i v1_6 = _mm_unpacklo_epi32(v2_3, v2_7); const __m128i v1_7 = _mm_unpackhi_epi32(v2_3, v2_7); const __m128i v1_8 = _mm_unpacklo_epi32(v2_8, v2_12); const __m128i v1_9 = _mm_unpackhi_epi32(v2_8, v2_12); const __m128i v1_10 = _mm_unpacklo_epi32(v2_9, v2_13); const __m128i v1_11 = _mm_unpackhi_epi32(v2_9, v2_13); const __m128i v1_12 = _mm_unpacklo_epi32(v2_10, v2_14); const __m128i v1_13 = _mm_unpackhi_epi32(v2_10, v2_14); const __m128i v1_14 = _mm_unpacklo_epi32(v2_11, v2_15); const __m128i v1_15 = _mm_unpackhi_epi32(v2_11, v2_15); __m128i v0_0 = _mm_unpacklo_epi64(v1_0, v1_8); __m128i v0_1 = _mm_unpackhi_epi64(v1_0, v1_8); __m128i v0_2 = _mm_unpacklo_epi64(v1_1, v1_9); __m128i v0_3 = _mm_unpackhi_epi64(v1_1, v1_9); __m128i v0_4 = _mm_unpacklo_epi64(v1_2, v1_10); __m128i v0_5 = _mm_unpackhi_epi64(v1_2, v1_10); __m128i v0_6 = _mm_unpacklo_epi64(v1_3, v1_11); __m128i v0_7 = _mm_unpackhi_epi64(v1_3, v1_11); __m128i v0_8 = _mm_unpacklo_epi64(v1_4, v1_12); __m128i v0_9 = _mm_unpackhi_epi64(v1_4, v1_12); __m128i v0_10 = _mm_unpacklo_epi64(v1_5, v1_13); __m128i v0_11 = _mm_unpackhi_epi64(v1_5, v1_13); __m128i v0_12 = _mm_unpacklo_epi64(v1_6, v1_14); __m128i v0_13 = _mm_unpackhi_epi64(v1_6, v1_14); __m128i v0_14 = _mm_unpacklo_epi64(v1_7, v1_15); __m128i v0_15 = _mm_unpackhi_epi64(v1_7, v1_15); if (bh & 8) { o = (uint8_t*) ((uintptr_t) o + oN_stride); _mm_storel_epi64((__m128i*) o, v0_15); uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 15) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_14); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 15) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_13); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 13) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_12); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 13) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_11); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 11) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_10); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 11) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_9); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 9) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_8); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 9) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_7); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 7) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_6); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 7) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_5); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 5) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_4); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 5) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_3); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 3) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_2); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 3) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_1); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 1) { o = oN; } _mm_storel_epi64((__m128i*) o, v0_0); o += 8; v0_0 = _mm_unpackhi_epi64(v0_0, v0_0); v0_1 = _mm_unpackhi_epi64(v0_1, v0_1); v0_2 = _mm_unpackhi_epi64(v0_2, v0_2); v0_3 = _mm_unpackhi_epi64(v0_3, v0_3); v0_4 = _mm_unpackhi_epi64(v0_4, v0_4); v0_5 = _mm_unpackhi_epi64(v0_5, v0_5); v0_6 = _mm_unpackhi_epi64(v0_6, v0_6); v0_7 = _mm_unpackhi_epi64(v0_7, v0_7); v0_8 = _mm_unpackhi_epi64(v0_8, v0_8); v0_9 = _mm_unpackhi_epi64(v0_9, v0_9); v0_10 = _mm_unpackhi_epi64(v0_10, v0_10); v0_11 = _mm_unpackhi_epi64(v0_11, v0_11); v0_12 = _mm_unpackhi_epi64(v0_12, v0_12); v0_13 = _mm_unpackhi_epi64(v0_13, v0_13); v0_14 = _mm_unpackhi_epi64(v0_14, v0_14); v0_15 = _mm_unpackhi_epi64(v0_15, v0_15); } if (bh & 4) { o = (uint8_t*) ((uintptr_t) o + oN_stride); unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_15)); uint8_t *oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 15) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_14)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 15) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_13)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 13) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_12)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 13) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_11)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 11) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_10)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 11) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_9)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 9) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_8)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 9) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_7)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 7) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_6)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 7) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_5)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 5) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_4)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 5) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_3)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 3) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_2)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 3) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_1)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 1) { o = oN; } unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(v0_0)); o += 4; v0_0 = _mm_srli_epi64(v0_0, 32); v0_1 = _mm_srli_epi64(v0_1, 32); v0_2 = _mm_srli_epi64(v0_2, 32); v0_3 = _mm_srli_epi64(v0_3, 32); v0_4 = _mm_srli_epi64(v0_4, 32); v0_5 = _mm_srli_epi64(v0_5, 32); v0_6 = _mm_srli_epi64(v0_6, 32); v0_7 = _mm_srli_epi64(v0_7, 32); v0_8 = _mm_srli_epi64(v0_8, 32); v0_9 = _mm_srli_epi64(v0_9, 32); v0_10 = _mm_srli_epi64(v0_10, 32); v0_11 = _mm_srli_epi64(v0_11, 32); v0_12 = _mm_srli_epi64(v0_12, 32); v0_13 = _mm_srli_epi64(v0_13, 32); v0_14 = _mm_srli_epi64(v0_14, 32); v0_15 = _mm_srli_epi64(v0_15, 32); } if (bh & 2) { o = (uint8_t*) ((uintptr_t) o + oN_stride); unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_15)); uint8_t* oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 15) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_14)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 15) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_13)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 13) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_12)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 13) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_11)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 11) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_10)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 11) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_9)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 9) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_8)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 9) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_7)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 7) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_6)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 7) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_5)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 5) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_4)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 5) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_3)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 3) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_2)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 3) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_1)); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 1) { o = oN; } unaligned_store_u16(o, (uint16_t) _mm_cvtsi128_si32(v0_0)); o += 2; v0_0 = _mm_srli_epi32(v0_0, 16); v0_1 = _mm_srli_epi32(v0_1, 16); v0_2 = _mm_srli_epi32(v0_2, 16); v0_3 = _mm_srli_epi32(v0_3, 16); v0_4 = _mm_srli_epi32(v0_4, 16); v0_5 = _mm_srli_epi32(v0_5, 16); v0_6 = _mm_srli_epi32(v0_6, 16); v0_7 = _mm_srli_epi32(v0_7, 16); v0_8 = _mm_srli_epi32(v0_8, 16); v0_9 = _mm_srli_epi32(v0_9, 16); v0_10 = _mm_srli_epi32(v0_10, 16); v0_11 = _mm_srli_epi32(v0_11, 16); v0_12 = _mm_srli_epi32(v0_12, 16); v0_13 = _mm_srli_epi32(v0_13, 16); v0_14 = _mm_srli_epi32(v0_14, 16); v0_15 = _mm_srli_epi32(v0_15, 16); } if (bh & 1) { o = (uint8_t*) ((uintptr_t) o + oN_stride); *o = (uint8_t) _mm_cvtsi128_si32(v0_15); uint8_t* oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 15) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_14); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 15) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_13); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 13) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_12); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 13) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_11); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 11) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_10); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 11) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_9); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 9) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_8); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 9) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_7); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 7) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_6); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 7) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_5); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 5) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_4); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 5) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_3); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 3) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_2); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width >= 3) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_1); oN = (uint8_t*) ((uintptr_t) o + minus_output_stride); if XNN_UNPREDICTABLE(block_width > 1) { o = oN; } *o = (uint8_t) _mm_cvtsi128_si32(v0_0); } } i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset); o = (uint8_t*) ((uintptr_t) o + output_reset); block_width = doz(block_width, tile_width); } while (block_width != 0); } void xnn_x8_zip_x2_ukernel__sse2( size_t n, const uint8_t* input, uint8_t* output) { const uint8_t* x = input; const uint8_t* y = (const uint8_t*) ((uintptr_t) x + n); uint8_t* o = output; if (n >= 16) { do { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 16; const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy); _mm_storeu_si128((__m128i*) o, vxy_lo); _mm_storeu_si128((__m128i*) (o + 16), vxy_hi); o = (void*) ((uintptr_t) o + 32); n -= 16; } while (n >= 16); if (n != 0) { const size_t address_increment = n - 16; const __m128i vx = _mm_loadu_si128((const __m128i*) ((uintptr_t) x + address_increment)); const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment)); const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy); o = (void*) ((uintptr_t) o + address_increment * 2); _mm_storeu_si128((__m128i*) o, vxy_lo); _mm_storeu_si128((__m128i*) o + 1, vxy_hi); } } else { do { const uint8_t vx = *x++; const uint8_t vy = *y++; o[0] = vx; o[1] = vy; o += 2; } while (--n != 0); } } void xnn_x8_zip_x3_ukernel__sse2( size_t n, const uint8_t* input, uint8_t* output) { const uint8_t* x = input; const uint8_t* y = (const uint8_t*) ((uintptr_t) x + n); const uint8_t* z = (const uint8_t*) ((uintptr_t) y + n); uint8_t* o = output; if (n >= 16) { const __m128i vmask0x00FF00FF = _mm_set1_epi16(0x00FF); const __m128i vmask0x0000FFFF = _mm_set1_epi32(0x0000FFFF); do { // vx = ( x15, x14, x13, x12, x11, x10, x9, x8, x7, x6, x5, x4, x3, x2, x1, x0 ) const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; // vy = ( y15, y14, y13, y12, y11, y10, y9, y8, y7, y6, y5, y4, y3, y2, y1, y0 ) const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 16; // vz = ( z15, z14, z13, z12, z11, z10, z9, z8, z7, z6, z5, z4, z3, z2, z1, z0 ) const __m128i vz = _mm_loadu_si128((const __m128i*) z); z += 16; // vxeye = ( y14, x14, y12, x12, y10, x10, y8, x8, y6, x6, y4, x4, y2, x2, y0, x0 ) const __m128i vxeye = _mm_or_si128(_mm_and_si128(vx, vmask0x00FF00FF), _mm_slli_epi16(vy, 8)); // vyozo = ( z15, y15, z13, y13, z11, y11, z9, y9, z7, y7, z5, y5, z3, y3, z1, y1 ) const __m128i vyozo = _mm_or_si128(_mm_andnot_si128(vmask0x00FF00FF, vz), _mm_srli_epi16(vy, 8)); // vzoxo = ( x15, z14, x13, z12, x11, z10, x9, z8, x7, z6, x5, z4, x3, z2, x1, z0 ) const __m128i vzexo = _mm_or_si128(_mm_and_si128(vz, vmask0x00FF00FF), _mm_andnot_si128(vmask0x00FF00FF, vx)); // vxeyezexo = ( x13, z12, y12, x12, x9, z8, y8, x8, x5, z4, y4, x4, x1, z0, y0, x0 ) const __m128i vxeyezexo = _mm_or_si128(_mm_and_si128(vxeye, vmask0x0000FFFF), _mm_slli_epi32(vzexo, 16)); // vyozoxeye = ( y14, x14, z13, y13, y10, x10, z9, y9, y6, x6, z5, y5, y2, x2, z1, y1 ) const __m128i vyozoxeye = _mm_or_si128(_mm_and_si128(vyozo, vmask0x0000FFFF), _mm_andnot_si128(vmask0x0000FFFF, vxeye)); // vzexoyozo = ( z15, y15, x15, z14, z11, y11, x11, z10, z7, y7, x7, z6, z3, y3, x3, z2 ) const __m128i vzexoyozo = _mm_or_si128(_mm_andnot_si128(vmask0x0000FFFF, vyozo), _mm_srli_epi32(vzexo, 16)); // vtemp0 = ( x13, z12, y12, x12, x5, z4, y4, x4, z11, y11, x11, z10, z3, y3, x3, z2 ) const __m128i vtemp0 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vzexoyozo), _mm_castsi128_ps(vxeyezexo), _MM_SHUFFLE(3, 1, 2, 0))); // vtemp1 = ( y10, x10, z9, y9, y2, x2, z1, y1, x9, z8, y8, x8, x1, z0, y0, x0 ) const __m128i vtemp1 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vxeyezexo), _mm_castsi128_ps(vyozoxeye), _MM_SHUFFLE(2, 0, 2, 0))); // vtemp2 = ( z15, y15, x15, z14, z7, y7, x7, z6, y14, x14, z13, y13, y6, x6, z5, y5 ) const __m128i vtemp2 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vyozoxeye), _mm_castsi128_ps(vzexoyozo), _MM_SHUFFLE(3, 1, 3, 1))); // vxyz0 = ( x5, z4, y4, x4, z3, y3, x3, z2, y2, x2, z1, y1, x1, z0, y0, x0 ) const __m128i vxyz0 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vtemp1), _mm_castsi128_ps(vtemp0), _MM_SHUFFLE(2, 0, 2, 0))); // vxyz1 = ( y10, x10, z9, y9, x9, z8, y8, x8, z7, y7, x7, z6, y6, x6, z5, y5 ) const __m128i vxyz1 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vtemp2), _mm_castsi128_ps(vtemp1), _MM_SHUFFLE(3, 1, 2, 0))); // vxyz2 = ( z15, y15, x15, z14, y14, x14, z13, y13, x13, z12, y12, x12, z11, y11, x11, z10 ) const __m128i vxyz2 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vtemp0), _mm_castsi128_ps(vtemp2), _MM_SHUFFLE(3, 1, 3, 1))); _mm_storeu_si128((__m128i*) o, vxyz0); _mm_storeu_si128((__m128i*) o + 1, vxyz1); _mm_storeu_si128((__m128i*) o + 2, vxyz2); o += 48; n -= 16; } while (n >= 16); if (n != 0) { const size_t address_increment = n - 16; // vx = ( x15, x14, x13, x12, x11, x10, x9, x8, x7, x6, x5, x4, x3, x2, x1, x0 ) const __m128i vx = _mm_loadu_si128((const __m128i*) ((uintptr_t) x + address_increment)); // vy = ( y15, y14, y13, y12, y11, y10, y9, y8, y7, y6, y5, y4, y3, y2, y1, y0 ) const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment)); // vz = ( z15, z14, z13, z12, z11, z10, z9, z8, z7, z6, z5, z4, z3, z2, z1, z0 ) const __m128i vz = _mm_loadu_si128((const __m128i*) ((uintptr_t) z + address_increment)); // vxeye = ( y14, x14, y12, x12, y10, x10, y8, x8, y6, x6, y4, x4, y2, x2, y0, x0 ) const __m128i vxeye = _mm_or_si128(_mm_and_si128(vx, vmask0x00FF00FF), _mm_slli_epi16(vy, 8)); // vyozo = ( z15, y15, z13, y13, z11, y11, z9, y9, z7, y7, z5, y5, z3, y3, z1, y1 ) const __m128i vyozo = _mm_or_si128(_mm_andnot_si128(vmask0x00FF00FF, vz), _mm_srli_epi16(vy, 8)); // vzoxo = ( x15, z14, x13, z12, x11, z10, x9, z8, x7, z6, x5, z4, x3, z2, x1, z0 ) const __m128i vzexo = _mm_or_si128(_mm_and_si128(vz, vmask0x00FF00FF), _mm_andnot_si128(vmask0x00FF00FF, vx)); // vxeyezexo = ( x13, z12, y12, x12, x9, z8, y8, x8, x5, z4, y4, x4, x1, z0, y0, x0 ) const __m128i vxeyezexo = _mm_or_si128(_mm_and_si128(vxeye, vmask0x0000FFFF), _mm_slli_epi32(vzexo, 16)); // vyozoxeye = ( y14, x14, z13, y13, y10, x10, z9, y9, y6, x6, z5, y5, y2, x2, z1, y1 ) const __m128i vyozoxeye = _mm_or_si128(_mm_and_si128(vyozo, vmask0x0000FFFF), _mm_andnot_si128(vmask0x0000FFFF, vxeye)); // vzexoyozo = ( z15, y15, x15, z14, z11, y11, x11, z10, z7, y7, x7, z6, z3, y3, x3, z2 ) const __m128i vzexoyozo = _mm_or_si128(_mm_andnot_si128(vmask0x0000FFFF, vyozo), _mm_srli_epi32(vzexo, 16)); // vtemp0 = ( x13, z12, y12, x12, x5, z4, y4, x4, z11, y11, x11, z10, z3, y3, x3, z2 ) const __m128i vtemp0 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vzexoyozo), _mm_castsi128_ps(vxeyezexo), _MM_SHUFFLE(3, 1, 2, 0))); // vtemp1 = ( y10, x10, z9, y9, y2, x2, z1, y1, x9, z8, y8, x8, x1, z0, y0, x0 ) const __m128i vtemp1 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vxeyezexo), _mm_castsi128_ps(vyozoxeye), _MM_SHUFFLE(2, 0, 2, 0))); // vtemp2 = ( z15, y15, x15, z14, z7, y7, x7, z6, y14, x14, z13, y13, y6, x6, z5, y5 ) const __m128i vtemp2 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vyozoxeye), _mm_castsi128_ps(vzexoyozo), _MM_SHUFFLE(3, 1, 3, 1))); // vxyz0 = ( x5, z4, y4, x4, z3, y3, x3, z2, y2, x2, z1, y1, x1, z0, y0, x0 ) const __m128i vxyz0 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vtemp1), _mm_castsi128_ps(vtemp0), _MM_SHUFFLE(2, 0, 2, 0))); // vxyz1 = ( y10, x10, z9, y9, x9, z8, y8, x8, z7, y7, x7, z6, y6, x6, z5, y5 ) const __m128i vxyz1 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vtemp2), _mm_castsi128_ps(vtemp1), _MM_SHUFFLE(3, 1, 2, 0))); // vxyz2 = ( z15, y15, x15, z14, y14, x14, z13, y13, x13, z12, y12, x12, z11, y11, x11, z10 ) const __m128i vxyz2 = _mm_castps_si128( _mm_shuffle_ps(_mm_castsi128_ps(vtemp0), _mm_castsi128_ps(vtemp2), _MM_SHUFFLE(3, 1, 3, 1))); o = (uint8_t*) ((uintptr_t) o + address_increment * 3); _mm_storeu_si128((__m128i*) o, vxyz0); _mm_storeu_si128((__m128i*) o + 1, vxyz1); _mm_storeu_si128((__m128i*) o + 2, vxyz2); } } else { do { const uint8_t vx = *x++; const uint8_t vy = *y++; const uint8_t vz = *z++; o[0] = vx; o[1] = vy; o[2] = vz; o += 3; } while (--n != 0); } } void xnn_x8_zip_x4_ukernel__sse2( size_t n, const uint8_t* input, uint8_t* output) { const uint8_t* x = input; const uint8_t* y = (const uint8_t*) ((uintptr_t) x + n); const uint8_t* z = (const uint8_t*) ((uintptr_t) y + n); const uint8_t* w = (const uint8_t*) ((uintptr_t) z + n); uint8_t* o = output; if (n >= 16) { do { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 16; const __m128i vz = _mm_loadu_si128((const __m128i*) z); z += 16; const __m128i vw = _mm_loadu_si128((const __m128i*) w); w += 16; const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy); const __m128i vzw_lo = _mm_unpacklo_epi8(vz, vw); const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); const __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo); const __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo); const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); _mm_storeu_si128((__m128i*) o, vxyzw0); _mm_storeu_si128((__m128i*) o + 1, vxyzw1); _mm_storeu_si128((__m128i*) o + 2, vxyzw2); _mm_storeu_si128((__m128i*) o + 3, vxyzw3); o = (void*) ((uintptr_t) o + 64); n -= 16; } while (n >= 16); if (n != 0) { const size_t address_increment = n - 16; const __m128i vx = _mm_loadu_si128((const __m128i*) ((uintptr_t) x + address_increment)); const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment)); const __m128i vz = _mm_loadu_si128((const __m128i*) ((uintptr_t) z + address_increment)); const __m128i vw = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + address_increment)); const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy); const __m128i vzw_lo = _mm_unpacklo_epi8(vz, vw); const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); const __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo); const __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo); const __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); const __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); o = (void*) ((uintptr_t) o + address_increment * 4); _mm_storeu_si128((__m128i*) o, vxyzw0); _mm_storeu_si128((__m128i*) o + 1, vxyzw1); _mm_storeu_si128((__m128i*) o + 2, vxyzw2); _mm_storeu_si128((__m128i*) o + 3, vxyzw3); } } else { do { const uint8_t vx = *x++; const uint8_t vy = *y++; const uint8_t vz = *z++; const uint8_t vw = *w++; o[0] = vx; o[1] = vy; o[2] = vz; o[3] = vw; o += 4; } while (--n != 0); } } void xnn_x8_zip_xm_ukernel__sse2( size_t n, size_t m, const uint8_t* input, uint8_t* output) { const uint8_t* w = input; const size_t input_increment = n * 3; const size_t output_increment = 4 - m * n; const uint8_t* last_input = w + n * (m - 1); uint8_t* last_output = (uint8_t*) ((uintptr_t) output + (m - 4)); if (n >= 8) { for (size_t i = 0; i < m; i += 4) { size_t k = n; w = (const uint8_t*) ((uintptr_t) w + input_increment); if (w >= last_input) { w = last_input; } const uint8_t* z = (const uint8_t*) ((uintptr_t) w - n); const uint8_t* y = (const uint8_t*) ((uintptr_t) z - n); const uint8_t* x = (const uint8_t*) ((uintptr_t) y - n); while (k >= 16) { const __m128i vx = _mm_loadu_si128((const __m128i*) x); x += 16; const __m128i vy = _mm_loadu_si128((const __m128i*) y); y += 16; const __m128i vz = _mm_loadu_si128((const __m128i*) z); z += 16; const __m128i vw = _mm_loadu_si128((const __m128i*) w); w += 16; const __m128i vxy_lo = _mm_unpacklo_epi8(vx, vy); const __m128i vxy_hi = _mm_unpackhi_epi8(vx, vy); const __m128i vzw_lo = _mm_unpacklo_epi8(vz, vw); const __m128i vzw_hi = _mm_unpackhi_epi8(vz, vw); __m128i vxyzw0 = _mm_unpacklo_epi16(vxy_lo, vzw_lo); __m128i vxyzw1 = _mm_unpackhi_epi16(vxy_lo, vzw_lo); __m128i vxyzw2 = _mm_unpacklo_epi16(vxy_hi, vzw_hi); __m128i vxyzw3 = _mm_unpackhi_epi16(vxy_hi, vzw_hi); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw1 = _mm_shufflelo_epi16(vxyzw1, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw1 = _mm_unpackhi_epi64(vxyzw1, vxyzw1); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw1 = _mm_shufflelo_epi16(vxyzw1, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw2)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw2 = _mm_shufflelo_epi16(vxyzw2, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw2)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw2 = _mm_unpackhi_epi64(vxyzw2, vxyzw2); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw2)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw2 = _mm_shufflelo_epi16(vxyzw2, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw2)); output = (uint8_t*) ((uintptr_t) output + m); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw3)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw3 = _mm_shufflelo_epi16(vxyzw3, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw3)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw3 = _mm_unpackhi_epi64(vxyzw3, vxyzw3); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw3)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw3 = _mm_shufflelo_epi16(vxyzw3, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw3)); output = (uint8_t*) ((uintptr_t) output + m); k -= 16; }; if (k >= 8) { const __m128i vx = _mm_loadl_epi64((const __m128i*) x); x += 8; const __m128i vy = _mm_loadl_epi64((const __m128i*) y); y += 8; const __m128i vz = _mm_loadl_epi64((const __m128i*) z); z += 8; const __m128i vw = _mm_loadl_epi64((const __m128i*) w); w += 8; const __m128i vxy = _mm_unpacklo_epi8(vx, vy); const __m128i vzw = _mm_unpacklo_epi8(vz, vw); __m128i vxyzw0 = _mm_unpacklo_epi16(vxy, vzw); __m128i vxyzw1 = _mm_unpackhi_epi16(vxy, vzw); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw1 = _mm_shufflelo_epi16(vxyzw1, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw1 = _mm_unpackhi_epi64(vxyzw1, vxyzw1); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw1 = _mm_shufflelo_epi16(vxyzw1, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw1)); output = (uint8_t*) ((uintptr_t) output + m); k -= 8; } if (k != 0) { const size_t address_decrement = 8 - k; x -= address_decrement; y -= address_decrement; z -= address_decrement; w -= address_decrement; const __m128i vshift = _mm_cvtsi32_si128((int) address_decrement * 8); const __m128i vx = _mm_srl_epi64(_mm_loadl_epi64((const __m128i*) x), vshift); const __m128i vy = _mm_srl_epi64(_mm_loadl_epi64((const __m128i*) y), vshift); const __m128i vz = _mm_srl_epi64(_mm_loadl_epi64((const __m128i*) z), vshift); const __m128i vw = _mm_srl_epi64(_mm_loadl_epi64((const __m128i*) w), vshift); w += 8; const __m128i vxy = _mm_unpacklo_epi8(vx, vy); const __m128i vzw = _mm_unpacklo_epi8(vz, vw); __m128i vxyzw0 = _mm_unpacklo_epi16(vxy, vzw); __m128i vxyzw1 = _mm_unpackhi_epi16(vxy, vzw); if (k & 4) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = vxyzw1; } if (k & 2) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_shufflelo_epi16(vxyzw0, _MM_SHUFFLE(3, 2, 3, 2)); unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); } if (k & 1) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vxyzw0)); output = (uint8_t*) ((uintptr_t) output + m); } } output = (uint8_t*) ((uintptr_t) output + output_increment); if (output > last_output) { output = last_output; } } } else { const uint8_t* i = input; uint8_t* o = output; size_t k = n; do { size_t l = m; const uint8_t* ii = i++; do { *o++ = *ii; ii += n; } while (--l != 0); } while (--k != 0); } } void xnn_xx_fill_ukernel__sse2_x64( size_t rows, size_t channels, void* output, size_t output_stride, const uint32_t fill_pattern) { assert(rows != 0); assert(channels != 0); const size_t output_increment = output_stride - channels; const __m128i vfill = _mm_shuffle_epi32(_mm_cvtsi32_si128(fill_pattern), _MM_SHUFFLE(0, 0, 0, 0)); do { size_t c = channels; for (; c >= 64 * sizeof(uint8_t); c -= 64 * sizeof(uint8_t)) { _mm_storeu_si128((__m128i*) output, vfill); _mm_storeu_si128((__m128i*) output + 1, vfill); _mm_storeu_si128((__m128i*) output + 2, vfill); _mm_storeu_si128((__m128i*) output + 3, vfill); output = ((uint8_t*) output + 64); } for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) { _mm_storeu_si128((__m128i*) output, vfill); output = ((uint8_t*) output + 16); } if XNN_UNLIKELY(c != 0) { if XNN_LIKELY(c & (8 * sizeof(uint8_t))) { _mm_storel_epi64(output, vfill); output = ((uint8_t*) output + 8); } if XNN_LIKELY(c & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, fill_pattern); output = ((uint8_t*) output + 4); } uint32_t vfill_subpattern = fill_pattern; if XNN_LIKELY(c & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) vfill_subpattern); vfill_subpattern >>= 16; output = ((uint8_t*) output + 2); } if XNN_LIKELY(c & (1 * sizeof(uint8_t))) { *((uint8_t*) output) = (uint8_t) vfill_subpattern; output = ((uint8_t*) output + 1); } } output = (void*) ((uintptr_t) output + output_increment); } while (--rows != 0); } void xnn_xx_pad_ukernel__sse2( size_t rows, size_t channels, size_t pre_padding, size_t post_padding, const void* input, size_t input_stride, void* output, size_t output_stride, const uint32_t fill_pattern) XNN_OOB_READS { const size_t input_increment = input_stride - channels; const size_t output_increment = output_stride - (pre_padding + channels + post_padding); const __m128i vfill_pattern = _mm_shuffle_epi32(_mm_cvtsi32_si128((int) fill_pattern), _MM_SHUFFLE(0, 0, 0, 0)); do { // Pre-pad input channels. size_t l = pre_padding; if XNN_LIKELY(l != 0) { for (; l >= 16 * sizeof(uint8_t); l -= 16 * sizeof(uint8_t)) { _mm_storeu_si128((__m128i*) output, vfill_pattern); output = (uint8_t*) output + 16; } if (l & (8 * sizeof(uint8_t))) { _mm_storel_epi64((__m128i*) output, vfill_pattern); output = (uint8_t*) output + 8; } uint32_t vfill_subpattern = fill_pattern; if (l & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, vfill_subpattern); output = (uint8_t*) output + 4; } if (l & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, vfill_subpattern); vfill_subpattern >>= 16; output = (uint8_t*) output + 2; } if (l & (1 * sizeof(uint8_t))) { *((uint8_t*) output) = (uint8_t) vfill_subpattern; output = (uint8_t*) output + 1; } } // Copy input channels. size_t c = channels; for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) { const __m128i vdata = _mm_loadu_si128((const __m128i*) input); input = (const uint8_t*) input + 16; _mm_storeu_si128((__m128i*) output, vdata); output = (uint8_t*) output + 16; } if XNN_UNLIKELY(c != 0) { __m128i vdata = _mm_loadu_si128((const __m128i*) input); input = (const void*) ((uintptr_t) input + c); if (c & (8 * sizeof(uint8_t))) { _mm_storel_epi64((__m128i*) output, vdata); vdata = _mm_unpackhi_epi64(vdata, vdata); output = (uint8_t*) output + 8; } if (c & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vdata)); vdata = _mm_srli_epi64(vdata, 32); output = (uint8_t*) output + 4; } uint32_t vsubdata = (uint32_t) _mm_cvtsi128_si32(vdata); if (c & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) vsubdata); vsubdata >>= 16; output = (uint8_t*) output + 2; } if (c & (1 * sizeof(uint8_t))) { *((uint8_t*) output) = (uint8_t) vsubdata; output = (uint8_t*) output + 1; } } // Post-pad input channels. size_t r = post_padding; if XNN_LIKELY(r != 0) { for (; r >= 16 * sizeof(uint8_t); r -= 16 * sizeof(uint8_t)) { _mm_storeu_si128((__m128i*) output, vfill_pattern); output = (uint8_t*) output + 16; } if (r & (8 * sizeof(uint8_t))) { _mm_storel_epi64((__m128i*) output, vfill_pattern); output = (uint8_t*) output + 8; } uint32_t vfill_subpattern = fill_pattern; if (r & (4 * sizeof(uint8_t))) { unaligned_store_u32(output, vfill_subpattern); output = (uint8_t*) output + 4; } if (r & (2 * sizeof(uint8_t))) { unaligned_store_u16(output, (uint16_t) vfill_subpattern); vfill_subpattern >>= 16; output = (uint8_t*) output + 2; } if (r & (1 * sizeof(uint8_t))) { *((uint8_t*) output) = (uint8_t) vfill_subpattern; output = (uint8_t*) output + 1; } } input = (const void*) ((uintptr_t) input + input_increment); output = (void*) ((uintptr_t) output + output_increment); } while (--rows != 0); }