// Tencent is pleased to support the open source community by making ncnn available. // // Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved. // // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except // in compliance with the License. You may obtain a copy of the License at // // https://opensource.org/licenses/BSD-3-Clause // // Unless required by applicable law or agreed to in writing, software distributed // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR // CONDITIONS OF ANY KIND, either express or implied. See the License for the // specific language governing permissions and limitations under the License. static void conv3x3s1_pack1to4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt) { int inch = bottom_blob.c; int outw = top_blob.w; int outh = top_blob.h; int outch = top_blob.c; const float* bias = _bias; int nn_outch = outch >> 1; int remain_outch_start = nn_outch << 1; #pragma omp parallel for num_threads(opt.num_threads) for (int pp = 0; pp < nn_outch; pp++) { int p = pp * 2; Mat out0 = top_blob.channel(p); Mat out1 = top_blob.channel(p + 1); __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f); __m128 _bias1 = bias ? _mm_loadu_ps((const float*)bias + (p + 1) * 4) : _mm_set1_ps(0.f); out0.fill(_bias0); out1.fill(_bias1); const float* k0 = kernel.channel(p); const float* k1 = kernel.channel(p + 1); for (int q = 0; q < inch; q++) { float* outptr0 = out0; float* outptr1 = out1; const Mat img0 = bottom_blob.channel(q); const float* r0 = img0.row(0); const float* r1 = img0.row(1); const float* r2 = img0.row(2); __m128 _k00_0 = _mm_loadu_ps(k0); __m128 _k01_0 = _mm_loadu_ps(k0 + 4); __m128 _k02_0 = _mm_loadu_ps(k0 + 8); __m128 _k10_0 = _mm_loadu_ps(k0 + 12); __m128 _k11_0 = _mm_loadu_ps(k0 + 16); __m128 _k12_0 = _mm_loadu_ps(k0 + 20); __m128 _k20_0 = _mm_loadu_ps(k0 + 24); __m128 _k21_0 = _mm_loadu_ps(k0 + 28); __m128 _k22_0 = _mm_loadu_ps(k0 + 32); __m128 _k00_1 = _mm_loadu_ps(k1); __m128 _k01_1 = _mm_loadu_ps(k1 + 4); __m128 _k02_1 = _mm_loadu_ps(k1 + 8); __m128 _k10_1 = _mm_loadu_ps(k1 + 12); __m128 _k11_1 = _mm_loadu_ps(k1 + 16); __m128 _k12_1 = _mm_loadu_ps(k1 + 20); __m128 _k20_1 = _mm_loadu_ps(k1 + 24); __m128 _k21_1 = _mm_loadu_ps(k1 + 28); __m128 _k22_1 = _mm_loadu_ps(k1 + 32); int i = 0; for (; i < outh; i++) { int j = 0; for (; j + 3 < outw; j += 4) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _sum10 = _mm_loadu_ps(outptr1); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10); _mm_storeu_ps(outptr0, _sum00); _mm_storeu_ps(outptr1, _sum10); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _sum11 = _mm_loadu_ps(outptr1 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); _sum01 = _mm_comp_fmadd_ps(_r02, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r03, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r12, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r22, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k22_0, _sum01); _sum11 = _mm_comp_fmadd_ps(_r02, _k00_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r03, _k01_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r04, _k02_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r12, _k10_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r13, _k11_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r14, _k12_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r22, _k20_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r23, _k21_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r24, _k22_1, _sum11); _mm_storeu_ps(outptr0 + 4, _sum01); _mm_storeu_ps(outptr1 + 4, _sum11); __m128 _sum02 = _mm_loadu_ps(outptr0 + 8); __m128 _sum12 = _mm_loadu_ps(outptr1 + 8); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _sum02 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum02); _sum12 = _mm_comp_fmadd_ps(_r03, _k00_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r04, _k01_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r05, _k02_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r13, _k10_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r14, _k11_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r15, _k12_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r23, _k20_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r24, _k21_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r25, _k22_1, _sum12); _mm_storeu_ps(outptr0 + 8, _sum02); _mm_storeu_ps(outptr1 + 8, _sum12); __m128 _r06 = _mm_set1_ps(*(r0 + 5)); __m128 _r16 = _mm_set1_ps(*(r1 + 5)); __m128 _r26 = _mm_set1_ps(*(r2 + 5)); __m128 _sum03 = _mm_loadu_ps(outptr0 + 12); __m128 _sum13 = _mm_loadu_ps(outptr1 + 12); _sum03 = _mm_comp_fmadd_ps(_r04, _k00_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r05, _k01_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r06, _k02_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r14, _k10_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r15, _k11_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r16, _k12_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r24, _k20_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r25, _k21_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r26, _k22_0, _sum03); _sum13 = _mm_comp_fmadd_ps(_r04, _k00_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r05, _k01_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r06, _k02_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r14, _k10_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r15, _k11_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r16, _k12_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r24, _k20_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r25, _k21_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r26, _k22_1, _sum13); _mm_storeu_ps(outptr0 + 12, _sum03); _mm_storeu_ps(outptr1 + 12, _sum13); r0 += 4; r1 += 4; r2 += 4; outptr0 += 16; outptr1 += 16; } for (; j + 1 < outw; j += 2) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _sum10 = _mm_loadu_ps(outptr1); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10); _mm_storeu_ps(outptr0, _sum00); _mm_storeu_ps(outptr1, _sum10); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _sum11 = _mm_loadu_ps(outptr1 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); _sum01 = _mm_comp_fmadd_ps(_r02, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r03, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r12, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r22, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k22_0, _sum01); _sum11 = _mm_comp_fmadd_ps(_r02, _k00_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r03, _k01_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r04, _k02_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r12, _k10_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r13, _k11_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r14, _k12_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r22, _k20_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r23, _k21_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r24, _k22_1, _sum11); _mm_storeu_ps(outptr0 + 4, _sum01); _mm_storeu_ps(outptr1 + 4, _sum11); r0 += 2; r1 += 2; r2 += 2; outptr0 += 8; outptr1 += 8; } for (; j < outw; j++) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _sum10 = _mm_loadu_ps(outptr1); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10); _mm_storeu_ps(outptr0, _sum00); _mm_storeu_ps(outptr1, _sum10); r0 += 1; r1 += 1; r2 += 1; outptr0 += 4; outptr1 += 4; } r0 += 2; r1 += 2; r2 += 2; } k0 += 9 * 4; k1 += 9 * 4; } } #pragma omp parallel for num_threads(opt.num_threads) for (int p = remain_outch_start; p < outch; p++) { Mat out0 = top_blob.channel(p); __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f); out0.fill(_bias0); const float* k0 = kernel.channel(p); for (int q = 0; q < inch; q++) { float* outptr0 = out0.row(0); const Mat img0 = bottom_blob.channel(q); const float* r0 = img0.row(0); const float* r1 = img0.row(1); const float* r2 = img0.row(2); __m128 _k00 = _mm_loadu_ps(k0); __m128 _k01 = _mm_loadu_ps(k0 + 4); __m128 _k02 = _mm_loadu_ps(k0 + 8); __m128 _k10 = _mm_loadu_ps(k0 + 12); __m128 _k11 = _mm_loadu_ps(k0 + 16); __m128 _k12 = _mm_loadu_ps(k0 + 20); __m128 _k20 = _mm_loadu_ps(k0 + 24); __m128 _k21 = _mm_loadu_ps(k0 + 28); __m128 _k22 = _mm_loadu_ps(k0 + 32); int i = 0; for (; i < outh; i++) { int j = 0; for (; j + 3 < outw; j += 4) { __m128 _sum0 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum0 = _mm_comp_fmadd_ps(_r01, _k00, _sum0); _sum0 = _mm_comp_fmadd_ps(_r02, _k01, _sum0); _sum0 = _mm_comp_fmadd_ps(_r03, _k02, _sum0); _sum0 = _mm_comp_fmadd_ps(_r11, _k10, _sum0); _sum0 = _mm_comp_fmadd_ps(_r12, _k11, _sum0); _sum0 = _mm_comp_fmadd_ps(_r13, _k12, _sum0); _sum0 = _mm_comp_fmadd_ps(_r21, _k20, _sum0); _sum0 = _mm_comp_fmadd_ps(_r22, _k21, _sum0); _sum0 = _mm_comp_fmadd_ps(_r23, _k22, _sum0); __m128 _sum1 = _mm_loadu_ps(outptr0 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); _mm_storeu_ps(outptr0, _sum0); _sum1 = _mm_comp_fmadd_ps(_r02, _k00, _sum1); _sum1 = _mm_comp_fmadd_ps(_r03, _k01, _sum1); _sum1 = _mm_comp_fmadd_ps(_r04, _k02, _sum1); _sum1 = _mm_comp_fmadd_ps(_r12, _k10, _sum1); _sum1 = _mm_comp_fmadd_ps(_r13, _k11, _sum1); _sum1 = _mm_comp_fmadd_ps(_r14, _k12, _sum1); _sum1 = _mm_comp_fmadd_ps(_r22, _k20, _sum1); _sum1 = _mm_comp_fmadd_ps(_r23, _k21, _sum1); _sum1 = _mm_comp_fmadd_ps(_r24, _k22, _sum1); __m128 _sum2 = _mm_loadu_ps(outptr0 + 8); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _mm_storeu_ps(outptr0 + 4, _sum1); _sum2 = _mm_comp_fmadd_ps(_r03, _k00, _sum2); _sum2 = _mm_comp_fmadd_ps(_r04, _k01, _sum2); _sum2 = _mm_comp_fmadd_ps(_r05, _k02, _sum2); _sum2 = _mm_comp_fmadd_ps(_r13, _k10, _sum2); _sum2 = _mm_comp_fmadd_ps(_r14, _k11, _sum2); _sum2 = _mm_comp_fmadd_ps(_r15, _k12, _sum2); _sum2 = _mm_comp_fmadd_ps(_r23, _k20, _sum2); _sum2 = _mm_comp_fmadd_ps(_r24, _k21, _sum2); _sum2 = _mm_comp_fmadd_ps(_r25, _k22, _sum2); __m128 _sum3 = _mm_loadu_ps(outptr0 + 12); __m128 _r06 = _mm_set1_ps(*(r0 + 5)); __m128 _r16 = _mm_set1_ps(*(r1 + 5)); __m128 _r26 = _mm_set1_ps(*(r2 + 5)); _mm_storeu_ps(outptr0 + 8, _sum2); _sum3 = _mm_comp_fmadd_ps(_r04, _k00, _sum3); _sum3 = _mm_comp_fmadd_ps(_r05, _k01, _sum3); _sum3 = _mm_comp_fmadd_ps(_r06, _k02, _sum3); _sum3 = _mm_comp_fmadd_ps(_r14, _k10, _sum3); _sum3 = _mm_comp_fmadd_ps(_r15, _k11, _sum3); _sum3 = _mm_comp_fmadd_ps(_r16, _k12, _sum3); _sum3 = _mm_comp_fmadd_ps(_r24, _k20, _sum3); _sum3 = _mm_comp_fmadd_ps(_r25, _k21, _sum3); _sum3 = _mm_comp_fmadd_ps(_r26, _k22, _sum3); _mm_storeu_ps(outptr0 + 12, _sum3); r0 += 4; r1 += 4; r2 += 4; outptr0 += 16; } for (; j + 1 < outw; j += 2) { __m128 _sum0 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum0 = _mm_comp_fmadd_ps(_r01, _k00, _sum0); _sum0 = _mm_comp_fmadd_ps(_r02, _k01, _sum0); _sum0 = _mm_comp_fmadd_ps(_r03, _k02, _sum0); _sum0 = _mm_comp_fmadd_ps(_r11, _k10, _sum0); _sum0 = _mm_comp_fmadd_ps(_r12, _k11, _sum0); _sum0 = _mm_comp_fmadd_ps(_r13, _k12, _sum0); _sum0 = _mm_comp_fmadd_ps(_r21, _k20, _sum0); _sum0 = _mm_comp_fmadd_ps(_r22, _k21, _sum0); _sum0 = _mm_comp_fmadd_ps(_r23, _k22, _sum0); __m128 _sum1 = _mm_loadu_ps(outptr0 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); _mm_storeu_ps(outptr0, _sum0); _sum1 = _mm_comp_fmadd_ps(_r02, _k00, _sum1); _sum1 = _mm_comp_fmadd_ps(_r03, _k01, _sum1); _sum1 = _mm_comp_fmadd_ps(_r04, _k02, _sum1); _sum1 = _mm_comp_fmadd_ps(_r12, _k10, _sum1); _sum1 = _mm_comp_fmadd_ps(_r13, _k11, _sum1); _sum1 = _mm_comp_fmadd_ps(_r14, _k12, _sum1); _sum1 = _mm_comp_fmadd_ps(_r22, _k20, _sum1); _sum1 = _mm_comp_fmadd_ps(_r23, _k21, _sum1); _sum1 = _mm_comp_fmadd_ps(_r24, _k22, _sum1); _mm_storeu_ps(outptr0 + 4, _sum1); r0 += 2; r1 += 2; r2 += 2; outptr0 += 8; } for (; j < outw; j++) { __m128 _sum0 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum0 = _mm_comp_fmadd_ps(_r01, _k00, _sum0); _sum0 = _mm_comp_fmadd_ps(_r02, _k01, _sum0); _sum0 = _mm_comp_fmadd_ps(_r03, _k02, _sum0); _sum0 = _mm_comp_fmadd_ps(_r11, _k10, _sum0); _sum0 = _mm_comp_fmadd_ps(_r12, _k11, _sum0); _sum0 = _mm_comp_fmadd_ps(_r13, _k12, _sum0); _sum0 = _mm_comp_fmadd_ps(_r21, _k20, _sum0); _sum0 = _mm_comp_fmadd_ps(_r22, _k21, _sum0); _sum0 = _mm_comp_fmadd_ps(_r23, _k22, _sum0); _mm_storeu_ps(outptr0, _sum0); r0 += 1; r1 += 1; r2 += 1; outptr0 += 4; } r0 += 2; r1 += 2; r2 += 2; } k0 += 9 * 4; } } } static void conv3x3s2_pack1to4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt) { int w = bottom_blob.w; int inch = bottom_blob.c; int outw = top_blob.w; int outh = top_blob.h; int outch = top_blob.c; const int tailstep = w - 2 * outw + w; const float* bias = _bias; int nn_outch = outch >> 1; int remain_outch_start = nn_outch << 1; #pragma omp parallel for num_threads(opt.num_threads) for (int pp = 0; pp < nn_outch; pp++) { int p = pp * 2; Mat out0 = top_blob.channel(p); Mat out1 = top_blob.channel(p + 1); __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f); __m128 _bias1 = bias ? _mm_loadu_ps((const float*)bias + (p + 1) * 4) : _mm_set1_ps(0.f); out0.fill(_bias0); out1.fill(_bias1); const float* k0 = kernel.channel(p); const float* k1 = kernel.channel(p + 1); for (int q = 0; q < inch; q++) { float* outptr0 = out0; float* outptr1 = out1; const Mat img0 = bottom_blob.channel(q); const float* r0 = img0.row(0); const float* r1 = img0.row(1); const float* r2 = img0.row(2); __m128 _k00_0 = _mm_loadu_ps(k0); __m128 _k01_0 = _mm_loadu_ps(k0 + 4); __m128 _k02_0 = _mm_loadu_ps(k0 + 8); __m128 _k10_0 = _mm_loadu_ps(k0 + 12); __m128 _k11_0 = _mm_loadu_ps(k0 + 16); __m128 _k12_0 = _mm_loadu_ps(k0 + 20); __m128 _k20_0 = _mm_loadu_ps(k0 + 24); __m128 _k21_0 = _mm_loadu_ps(k0 + 28); __m128 _k22_0 = _mm_loadu_ps(k0 + 32); __m128 _k00_1 = _mm_loadu_ps(k1); __m128 _k01_1 = _mm_loadu_ps(k1 + 4); __m128 _k02_1 = _mm_loadu_ps(k1 + 8); __m128 _k10_1 = _mm_loadu_ps(k1 + 12); __m128 _k11_1 = _mm_loadu_ps(k1 + 16); __m128 _k12_1 = _mm_loadu_ps(k1 + 20); __m128 _k20_1 = _mm_loadu_ps(k1 + 24); __m128 _k21_1 = _mm_loadu_ps(k1 + 28); __m128 _k22_1 = _mm_loadu_ps(k1 + 32); int i = 0; for (; i < outh; i++) { int j = 0; for (; j + 3 < outw; j += 4) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _sum10 = _mm_loadu_ps(outptr1); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10); _mm_storeu_ps(outptr0, _sum00); _mm_storeu_ps(outptr1, _sum10); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _sum11 = _mm_loadu_ps(outptr1 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01); _sum11 = _mm_comp_fmadd_ps(_r03, _k00_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r04, _k01_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r05, _k02_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r13, _k10_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r14, _k11_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r15, _k12_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r23, _k20_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r24, _k21_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r25, _k22_1, _sum11); _mm_storeu_ps(outptr0 + 4, _sum01); _mm_storeu_ps(outptr1 + 4, _sum11); __m128 _sum02 = _mm_loadu_ps(outptr0 + 8); __m128 _sum12 = _mm_loadu_ps(outptr1 + 8); __m128 _r06 = _mm_set1_ps(*(r0 + 5)); __m128 _r16 = _mm_set1_ps(*(r1 + 5)); __m128 _r26 = _mm_set1_ps(*(r2 + 5)); __m128 _r07 = _mm_set1_ps(*(r0 + 6)); __m128 _r17 = _mm_set1_ps(*(r1 + 6)); __m128 _r27 = _mm_set1_ps(*(r2 + 6)); _sum02 = _mm_comp_fmadd_ps(_r05, _k00_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r06, _k01_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r07, _k02_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r15, _k10_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r16, _k11_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r17, _k12_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r25, _k20_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r26, _k21_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r27, _k22_0, _sum02); _sum12 = _mm_comp_fmadd_ps(_r05, _k00_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r06, _k01_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r07, _k02_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r15, _k10_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r16, _k11_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r17, _k12_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r25, _k20_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r26, _k21_1, _sum12); _sum12 = _mm_comp_fmadd_ps(_r27, _k22_1, _sum12); _mm_storeu_ps(outptr0 + 8, _sum02); _mm_storeu_ps(outptr1 + 8, _sum12); __m128 _r08 = _mm_set1_ps(*(r0 + 7)); __m128 _r18 = _mm_set1_ps(*(r1 + 7)); __m128 _r28 = _mm_set1_ps(*(r2 + 7)); __m128 _r09 = _mm_set1_ps(*(r0 + 8)); __m128 _r19 = _mm_set1_ps(*(r1 + 8)); __m128 _r29 = _mm_set1_ps(*(r2 + 8)); __m128 _sum03 = _mm_loadu_ps(outptr0 + 12); __m128 _sum13 = _mm_loadu_ps(outptr1 + 12); _sum03 = _mm_comp_fmadd_ps(_r07, _k00_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r08, _k01_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r09, _k02_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r17, _k10_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r18, _k11_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r19, _k12_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r27, _k20_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r28, _k21_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r29, _k22_0, _sum03); _sum13 = _mm_comp_fmadd_ps(_r07, _k00_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r08, _k01_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r09, _k02_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r17, _k10_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r18, _k11_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r19, _k12_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r27, _k20_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r28, _k21_1, _sum13); _sum13 = _mm_comp_fmadd_ps(_r29, _k22_1, _sum13); _mm_storeu_ps(outptr0 + 12, _sum03); _mm_storeu_ps(outptr1 + 12, _sum13); r0 += 8; r1 += 8; r2 += 8; outptr0 += 16; outptr1 += 16; } for (; j + 1 < outw; j += 2) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _sum10 = _mm_loadu_ps(outptr1); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10); _mm_storeu_ps(outptr0, _sum00); _mm_storeu_ps(outptr1, _sum10); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _sum11 = _mm_loadu_ps(outptr1 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01); _sum11 = _mm_comp_fmadd_ps(_r03, _k00_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r04, _k01_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r05, _k02_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r13, _k10_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r14, _k11_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r15, _k12_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r23, _k20_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r24, _k21_1, _sum11); _sum11 = _mm_comp_fmadd_ps(_r25, _k22_1, _sum11); _mm_storeu_ps(outptr0 + 4, _sum01); _mm_storeu_ps(outptr1 + 4, _sum11); r0 += 4; r1 += 4; r2 += 4; outptr0 += 8; outptr1 += 8; } for (; j < outw; j++) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _sum10 = _mm_loadu_ps(outptr1); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10); _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10); _mm_storeu_ps(outptr0, _sum00); _mm_storeu_ps(outptr1, _sum10); r0 += 2; r1 += 2; r2 += 2; outptr0 += 4; outptr1 += 4; } r0 += tailstep; r1 += tailstep; r2 += tailstep; } k0 += 9 * 4; k1 += 9 * 4; } } #pragma omp parallel for num_threads(opt.num_threads) for (int p = remain_outch_start; p < outch; p++) { Mat out0 = top_blob.channel(p); __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f); out0.fill(_bias0); const float* k0 = kernel.channel(p); for (int q = 0; q < inch; q++) { float* outptr0 = out0.row(0); const Mat img0 = bottom_blob.channel(q); const float* r0 = img0.row(0); const float* r1 = img0.row(1); const float* r2 = img0.row(2); __m128 _k00_0 = _mm_loadu_ps(k0); __m128 _k01_0 = _mm_loadu_ps(k0 + 4); __m128 _k02_0 = _mm_loadu_ps(k0 + 8); __m128 _k10_0 = _mm_loadu_ps(k0 + 12); __m128 _k11_0 = _mm_loadu_ps(k0 + 16); __m128 _k12_0 = _mm_loadu_ps(k0 + 20); __m128 _k20_0 = _mm_loadu_ps(k0 + 24); __m128 _k21_0 = _mm_loadu_ps(k0 + 28); __m128 _k22_0 = _mm_loadu_ps(k0 + 32); int i = 0; for (; i < outh; i++) { int j = 0; for (; j + 7 < outw; j += 8) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _mm_storeu_ps(outptr0, _sum00); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01); _mm_storeu_ps(outptr0 + 4, _sum01); __m128 _sum02 = _mm_loadu_ps(outptr0 + 8); __m128 _r06 = _mm_set1_ps(*(r0 + 5)); __m128 _r16 = _mm_set1_ps(*(r1 + 5)); __m128 _r26 = _mm_set1_ps(*(r2 + 5)); __m128 _r07 = _mm_set1_ps(*(r0 + 6)); __m128 _r17 = _mm_set1_ps(*(r1 + 6)); __m128 _r27 = _mm_set1_ps(*(r2 + 6)); _sum02 = _mm_comp_fmadd_ps(_r05, _k00_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r06, _k01_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r07, _k02_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r15, _k10_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r16, _k11_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r17, _k12_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r25, _k20_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r26, _k21_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r27, _k22_0, _sum02); _mm_storeu_ps(outptr0 + 8, _sum02); __m128 _r08 = _mm_set1_ps(*(r0 + 7)); __m128 _r18 = _mm_set1_ps(*(r1 + 7)); __m128 _r28 = _mm_set1_ps(*(r2 + 7)); __m128 _r09 = _mm_set1_ps(*(r0 + 8)); __m128 _r19 = _mm_set1_ps(*(r1 + 8)); __m128 _r29 = _mm_set1_ps(*(r2 + 8)); __m128 _sum03 = _mm_loadu_ps(outptr0 + 12); _sum03 = _mm_comp_fmadd_ps(_r07, _k00_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r08, _k01_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r09, _k02_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r17, _k10_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r18, _k11_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r19, _k12_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r27, _k20_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r28, _k21_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r29, _k22_0, _sum03); _mm_storeu_ps(outptr0 + 12, _sum03); __m128 _r010 = _mm_set1_ps(*(r0 + 9)); __m128 _r110 = _mm_set1_ps(*(r1 + 9)); __m128 _r210 = _mm_set1_ps(*(r2 + 9)); __m128 _r011 = _mm_set1_ps(*(r0 + 10)); __m128 _r111 = _mm_set1_ps(*(r1 + 10)); __m128 _r211 = _mm_set1_ps(*(r2 + 10)); __m128 _sum04 = _mm_loadu_ps(outptr0 + 16); _sum04 = _mm_comp_fmadd_ps(_r09, _k00_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r010, _k01_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r011, _k02_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r19, _k10_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r110, _k11_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r111, _k12_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r29, _k20_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r210, _k21_0, _sum04); _sum04 = _mm_comp_fmadd_ps(_r211, _k22_0, _sum04); _mm_storeu_ps(outptr0 + 16, _sum04); __m128 _r012 = _mm_set1_ps(*(r0 + 11)); __m128 _r112 = _mm_set1_ps(*(r1 + 11)); __m128 _r212 = _mm_set1_ps(*(r2 + 11)); __m128 _r013 = _mm_set1_ps(*(r0 + 12)); __m128 _r113 = _mm_set1_ps(*(r1 + 12)); __m128 _r213 = _mm_set1_ps(*(r2 + 12)); __m128 _sum05 = _mm_loadu_ps(outptr0 + 20); _sum05 = _mm_comp_fmadd_ps(_r011, _k00_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r012, _k01_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r013, _k02_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r111, _k10_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r112, _k11_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r113, _k12_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r211, _k20_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r212, _k21_0, _sum05); _sum05 = _mm_comp_fmadd_ps(_r213, _k22_0, _sum05); _mm_storeu_ps(outptr0 + 20, _sum05); __m128 _r014 = _mm_set1_ps(*(r0 + 13)); __m128 _r114 = _mm_set1_ps(*(r1 + 13)); __m128 _r214 = _mm_set1_ps(*(r2 + 13)); __m128 _r015 = _mm_set1_ps(*(r0 + 14)); __m128 _r115 = _mm_set1_ps(*(r1 + 14)); __m128 _r215 = _mm_set1_ps(*(r2 + 14)); __m128 _sum06 = _mm_loadu_ps(outptr0 + 24); _sum06 = _mm_comp_fmadd_ps(_r013, _k00_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r014, _k01_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r015, _k02_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r113, _k10_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r114, _k11_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r115, _k12_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r213, _k20_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r214, _k21_0, _sum06); _sum06 = _mm_comp_fmadd_ps(_r215, _k22_0, _sum06); _mm_storeu_ps(outptr0 + 24, _sum06); __m128 _r016 = _mm_set1_ps(*(r0 + 15)); __m128 _r116 = _mm_set1_ps(*(r1 + 15)); __m128 _r216 = _mm_set1_ps(*(r2 + 15)); __m128 _r017 = _mm_set1_ps(*(r0 + 16)); __m128 _r117 = _mm_set1_ps(*(r1 + 16)); __m128 _r217 = _mm_set1_ps(*(r2 + 16)); __m128 _sum07 = _mm_loadu_ps(outptr0 + 28); _sum07 = _mm_comp_fmadd_ps(_r015, _k00_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r016, _k01_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r017, _k02_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r115, _k10_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r116, _k11_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r117, _k12_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r215, _k20_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r216, _k21_0, _sum07); _sum07 = _mm_comp_fmadd_ps(_r217, _k22_0, _sum07); _mm_storeu_ps(outptr0 + 28, _sum07); r0 += 16; r1 += 16; r2 += 16; outptr0 += 32; } for (; j + 3 < outw; j += 4) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _mm_storeu_ps(outptr0, _sum00); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01); _mm_storeu_ps(outptr0 + 4, _sum01); __m128 _sum02 = _mm_loadu_ps(outptr0 + 8); __m128 _r06 = _mm_set1_ps(*(r0 + 5)); __m128 _r16 = _mm_set1_ps(*(r1 + 5)); __m128 _r26 = _mm_set1_ps(*(r2 + 5)); __m128 _r07 = _mm_set1_ps(*(r0 + 6)); __m128 _r17 = _mm_set1_ps(*(r1 + 6)); __m128 _r27 = _mm_set1_ps(*(r2 + 6)); _sum02 = _mm_comp_fmadd_ps(_r05, _k00_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r06, _k01_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r07, _k02_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r15, _k10_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r16, _k11_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r17, _k12_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r25, _k20_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r26, _k21_0, _sum02); _sum02 = _mm_comp_fmadd_ps(_r27, _k22_0, _sum02); _mm_storeu_ps(outptr0 + 8, _sum02); __m128 _r08 = _mm_set1_ps(*(r0 + 7)); __m128 _r18 = _mm_set1_ps(*(r1 + 7)); __m128 _r28 = _mm_set1_ps(*(r2 + 7)); __m128 _r09 = _mm_set1_ps(*(r0 + 8)); __m128 _r19 = _mm_set1_ps(*(r1 + 8)); __m128 _r29 = _mm_set1_ps(*(r2 + 8)); __m128 _sum03 = _mm_loadu_ps(outptr0 + 12); _sum03 = _mm_comp_fmadd_ps(_r07, _k00_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r08, _k01_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r09, _k02_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r17, _k10_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r18, _k11_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r19, _k12_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r27, _k20_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r28, _k21_0, _sum03); _sum03 = _mm_comp_fmadd_ps(_r29, _k22_0, _sum03); _mm_storeu_ps(outptr0 + 12, _sum03); r0 += 8; r1 += 8; r2 += 8; outptr0 += 16; } for (; j + 1 < outw; j += 2) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _mm_storeu_ps(outptr0, _sum00); __m128 _sum01 = _mm_loadu_ps(outptr0 + 4); __m128 _r04 = _mm_set1_ps(*(r0 + 3)); __m128 _r14 = _mm_set1_ps(*(r1 + 3)); __m128 _r24 = _mm_set1_ps(*(r2 + 3)); __m128 _r05 = _mm_set1_ps(*(r0 + 4)); __m128 _r15 = _mm_set1_ps(*(r1 + 4)); __m128 _r25 = _mm_set1_ps(*(r2 + 4)); _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01); _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01); _mm_storeu_ps(outptr0 + 4, _sum01); r0 += 4; r1 += 4; r2 += 4; outptr0 += 8; } for (; j < outw; j++) { __m128 _sum00 = _mm_loadu_ps(outptr0); __m128 _r01 = _mm_set1_ps(*(r0)); __m128 _r02 = _mm_set1_ps(*(r0 + 1)); __m128 _r03 = _mm_set1_ps(*(r0 + 2)); __m128 _r11 = _mm_set1_ps(*(r1)); __m128 _r12 = _mm_set1_ps(*(r1 + 1)); __m128 _r13 = _mm_set1_ps(*(r1 + 2)); __m128 _r21 = _mm_set1_ps(*(r2)); __m128 _r22 = _mm_set1_ps(*(r2 + 1)); __m128 _r23 = _mm_set1_ps(*(r2 + 2)); _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00); _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00); _mm_storeu_ps(outptr0, _sum00); r0 += 2; r1 += 2; r2 += 2; outptr0 += 4; } r0 += tailstep; r1 += tailstep; r2 += tailstep; } k0 += 9 * 4; } } }