// Auto-generated file. Do not edit! // Template: src/f32-velu/wasmsimd-rr2-lut16-p3.c.in // Generator: tools/xngen // // Copyright 2020 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 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16]; void xnn_f32_velu_ukernel__wasmsimd_arm_rr2_lut16_p3_x16( 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 v128_t vprescale = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.prescale); const v128_t valpha = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.alpha); const v128_t vbeta = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.beta); const v128_t vsat_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.sat_cutoff); const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.magic_bias); const v128_t vlog2e = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.log2e); const v128_t vindex_mask = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.index_mask); const v128_t vminus_ln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.minus_ln2_hi); const v128_t vminus_ln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.minus_ln2_lo); const v128_t vc3 = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.c3); const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.c2); const v128_t vone = wasm_v128_load64_splat(params->wasmsimd_rr2_lut16_p3.one); for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { v128_t vx0123 = wasm_v128_load(x); v128_t vx4567 = wasm_v128_load(x + 4); v128_t vx89AB = wasm_v128_load(x + 8); v128_t vxCDEF = wasm_v128_load(x + 12); x += 16; const v128_t vz0123 = wasm_f32x4_max(wasm_f32x4_mul(vx0123, vprescale), vsat_cutoff); const v128_t vz4567 = wasm_f32x4_max(wasm_f32x4_mul(vx4567, vprescale), vsat_cutoff); const v128_t vz89AB = wasm_f32x4_max(wasm_f32x4_mul(vx89AB, vprescale), vsat_cutoff); const v128_t vzCDEF = wasm_f32x4_max(wasm_f32x4_mul(vxCDEF, vprescale), vsat_cutoff); v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias); v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias); v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias); v128_t vnCDEF = wasm_f32x4_add(wasm_f32x4_mul(vzCDEF, vlog2e), vmagic_bias); const v128_t vidx0123 = wasm_i32x4_shl(wasm_v128_and(vn0123, vindex_mask), 2); const v128_t ven0123 = wasm_i32x4_shl(vn0123, 19); const v128_t vidx4567 = wasm_i32x4_shl(wasm_v128_and(vn4567, vindex_mask), 2); const v128_t ven4567 = wasm_i32x4_shl(vn4567, 19); const v128_t vidx89AB = wasm_i32x4_shl(wasm_v128_and(vn89AB, vindex_mask), 2); const v128_t ven89AB = wasm_i32x4_shl(vn89AB, 19); const v128_t vidxCDEF = wasm_i32x4_shl(wasm_v128_and(vnCDEF, vindex_mask), 2); const v128_t venCDEF = wasm_i32x4_shl(vnCDEF, 19); const uint64_t vidx01 = wasm_i64x2_extract_lane(vidx0123, 0); const uint64_t vidx23 = wasm_i64x2_extract_lane(vidx0123, 1); const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01)); const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32))); const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)); const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32))); const v128_t vl0123 = wasm_f32x4_make(vl0, vl1, vl2, vl3); const uint64_t vidx45 = wasm_i64x2_extract_lane(vidx4567, 0); const uint64_t vidx67 = wasm_i64x2_extract_lane(vidx4567, 1); const float vl4 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45)); const float vl5 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32))); const float vl6 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)); const float vl7 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32))); const v128_t vl4567 = wasm_f32x4_make(vl4, vl5, vl6, vl7); const uint64_t vidx89 = wasm_i64x2_extract_lane(vidx89AB, 0); const uint64_t vidxAB = wasm_i64x2_extract_lane(vidx89AB, 1); const float vl8 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89)); const float vl9 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32))); const float vlA = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB)); const float vlB = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32))); const v128_t vl89AB = wasm_f32x4_make(vl8, vl9, vlA, vlB); const uint64_t vidxCD = wasm_i64x2_extract_lane(vidxCDEF, 0); const uint64_t vidxEF = wasm_i64x2_extract_lane(vidxCDEF, 1); const float vlC = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD)); const float vlD = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32))); const float vlE = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF)); const float vlF = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32))); const v128_t vlCDEF = wasm_f32x4_make(vlC, vlD, vlE, vlF); vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias); v128_t vs0123 = wasm_i32x4_add(vl0123, ven0123); vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); v128_t vs4567 = wasm_i32x4_add(vl4567, ven4567); vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias); v128_t vs89AB = wasm_i32x4_add(vl89AB, ven89AB); vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias); v128_t vsCDEF = wasm_i32x4_add(vlCDEF, venCDEF); v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123); v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567); v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB); v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF); vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123); vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567); vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB); vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF); v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt0123), vc2); v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt4567), vc2); v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt89AB), vc2); v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc3, vtCDEF), vc2); vp0123 = wasm_f32x4_mul(vp0123, vt0123); vp4567 = wasm_f32x4_mul(vp4567, vt4567); vp89AB = wasm_f32x4_mul(vp89AB, vt89AB); vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF); vt0123 = wasm_f32x4_mul(vt0123, vs0123); vs0123 = wasm_f32x4_sub(vs0123, vone); vt4567 = wasm_f32x4_mul(vt4567, vs4567); vs4567 = wasm_f32x4_sub(vs4567, vone); vt89AB = wasm_f32x4_mul(vt89AB, vs89AB); vs89AB = wasm_f32x4_sub(vs89AB, vone); vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF); vsCDEF = wasm_f32x4_sub(vsCDEF, vone); vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123); vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567); vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB); vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF); const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha); const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha); const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha); const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha); const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31); vx0123 = wasm_f32x4_mul(vx0123, vbeta); const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31); vx4567 = wasm_f32x4_mul(vx4567, vbeta); const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31); vx89AB = wasm_f32x4_mul(vx89AB, vbeta); const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31); vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta); const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123); const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567); const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB); const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF); wasm_v128_store(y, vy0123); wasm_v128_store(y + 4, vy4567); wasm_v128_store(y + 8, vy89AB); wasm_v128_store(y + 12, vyCDEF); y += 16; } for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { v128_t vx = wasm_v128_load(x); x += 4; const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff); v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias); const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2); const v128_t ven = wasm_i32x4_shl(vn, 19); const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0); const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1); const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)); const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))); const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)); const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))); const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3); v128_t vs = wasm_i32x4_add(vl, ven); vn = wasm_f32x4_sub(vn, vmagic_bias); v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz); vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt); v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2); vp = wasm_f32x4_mul(vp, vt); vt = wasm_f32x4_mul(vt, vs); vs = wasm_f32x4_sub(vs, vone); vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt); const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha); const v128_t vsignm = wasm_i32x4_shr(vx, 31); vx = wasm_f32x4_mul(vx, vbeta); const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm); wasm_v128_store(y, vy); y += 4; } if XNN_UNLIKELY(n != 0) { v128_t vx = wasm_v128_load(x); const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff); v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias); const v128_t vidx = wasm_i32x4_shl(wasm_v128_and(vn, vindex_mask), 2); const v128_t ven = wasm_i32x4_shl(vn, 19); const uint64_t vidx_lo = wasm_i64x2_extract_lane(vidx, 0); const uint64_t vidx_hi = wasm_i64x2_extract_lane(vidx, 1); const float vl0 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)); const float vl1 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))); const float vl2 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)); const float vl3 = *((const float*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))); const v128_t vl = wasm_f32x4_make(vl0, vl1, vl2, vl3); v128_t vs = wasm_i32x4_add(vl, ven); vn = wasm_f32x4_sub(vn, vmagic_bias); v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz); vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt); v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc3, vt), vc2); vp = wasm_f32x4_mul(vp, vt); vt = wasm_f32x4_mul(vt, vs); vs = wasm_f32x4_sub(vs, vone); vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt); const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha); const v128_t vsignm = wasm_i32x4_shr(vx, 31); vx = wasm_f32x4_mul(vx, vbeta); v128_t vy = wasm_v128_bitselect(ve, vx, vsignm); if (n & (2 * sizeof(float))) { *((double*) y) = wasm_f64x2_extract_lane(vy, 0); vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3); y += 2; } if (n & (1 * sizeof(float))) { *y = wasm_f32x4_extract_lane(vy, 0); } } }