// Copyright (c) Facebook, Inc. and its affiliates. // All rights reserved. // // Copyright 2019 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 void xnn_qs8_requantize_gemmlowp__sse2( size_t n, const int32_t* input, float scale, int8_t zero_point, int8_t qmin, int8_t qmax, int8_t* output) { assert(n % 16 == 0); assert(scale < 1.0f); assert(scale >= 0x1.0p-32f); // Compute requantization parameters. const uint32_t scale_bits = float_as_uint32(scale); // Multiplier is in [0x40000000, 0x7FFFFF80] range. const int32_t multiplier = (int32_t)(((scale_bits & UINT32_C(0x007FFFFF)) | UINT32_C(0x00800000)) << 7); assert(multiplier >= INT32_C(0x40000000)); assert(multiplier <= INT32_C(0x7FFFFF80)); // Shift is in [0, 31] range. const int32_t shift = 127 + 31 - 32 - (float_as_uint32(scale) >> 23); assert(shift >= 0); assert(shift < 32); const __m128i vmultiplier = _mm_set1_epi32(multiplier); const __m128i vzero_point = _mm_set1_epi16((short) zero_point); const __m128i vqmin = _mm_set1_epi16((short) qmin); const __m128i vqmax = _mm_set1_epi16((short) qmax); const __m128i vshift = _mm_cvtsi32_si128((int) shift); const uint32_t remainder_mask = (UINT32_C(1) << shift) - UINT32_C(1); const __m128i vremainder_mask = _mm_set1_epi32((int) remainder_mask); const __m128i vthreshold = _mm_set1_epi32((int) (remainder_mask >> 1)); const __m128i vq31rounding = _mm_set1_epi64x(UINT64_C(0x40000000)); for (; n != 0; n -= 16) { const __m128i x = _mm_loadu_si128((const __m128i*) input); const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4)); const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8)); const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12)); input += 16; const __m128i x_neg_mask = _mm_cmpgt_epi32(_mm_setzero_si128(), x); const __m128i y_neg_mask = _mm_cmpgt_epi32(_mm_setzero_si128(), y); const __m128i z_neg_mask = _mm_cmpgt_epi32(_mm_setzero_si128(), z); const __m128i w_neg_mask = _mm_cmpgt_epi32(_mm_setzero_si128(), w); const __m128i x_abs = _mm_sub_epi32(_mm_xor_si128(x, x_neg_mask), x_neg_mask); const __m128i y_abs = _mm_sub_epi32(_mm_xor_si128(y, y_neg_mask), y_neg_mask); const __m128i z_abs = _mm_sub_epi32(_mm_xor_si128(z, z_neg_mask), z_neg_mask); const __m128i w_abs = _mm_sub_epi32(_mm_xor_si128(w, w_neg_mask), w_neg_mask); const __m128i x_abs_rev = _mm_shuffle_epi32(x_abs, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i y_abs_rev = _mm_shuffle_epi32(y_abs, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i z_abs_rev = _mm_shuffle_epi32(z_abs, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i w_abs_rev = _mm_shuffle_epi32(w_abs, _MM_SHUFFLE(2, 3, 0, 1)); const __m128i x_abs_product_even = _mm_mul_epu32(x_abs, vmultiplier); const __m128i y_abs_product_even = _mm_mul_epu32(y_abs, vmultiplier); const __m128i z_abs_product_even = _mm_mul_epu32(z_abs, vmultiplier); const __m128i w_abs_product_even = _mm_mul_epu32(w_abs, vmultiplier); const __m128i x_neg_mask_even = _mm_shuffle_epi32(x_neg_mask, _MM_SHUFFLE(2, 2, 0, 0)); const __m128i y_neg_mask_even = _mm_shuffle_epi32(y_neg_mask, _MM_SHUFFLE(2, 2, 0, 0)); const __m128i z_neg_mask_even = _mm_shuffle_epi32(z_neg_mask, _MM_SHUFFLE(2, 2, 0, 0)); const __m128i w_neg_mask_even = _mm_shuffle_epi32(w_neg_mask, _MM_SHUFFLE(2, 2, 0, 0)); const __m128i x_product_even = _mm_sub_epi64(_mm_xor_si128(x_abs_product_even, x_neg_mask_even), x_neg_mask_even); const __m128i y_product_even = _mm_sub_epi64(_mm_xor_si128(y_abs_product_even, y_neg_mask_even), y_neg_mask_even); const __m128i z_product_even = _mm_sub_epi64(_mm_xor_si128(z_abs_product_even, z_neg_mask_even), z_neg_mask_even); const __m128i w_product_even = _mm_sub_epi64(_mm_xor_si128(w_abs_product_even, w_neg_mask_even), w_neg_mask_even); const __m128i x_rounded_product_even = _mm_add_epi64(x_product_even, vq31rounding); const __m128i y_rounded_product_even = _mm_add_epi64(y_product_even, vq31rounding); const __m128i z_rounded_product_even = _mm_add_epi64(z_product_even, vq31rounding); const __m128i w_rounded_product_even = _mm_add_epi64(w_product_even, vq31rounding); const __m128i x_abs_product_odd = _mm_mul_epu32(x_abs_rev, vmultiplier); const __m128i y_abs_product_odd = _mm_mul_epu32(y_abs_rev, vmultiplier); const __m128i z_abs_product_odd = _mm_mul_epu32(z_abs_rev, vmultiplier); const __m128i w_abs_product_odd = _mm_mul_epu32(w_abs_rev, vmultiplier); const __m128i x_neg_mask_odd = _mm_shuffle_epi32(x_neg_mask, _MM_SHUFFLE(3, 3, 1, 1)); const __m128i y_neg_mask_odd = _mm_shuffle_epi32(y_neg_mask, _MM_SHUFFLE(3, 3, 1, 1)); const __m128i z_neg_mask_odd = _mm_shuffle_epi32(z_neg_mask, _MM_SHUFFLE(3, 3, 1, 1)); const __m128i w_neg_mask_odd = _mm_shuffle_epi32(w_neg_mask, _MM_SHUFFLE(3, 3, 1, 1)); const __m128i x_product_odd = _mm_sub_epi64(_mm_xor_si128(x_abs_product_odd, x_neg_mask_odd), x_neg_mask_odd); const __m128i y_product_odd = _mm_sub_epi64(_mm_xor_si128(y_abs_product_odd, y_neg_mask_odd), y_neg_mask_odd); const __m128i z_product_odd = _mm_sub_epi64(_mm_xor_si128(z_abs_product_odd, z_neg_mask_odd), z_neg_mask_odd); const __m128i w_product_odd = _mm_sub_epi64(_mm_xor_si128(w_abs_product_odd, w_neg_mask_odd), w_neg_mask_odd); const __m128i x_rounded_product_odd = _mm_add_epi64(x_product_odd, vq31rounding); const __m128i y_rounded_product_odd = _mm_add_epi64(y_product_odd, vq31rounding); const __m128i z_rounded_product_odd = _mm_add_epi64(z_product_odd, vq31rounding); const __m128i w_rounded_product_odd = _mm_add_epi64(w_product_odd, vq31rounding); const __m128i x_q31product_even = _mm_srli_epi64(x_rounded_product_even, 31); const __m128i x_q31product_odd = _mm_srli_epi64(x_rounded_product_odd, 31); const __m128i y_q31product_even = _mm_srli_epi64(y_rounded_product_even, 31); const __m128i y_q31product_odd = _mm_srli_epi64(y_rounded_product_odd, 31); const __m128i z_q31product_even = _mm_srli_epi64(z_rounded_product_even, 31); const __m128i z_q31product_odd = _mm_srli_epi64(z_rounded_product_odd, 31); const __m128i w_q31product_even = _mm_srli_epi64(w_rounded_product_even, 31); const __m128i w_q31product_odd = _mm_srli_epi64(w_rounded_product_odd, 31); const __m128i x_q31product_0213 = _mm_castps_si128(_mm_shuffle_ps( _mm_castsi128_ps(x_q31product_even), _mm_castsi128_ps(x_q31product_odd), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i y_q31product_0213 = _mm_castps_si128(_mm_shuffle_ps( _mm_castsi128_ps(y_q31product_even), _mm_castsi128_ps(y_q31product_odd), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i z_q31product_0213 = _mm_castps_si128(_mm_shuffle_ps( _mm_castsi128_ps(z_q31product_even), _mm_castsi128_ps(z_q31product_odd), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i w_q31product_0213 = _mm_castps_si128(_mm_shuffle_ps( _mm_castsi128_ps(w_q31product_even), _mm_castsi128_ps(w_q31product_odd), _MM_SHUFFLE(2, 0, 2, 0))); const __m128i x_q31product = _mm_shuffle_epi32(x_q31product_0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i y_q31product = _mm_shuffle_epi32(y_q31product_0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i z_q31product = _mm_shuffle_epi32(z_q31product_0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i w_q31product = _mm_shuffle_epi32(w_q31product_0213, _MM_SHUFFLE(3, 1, 2, 0)); const __m128i x_remainder = _mm_add_epi32(_mm_and_si128(x_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), x_q31product)); const __m128i y_remainder = _mm_add_epi32(_mm_and_si128(y_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), y_q31product)); const __m128i z_remainder = _mm_add_epi32(_mm_and_si128(z_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), z_q31product)); const __m128i w_remainder = _mm_add_epi32(_mm_and_si128(w_q31product, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), w_q31product)); const __m128i x_scaled = _mm_sub_epi32(_mm_sra_epi32(x_q31product, vshift), _mm_cmpgt_epi32(x_remainder, vthreshold)); const __m128i y_scaled = _mm_sub_epi32(_mm_sra_epi32(y_q31product, vshift), _mm_cmpgt_epi32(y_remainder, vthreshold)); const __m128i z_scaled = _mm_sub_epi32(_mm_sra_epi32(z_q31product, vshift), _mm_cmpgt_epi32(z_remainder, vthreshold)); const __m128i w_scaled = _mm_sub_epi32(_mm_sra_epi32(w_q31product, vshift), _mm_cmpgt_epi32(w_remainder, vthreshold)); const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_scaled, y_scaled), vzero_point); const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_scaled, w_scaled), vzero_point); const __m128i xy_clamped = _mm_max_epi16(_mm_min_epi16(xy_packed, vqmax), vqmin); const __m128i zw_clamped = _mm_max_epi16(_mm_min_epi16(zw_packed, vqmax), vqmin); const __m128i xyzw_clamped = _mm_packs_epi16(xy_clamped, zw_clamped); // 16x PSHUFD // 4x SHUFPS // 8x PMULUDQ // 8x PXOR (setzero) // 12x PXOR // 4x PAND // 8x PADDQ // 4x PADDD // 8x PSUBQ // 8x PSUBD // 8x PSRLQ (immediate) // 4x PSRAD (register) // 12x PCMPGTD // 2x PACKSSDW // 2x PADDSW // 2x PMAXSW // 2x PMINSW // 1x PACKSSWB // --------------------- // 113 instructions total _mm_storeu_si128((__m128i*) output, xyzw_clamped); output += 16; } }