// Tencent is pleased to support the open source community by making ncnn available. // // Copyright (C) 2023 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 convolution1d_transform_kernel_packed_fp16s(const Mat& kernel, Mat& kernel_tm, int inh, int outh, int kernel_w) { // src = kw-inh-outh // dst = pb-pa-kw-inh/pa-outh/pb // clang-format off // *INDENT-OFF* if (outh >= 8) { if (inh >= 8) kernel_tm.create(8 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); else if (inh >= 4) kernel_tm.create(8 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); else if (inh >= 2) kernel_tm.create(8 * 2 * kernel_w, inh / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); else kernel_tm.create(8 * kernel_w, inh, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); } else if (outh >= 4) { if (inh >= 8) kernel_tm.create(4 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); else if (inh >= 4) kernel_tm.create(4 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); else if (inh >= 2) kernel_tm.create(4 * 2 * kernel_w, inh / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); else kernel_tm.create(4 * kernel_w, inh, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u); } else if (outh >= 2) { if (inh >= 8) kernel_tm.create(2 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u); else if (inh >= 4) kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u); else if (inh >= 2) kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u); else kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2, (size_t)2u); } else { if (inh >= 8) kernel_tm.create(8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u); else if (inh >= 4) kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u); else if (inh >= 2) kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh, (size_t)2u); else kernel_tm.create(kernel_w, inh, outh, (size_t)2u); } // *INDENT-ON* // clang-format on int q = 0; for (; q + 7 < outh; q += 8) { const float* kptr0 = (const float*)kernel + q * inh * kernel_w; const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w; const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w; const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w; const float* kptr4 = (const float*)kernel + (q + 4) * inh * kernel_w; const float* kptr5 = (const float*)kernel + (q + 5) * inh * kernel_w; const float* kptr6 = (const float*)kernel + (q + 6) * inh * kernel_w; const float* kptr7 = (const float*)kernel + (q + 7) * inh * kernel_w; __fp16* g00 = kernel_tm.channel(q / 8); int p = 0; for (; p + 7 < inh; p += 8) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; const float* k4 = kptr4 + p * kernel_w; const float* k5 = kptr5 + p * kernel_w; const float* k6 = kptr6 + p * kernel_w; const float* k7 = kptr7 + p * kernel_w; for (int i = 0; i < 8; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; g00[4] = (__fp16)k4[k]; g00[5] = (__fp16)k5[k]; g00[6] = (__fp16)k6[k]; g00[7] = (__fp16)k7[k]; k0 += kernel_w; k1 += kernel_w; k2 += kernel_w; k3 += kernel_w; k4 += kernel_w; k5 += kernel_w; k6 += kernel_w; k7 += kernel_w; g00 += 8; } } } for (; p + 3 < inh; p += 4) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; const float* k4 = kptr4 + p * kernel_w; const float* k5 = kptr5 + p * kernel_w; const float* k6 = kptr6 + p * kernel_w; const float* k7 = kptr7 + p * kernel_w; for (int i = 0; i < 4; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; g00[4] = (__fp16)k4[k]; g00[5] = (__fp16)k5[k]; g00[6] = (__fp16)k6[k]; g00[7] = (__fp16)k7[k]; k0 += kernel_w; k1 += kernel_w; k2 += kernel_w; k3 += kernel_w; k4 += kernel_w; k5 += kernel_w; k6 += kernel_w; k7 += kernel_w; g00 += 8; } } } for (; p + 1 < inh; p += 2) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; const float* k4 = kptr4 + p * kernel_w; const float* k5 = kptr5 + p * kernel_w; const float* k6 = kptr6 + p * kernel_w; const float* k7 = kptr7 + p * kernel_w; for (int i = 0; i < 2; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; g00[4] = (__fp16)k4[k]; g00[5] = (__fp16)k5[k]; g00[6] = (__fp16)k6[k]; g00[7] = (__fp16)k7[k]; k0 += kernel_w; k1 += kernel_w; k2 += kernel_w; k3 += kernel_w; k4 += kernel_w; k5 += kernel_w; k6 += kernel_w; k7 += kernel_w; g00 += 8; } } } for (; p < inh; p++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; const float* k4 = kptr4 + p * kernel_w; const float* k5 = kptr5 + p * kernel_w; const float* k6 = kptr6 + p * kernel_w; const float* k7 = kptr7 + p * kernel_w; for (int k = 0; k < kernel_w; k++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; g00[4] = (__fp16)k4[k]; g00[5] = (__fp16)k5[k]; g00[6] = (__fp16)k6[k]; g00[7] = (__fp16)k7[k]; g00 += 8; } } } for (; q + 3 < outh; q += 4) { const float* kptr0 = (const float*)kernel + q * inh * kernel_w; const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w; const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w; const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w; __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4); int p = 0; for (; p + 7 < inh; p += 8) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; for (int i = 0; i < 8; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; k0 += kernel_w; k1 += kernel_w; k2 += kernel_w; k3 += kernel_w; g00 += 4; } } } for (; p + 3 < inh; p += 4) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; for (int i = 0; i < 4; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; k0 += kernel_w; k1 += kernel_w; k2 += kernel_w; k3 += kernel_w; g00 += 4; } } } for (; p + 1 < inh; p += 2) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; for (int i = 0; i < 2; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; k0 += kernel_w; k1 += kernel_w; k2 += kernel_w; k3 += kernel_w; g00 += 4; } } } for (; p < inh; p++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; const float* k2 = kptr2 + p * kernel_w; const float* k3 = kptr3 + p * kernel_w; for (int k = 0; k < kernel_w; k++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00[2] = (__fp16)k2[k]; g00[3] = (__fp16)k3[k]; g00 += 4; } } } for (; q + 1 < outh; q += 2) { const float* kptr0 = (const float*)kernel + q * inh * kernel_w; const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w; __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2); int p = 0; for (; p + 7 < inh; p += 8) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w + k; const float* k1 = kptr1 + p * kernel_w + k; g00[0] = (__fp16)k0[0]; g00[1] = (__fp16)k0[kernel_w]; g00[2] = (__fp16)k0[kernel_w * 2]; g00[3] = (__fp16)k0[kernel_w * 3]; g00[4] = (__fp16)k0[kernel_w * 4]; g00[5] = (__fp16)k0[kernel_w * 5]; g00[6] = (__fp16)k0[kernel_w * 6]; g00[7] = (__fp16)k0[kernel_w * 7]; g00[8] = (__fp16)k1[0]; g00[9] = (__fp16)k1[kernel_w]; g00[10] = (__fp16)k1[kernel_w * 2]; g00[11] = (__fp16)k1[kernel_w * 3]; g00[12] = (__fp16)k1[kernel_w * 4]; g00[13] = (__fp16)k1[kernel_w * 5]; g00[14] = (__fp16)k1[kernel_w * 6]; g00[15] = (__fp16)k1[kernel_w * 7]; g00 += 16; } } for (; p + 3 < inh; p += 4) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w + k; const float* k1 = kptr1 + p * kernel_w + k; g00[0] = (__fp16)k0[0]; g00[1] = (__fp16)k0[kernel_w]; g00[2] = (__fp16)k0[kernel_w * 2]; g00[3] = (__fp16)k0[kernel_w * 3]; g00[4] = (__fp16)k1[0]; g00[5] = (__fp16)k1[kernel_w]; g00[6] = (__fp16)k1[kernel_w * 2]; g00[7] = (__fp16)k1[kernel_w * 3]; g00 += 8; } } for (; p + 1 < inh; p += 2) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; for (int i = 0; i < 2; i++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; k0 += kernel_w; k1 += kernel_w; g00 += 2; } } } for (; p < inh; p++) { const float* k0 = kptr0 + p * kernel_w; const float* k1 = kptr1 + p * kernel_w; for (int k = 0; k < kernel_w; k++) { g00[0] = (__fp16)k0[k]; g00[1] = (__fp16)k1[k]; g00 += 2; } } } for (; q < outh; q++) { const float* kptr = (const float*)kernel + q * inh * kernel_w; __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2); int p = 0; for (; p + 7 < inh; p += 8) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr + p * kernel_w; for (int i = 0; i < 8; i++) { g00[0] = (__fp16)k0[k]; k0 += kernel_w; g00 += 1; } } } for (; p + 3 < inh; p += 4) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr + p * kernel_w; for (int i = 0; i < 4; i++) { g00[0] = (__fp16)k0[k]; k0 += kernel_w; g00 += 1; } } } for (; p + 1 < inh; p += 2) { for (int k = 0; k < kernel_w; k++) { const float* k0 = kptr + p * kernel_w; for (int i = 0; i < 2; i++) { g00[0] = (__fp16)k0[k]; k0 += kernel_w; g00 += 1; } } } for (; p < inh; p++) { const float* k0 = kptr + p * kernel_w; for (int k = 0; k < kernel_w; k++) { g00[0] = (__fp16)k0[k]; g00++; } } } } static void convolution1d_packed_fp16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt) { const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int N = bottom_blob.w * elempack; const int outw = top_blob.w; const int out_elempack = top_blob.elempack; const int outh = top_blob.h * out_elempack; const int M = top_blob.w * out_elempack; const float* bias_data_ptr = bias_data; int nn_outh = 0; int remain_outh_start = 0; nn_outh = (outh - remain_outh_start) / 8; #pragma omp parallel for num_threads(opt.num_threads) for (int pp = 0; pp < nn_outh; pp++) { const int p = remain_outh_start + pp * 8; // shadowed variable for less openmp task args const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int outw = top_blob.w; const int out_elempack = top_blob.elempack; __fp16* outptr = top_blob.row<__fp16>(p / out_elempack); for (int j = 0; j < outw; j++) { float32x4_t _sum0 = vdupq_n_f32(0.f); float32x4_t _sum1 = vdupq_n_f32(0.f); float32x4_t _sum2 = vdupq_n_f32(0.f); float32x4_t _sum3 = vdupq_n_f32(0.f); float32x4_t _sum4 = vdupq_n_f32(0.f); float32x4_t _sum5 = vdupq_n_f32(0.f); float32x4_t _sum6 = vdupq_n_f32(0.f); float32x4_t _sum7 = vdupq_n_f32(0.f); if (bias_data_ptr) { _sum0 = vld1q_f32(bias_data_ptr + p); _sum1 = vld1q_f32(bias_data_ptr + p + 4); } const __fp16* kptr = weight_data_tm.channel(p / 8); int q = 0; for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; float32x4_t _r1; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); _r1 = vcvt_f32_f16(vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x8_t _r_f16 = float16x8_t(); _r_f16 = vsetq_lane_f16(r0[0], _r_f16, 0); _r_f16 = vsetq_lane_f16(r0[N], _r_f16, 1); _r_f16 = vsetq_lane_f16(r0[N * 2], _r_f16, 2); _r_f16 = vsetq_lane_f16(r0[N * 3], _r_f16, 3); _r_f16 = vsetq_lane_f16(r0[N * 4], _r_f16, 4); _r_f16 = vsetq_lane_f16(r0[N * 5], _r_f16, 5); _r_f16 = vsetq_lane_f16(r0[N * 6], _r_f16, 6); _r_f16 = vsetq_lane_f16(r0[N * 7], _r_f16, 7); _r0 = vcvt_f32_f16(vget_low_f16(_r_f16)); _r1 = vcvt_f32_f16(vget_high_f16(_r_f16)); r0 += dilation_w; } float16x8_t _w01 = vld1q_f16(kptr); float16x8_t _w23 = vld1q_f16(kptr + 8); float16x8_t _w45 = vld1q_f16(kptr + 16); float16x8_t _w67 = vld1q_f16(kptr + 24); float16x8_t _w89 = vld1q_f16(kptr + 32); float16x8_t _wab = vld1q_f16(kptr + 40); float16x8_t _wcd = vld1q_f16(kptr + 48); float16x8_t _wef = vld1q_f16(kptr + 56); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01)); float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23)); float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23)); float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45)); float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45)); float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67)); float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67)); float32x4_t _w8 = vcvt_f32_f16(vget_low_f16(_w89)); float32x4_t _w9 = vcvt_f32_f16(vget_high_f16(_w89)); float32x4_t _wa = vcvt_f32_f16(vget_low_f16(_wab)); float32x4_t _wb = vcvt_f32_f16(vget_high_f16(_wab)); float32x4_t _wc = vcvt_f32_f16(vget_low_f16(_wcd)); float32x4_t _wd = vcvt_f32_f16(vget_high_f16(_wcd)); float32x4_t _we = vcvt_f32_f16(vget_low_f16(_wef)); float32x4_t _wf = vcvt_f32_f16(vget_high_f16(_wef)); _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0); _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0); _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1); _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1); _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2); _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2); _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3); _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3); _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0); _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0); _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1); _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1); _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2); _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2); _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3); _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3); kptr += 64; } } for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x4_t _r_f16 = float16x4_t(); _r_f16 = vset_lane_f16(r0[0], _r_f16, 0); _r_f16 = vset_lane_f16(r0[N], _r_f16, 1); _r_f16 = vset_lane_f16(r0[N * 2], _r_f16, 2); _r_f16 = vset_lane_f16(r0[N * 3], _r_f16, 3); _r0 = vcvt_f32_f16(_r_f16); r0 += dilation_w; } float16x8_t _w01 = vld1q_f16(kptr); float16x8_t _w23 = vld1q_f16(kptr + 8); float16x8_t _w45 = vld1q_f16(kptr + 16); float16x8_t _w67 = vld1q_f16(kptr + 24); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01)); float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23)); float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23)); float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45)); float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45)); float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67)); float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67)); _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0); _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0); _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1); _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1); _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2); _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2); _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3); _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3); kptr += 32; } } for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float val0; float val1; // if (elempack == 1) { val0 = (float)(r0[0]); val1 = (float)(r0[N]); r0 += dilation_w; } float16x8_t _w01 = vld1q_f16(kptr); float16x8_t _w23 = vld1q_f16(kptr + 8); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01)); float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23)); float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23)); _sum0 = vfmaq_n_f32(_sum0, _w0, val0); _sum1 = vfmaq_n_f32(_sum1, _w1, val0); _sum2 = vfmaq_n_f32(_sum2, _w2, val1); _sum3 = vfmaq_n_f32(_sum3, _w3, val1); kptr += 16; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float32x4_t _val; // if (elempack == 1) { _val = vcvt_f32_f16(vdup_n_f16(r0[0])); r0 += dilation_w; } float16x8_t _w = vld1q_f16(kptr); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w)); _sum0 = vfmaq_f32(_sum0, _w0, _val); _sum1 = vfmaq_f32(_sum1, _w1, _val); kptr += 8; } } _sum0 = vaddq_f32(_sum0, _sum2); _sum1 = vaddq_f32(_sum1, _sum3); _sum4 = vaddq_f32(_sum4, _sum6); _sum5 = vaddq_f32(_sum5, _sum7); _sum0 = vaddq_f32(_sum0, _sum4); _sum1 = vaddq_f32(_sum1, _sum5); _sum0 = activation_ps(_sum0, activation_type, activation_params); _sum1 = activation_ps(_sum1, activation_type, activation_params); if (out_elempack == 4) { vst1_f16(outptr, vcvt_f16_f32(_sum0)); vst1_f16(outptr + M, vcvt_f16_f32(_sum1)); outptr += 4; } else // if (out_elempack == 1) { float16x4_t _sum0_f16 = vcvt_f16_f32(_sum0); float16x4_t _sum1_f16 = vcvt_f16_f32(_sum1); outptr[0] = vget_lane_f16(_sum0_f16, 0); outptr[M] = vget_lane_f16(_sum0_f16, 1); outptr[M * 2] = vget_lane_f16(_sum0_f16, 2); outptr[M * 3] = vget_lane_f16(_sum0_f16, 3); outptr[M * 4] = vget_lane_f16(_sum1_f16, 0); outptr[M * 5] = vget_lane_f16(_sum1_f16, 1); outptr[M * 6] = vget_lane_f16(_sum1_f16, 2); outptr[M * 7] = vget_lane_f16(_sum1_f16, 3); outptr += 1; } } } remain_outh_start += nn_outh * 8; nn_outh = (outh - remain_outh_start) / 4; for (int pp = 0; pp < nn_outh; pp++) { const int p = remain_outh_start + pp * 4; // shadowed variable for less openmp task args const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int outw = top_blob.w; const int out_elempack = top_blob.elempack; __fp16* outptr = top_blob.row<__fp16>(p / out_elempack); for (int j = 0; j < outw; j++) { float32x4_t _sum0 = vdupq_n_f32(0.f); float32x4_t _sum1 = vdupq_n_f32(0.f); float32x4_t _sum2 = vdupq_n_f32(0.f); float32x4_t _sum3 = vdupq_n_f32(0.f); if (bias_data_ptr) { _sum0 = vld1q_f32(bias_data_ptr + p); } const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4); int q = 0; for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; float32x4_t _r1; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); _r1 = vcvt_f32_f16(vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x8_t _r_f16 = float16x8_t(); _r_f16 = vsetq_lane_f16(r0[0], _r_f16, 0); _r_f16 = vsetq_lane_f16(r0[N], _r_f16, 1); _r_f16 = vsetq_lane_f16(r0[N * 2], _r_f16, 2); _r_f16 = vsetq_lane_f16(r0[N * 3], _r_f16, 3); _r_f16 = vsetq_lane_f16(r0[N * 4], _r_f16, 4); _r_f16 = vsetq_lane_f16(r0[N * 5], _r_f16, 5); _r_f16 = vsetq_lane_f16(r0[N * 6], _r_f16, 6); _r_f16 = vsetq_lane_f16(r0[N * 7], _r_f16, 7); _r0 = vcvt_f32_f16(vget_low_f16(_r_f16)); _r1 = vcvt_f32_f16(vget_high_f16(_r_f16)); r0 += dilation_w; } float16x8_t _w01 = vld1q_f16(kptr); float16x8_t _w23 = vld1q_f16(kptr + 8); float16x8_t _w45 = vld1q_f16(kptr + 16); float16x8_t _w67 = vld1q_f16(kptr + 24); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01)); float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23)); float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23)); float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45)); float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45)); float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67)); float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67)); _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0); _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1); _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2); _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3); _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0); _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1); _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2); _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3); kptr += 32; } } for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x4_t _r_f16 = float16x4_t(); _r_f16 = vset_lane_f16(r0[0], _r_f16, 0); _r_f16 = vset_lane_f16(r0[N], _r_f16, 1); _r_f16 = vset_lane_f16(r0[N * 2], _r_f16, 2); _r_f16 = vset_lane_f16(r0[N * 3], _r_f16, 3); _r0 = vcvt_f32_f16(_r_f16); r0 += dilation_w; } float16x8_t _w01 = vld1q_f16(kptr); float16x8_t _w23 = vld1q_f16(kptr + 8); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01)); float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23)); float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23)); _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0); _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1); _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2); _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3); kptr += 16; } } for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float val0; float val1; // if (elempack == 1) { val0 = (float)(r0[0]); val1 = (float)(r0[N]); r0 += dilation_w; } float16x8_t _w = vld1q_f16(kptr); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w)); _sum0 = vfmaq_n_f32(_sum0, _w0, val0); _sum1 = vfmaq_n_f32(_sum1, _w1, val1); kptr += 8; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float32x4_t _val; // if (elempack == 1) { _val = vcvt_f32_f16(vdup_n_f16(r0[0])); r0 += dilation_w; } float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr)); _sum0 = vfmaq_f32(_sum0, _val, _w); kptr += 4; } } _sum0 = vaddq_f32(_sum0, _sum1); _sum2 = vaddq_f32(_sum2, _sum3); _sum0 = vaddq_f32(_sum0, _sum2); _sum0 = activation_ps(_sum0, activation_type, activation_params); if (out_elempack == 4) { vst1_f16(outptr, vcvt_f16_f32(_sum0)); outptr += 4; } else // if (out_elempack == 1) { float16x4_t _sum0_f16 = vcvt_f16_f32(_sum0); outptr[0] = vget_lane_f16(_sum0_f16, 0); outptr[M] = vget_lane_f16(_sum0_f16, 1); outptr[M * 2] = vget_lane_f16(_sum0_f16, 2); outptr[M * 3] = vget_lane_f16(_sum0_f16, 3); outptr += 1; } } } remain_outh_start += nn_outh * 4; nn_outh = (outh - remain_outh_start) / 2; for (int pp = 0; pp < nn_outh; pp++) { const int p = remain_outh_start + pp * 2; // shadowed variable for less openmp task args const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int outw = top_blob.w; __fp16* outptr0 = top_blob.row<__fp16>(p); __fp16* outptr1 = top_blob.row<__fp16>(p + 1); for (int j = 0; j < outw; j++) { float sum0 = 0.f; float sum1 = 0.f; if (bias_data_ptr) { sum0 = bias_data_ptr[p]; sum1 = bias_data_ptr[p + 1]; } const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2); int q = 0; float32x4_t _sum0 = vdupq_n_f32(0.f); float32x4_t _sum1 = vdupq_n_f32(0.f); float32x4_t _sum2 = vdupq_n_f32(0.f); float32x4_t _sum3 = vdupq_n_f32(0.f); for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; float32x4_t _r1; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); _r1 = vcvt_f32_f16(vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x8_t _r01_f16 = float16x8_t(); _r01_f16 = vsetq_lane_f16(r0[0], _r01_f16, 0); _r01_f16 = vsetq_lane_f16(r0[N], _r01_f16, 1); _r01_f16 = vsetq_lane_f16(r0[N * 2], _r01_f16, 2); _r01_f16 = vsetq_lane_f16(r0[N * 3], _r01_f16, 3); _r01_f16 = vsetq_lane_f16(r0[N * 4], _r01_f16, 4); _r01_f16 = vsetq_lane_f16(r0[N * 5], _r01_f16, 5); _r01_f16 = vsetq_lane_f16(r0[N * 6], _r01_f16, 6); _r01_f16 = vsetq_lane_f16(r0[N * 7], _r01_f16, 7); _r0 = vcvt_f32_f16(vget_low_f16(_r01_f16)); _r1 = vcvt_f32_f16(vget_high_f16(_r01_f16)); r0 += dilation_w; } float16x8_t _w01 = vld1q_f16(kptr); float16x8_t _w23 = vld1q_f16(kptr + 8); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01)); float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23)); float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23)); _sum0 = vfmaq_f32(_sum0, _r0, _w0); _sum1 = vfmaq_f32(_sum1, _r1, _w1); _sum2 = vfmaq_f32(_sum2, _r0, _w2); _sum3 = vfmaq_f32(_sum3, _r1, _w3); kptr += 16; } } _sum0 = vaddq_f32(_sum0, _sum1); _sum2 = vaddq_f32(_sum2, _sum3); sum0 += vaddvq_f32(_sum0); sum1 += vaddvq_f32(_sum2); _sum0 = vdupq_n_f32(0.f); _sum1 = vdupq_n_f32(0.f); for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x4_t _r0_f16 = float16x4_t(); _r0_f16 = vset_lane_f16(r0[0], _r0_f16, 0); _r0_f16 = vset_lane_f16(r0[N], _r0_f16, 1); _r0_f16 = vset_lane_f16(r0[N * 2], _r0_f16, 2); _r0_f16 = vset_lane_f16(r0[N * 3], _r0_f16, 3); _r0 = vcvt_f32_f16(_r0_f16); r0 += dilation_w; } float16x8_t _w = vld1q_f16(kptr); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w)); _sum0 = vfmaq_f32(_sum0, _r0, _w0); _sum1 = vfmaq_f32(_sum1, _r0, _w1); kptr += 8; } } sum0 += vaddvq_f32(_sum0); sum1 += vaddvq_f32(_sum1); for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float val0; float val1; // if (elempack == 1) { val0 = (float)(r0[0]); val1 = (float)(r0[N]); r0 += dilation_w; } sum0 += val0 * (float)(kptr[0]); sum1 += val0 * (float)(kptr[1]); sum0 += val1 * (float)(kptr[2]); sum1 += val1 * (float)(kptr[3]); kptr += 4; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float val; // if (elempack == 1) { val = (float)(r0[0]); r0 += dilation_w; } sum0 += val * (float)(kptr[0]); sum1 += val * (float)(kptr[1]); kptr += 2; } } sum0 = activation_ss(sum0, activation_type, activation_params); sum1 = activation_ss(sum1, activation_type, activation_params); outptr0[0] = (__fp16)(sum0); outptr1[0] = (__fp16)(sum1); outptr0 += 1; outptr1 += 1; } } remain_outh_start += nn_outh * 2; for (int p = remain_outh_start; p < outh; p++) { __fp16* outptr = top_blob.row<__fp16>(p); for (int j = 0; j < outw; j++) { float sum = 0.f; if (bias_data_ptr) { sum = bias_data_ptr[p]; } const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2); int q = 0; float32x4_t _sum0 = vdupq_n_f32(0.f); float32x4_t _sum1 = vdupq_n_f32(0.f); for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; float32x4_t _r1; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); _r1 = vcvt_f32_f16(vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x8_t _r01_f16 = float16x8_t(); _r01_f16 = vsetq_lane_f16(r0[0], _r01_f16, 0); _r01_f16 = vsetq_lane_f16(r0[N], _r01_f16, 1); _r01_f16 = vsetq_lane_f16(r0[N * 2], _r01_f16, 2); _r01_f16 = vsetq_lane_f16(r0[N * 3], _r01_f16, 3); _r01_f16 = vsetq_lane_f16(r0[N * 4], _r01_f16, 4); _r01_f16 = vsetq_lane_f16(r0[N * 5], _r01_f16, 5); _r01_f16 = vsetq_lane_f16(r0[N * 6], _r01_f16, 6); _r01_f16 = vsetq_lane_f16(r0[N * 7], _r01_f16, 7); _r0 = vcvt_f32_f16(vget_low_f16(_r01_f16)); _r1 = vcvt_f32_f16(vget_high_f16(_r01_f16)); r0 += dilation_w; } float16x8_t _w = vld1q_f16(kptr); float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w)); float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w)); _sum0 = vfmaq_f32(_sum0, _r0, _w0); _sum1 = vfmaq_f32(_sum1, _r1, _w1); kptr += 8; } } _sum0 = vaddq_f32(_sum0, _sum1); sum += vaddvq_f32(_sum0); float32x4_t _sum = vdupq_n_f32(0.f); for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float32x4_t _r0; if (elempack == 4) { _r0 = vcvt_f32_f16(vld1_f16(r0)); r0 += dilation_w * 4; } else // if (elempack == 1) { float16x4_t _r0_f16 = float16x4_t(); _r0_f16 = vset_lane_f16(r0[0], _r0_f16, 0); _r0_f16 = vset_lane_f16(r0[N], _r0_f16, 1); _r0_f16 = vset_lane_f16(r0[N * 2], _r0_f16, 2); _r0_f16 = vset_lane_f16(r0[N * 3], _r0_f16, 3); _r0 = vcvt_f32_f16(_r0_f16); r0 += dilation_w; } float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr)); _sum = vfmaq_f32(_sum, _r0, _w); kptr += 4; } } sum += vaddvq_f32(_sum); for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float val0; float val1; // if (elempack == 1) { val0 = (float)(r0[0]); val1 = (float)(r0[N]); r0 += dilation_w; } sum += val0 * (float)(kptr[0]); sum += val1 * (float)(kptr[1]); kptr += 2; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float val; // if (elempack == 1) { val = (float)(r0[0]); r0 += dilation_w; } sum += val * (float)(kptr[0]); kptr += 1; } } sum = activation_ss(sum, activation_type, activation_params); outptr[0] = (__fp16)(sum); outptr += 1; } } } static void convolution1d_packed_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt) { const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int N = bottom_blob.w * elempack; const int outw = top_blob.w; const int out_elempack = top_blob.elempack; const int outh = top_blob.h * out_elempack; const int M = top_blob.w * out_elempack; const __fp16* bias_data_ptr = bias_data; int nn_outh = 0; int remain_outh_start = 0; nn_outh = (outh - remain_outh_start) / 8; #pragma omp parallel for num_threads(opt.num_threads) for (int pp = 0; pp < nn_outh; pp++) { const int p = remain_outh_start + pp * 8; // shadowed variable for less openmp task args const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int outw = top_blob.w; const int out_elempack = top_blob.elempack; __fp16* outptr = top_blob.row<__fp16>(p / out_elempack); for (int j = 0; j < outw; j++) { float16x8_t _sum0 = vdupq_n_f16(0.f); float16x8_t _sum1 = vdupq_n_f16(0.f); float16x8_t _sum2 = vdupq_n_f16(0.f); float16x8_t _sum3 = vdupq_n_f16(0.f); if (bias_data_ptr) { _sum0 = vld1q_f16(bias_data_ptr + p); } const __fp16* kptr = weight_data_tm.channel(p / 8); int q = 0; for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x8_t _r0; if (elempack == 8) { _r0 = vld1q_f16(r0); r0 += dilation_w * 8; } else if (elempack == 4) { _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x8_t(); _r0 = vsetq_lane_f16(r0[0], _r0, 0); _r0 = vsetq_lane_f16(r0[N], _r0, 1); _r0 = vsetq_lane_f16(r0[N * 2], _r0, 2); _r0 = vsetq_lane_f16(r0[N * 3], _r0, 3); _r0 = vsetq_lane_f16(r0[N * 4], _r0, 4); _r0 = vsetq_lane_f16(r0[N * 5], _r0, 5); _r0 = vsetq_lane_f16(r0[N * 6], _r0, 6); _r0 = vsetq_lane_f16(r0[N * 7], _r0, 7); r0 += dilation_w; } float16x8_t _w0 = vld1q_f16(kptr); float16x8_t _w1 = vld1q_f16(kptr + 8); float16x8_t _w2 = vld1q_f16(kptr + 8 * 2); float16x8_t _w3 = vld1q_f16(kptr + 8 * 3); float16x8_t _w4 = vld1q_f16(kptr + 8 * 4); float16x8_t _w5 = vld1q_f16(kptr + 8 * 5); float16x8_t _w6 = vld1q_f16(kptr + 8 * 6); float16x8_t _w7 = vld1q_f16(kptr + 8 * 7); _sum0 = vfmaq_laneq_f16(_sum0, _w0, _r0, 0); _sum1 = vfmaq_laneq_f16(_sum1, _w1, _r0, 1); _sum2 = vfmaq_laneq_f16(_sum2, _w2, _r0, 2); _sum3 = vfmaq_laneq_f16(_sum3, _w3, _r0, 3); _sum0 = vfmaq_laneq_f16(_sum0, _w4, _r0, 4); _sum1 = vfmaq_laneq_f16(_sum1, _w5, _r0, 5); _sum2 = vfmaq_laneq_f16(_sum2, _w6, _r0, 6); _sum3 = vfmaq_laneq_f16(_sum3, _w7, _r0, 7); kptr += 64; } } for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x4_t _r0; if (elempack == 4) { _r0 = vld1_f16(r0); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x4_t(); _r0 = vset_lane_f16(r0[0], _r0, 0); _r0 = vset_lane_f16(r0[N], _r0, 1); _r0 = vset_lane_f16(r0[N * 2], _r0, 2); _r0 = vset_lane_f16(r0[N * 3], _r0, 3); r0 += dilation_w; } float16x8_t _w0 = vld1q_f16(kptr); float16x8_t _w1 = vld1q_f16(kptr + 8); float16x8_t _w2 = vld1q_f16(kptr + 8 * 2); float16x8_t _w3 = vld1q_f16(kptr + 8 * 3); _sum0 = vfmaq_lane_f16(_sum0, _w0, _r0, 0); _sum1 = vfmaq_lane_f16(_sum1, _w1, _r0, 1); _sum2 = vfmaq_lane_f16(_sum2, _w2, _r0, 2); _sum3 = vfmaq_lane_f16(_sum3, _w3, _r0, 3); kptr += 32; } } for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { __fp16 val0; __fp16 val1; // if (elempack == 1) { val0 = r0[0]; val1 = r0[N]; r0 += dilation_w; } float16x8_t _w0 = vld1q_f16(kptr); float16x8_t _w1 = vld1q_f16(kptr + 8); _sum0 = vfmaq_n_f16(_sum0, _w0, val0); _sum1 = vfmaq_n_f16(_sum1, _w1, val1); kptr += 16; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float16x8_t _val; // if (elempack == 1) { _val = vdupq_n_f16(r0[0]); r0 += dilation_w; } float16x8_t _w0 = vld1q_f16(kptr); _sum0 = vfmaq_f16(_sum0, _w0, _val); kptr += 8; } } _sum0 = vaddq_f16(_sum0, _sum1); _sum2 = vaddq_f16(_sum2, _sum3); _sum0 = vaddq_f16(_sum0, _sum2); _sum0 = activation_ps(_sum0, activation_type, activation_params); if (out_elempack == 8) { vst1q_f16(outptr, _sum0); outptr += 8; } else if (out_elempack == 4) { vst1_f16(outptr, vget_low_f16(_sum0)); vst1_f16(outptr + M, vget_high_f16(_sum0)); outptr += 4; } else // if (out_elempack == 1) { outptr[0] = vgetq_lane_f16(_sum0, 0); outptr[M] = vgetq_lane_f16(_sum0, 1); outptr[M * 2] = vgetq_lane_f16(_sum0, 2); outptr[M * 3] = vgetq_lane_f16(_sum0, 3); outptr[M * 4] = vgetq_lane_f16(_sum0, 4); outptr[M * 5] = vgetq_lane_f16(_sum0, 5); outptr[M * 6] = vgetq_lane_f16(_sum0, 6); outptr[M * 7] = vgetq_lane_f16(_sum0, 7); outptr += 1; } } } remain_outh_start += nn_outh * 8; nn_outh = (outh - remain_outh_start) / 4; for (int pp = 0; pp < nn_outh; pp++) { const int p = remain_outh_start + pp * 4; // shadowed variable for less openmp task args const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int outw = top_blob.w; const int out_elempack = top_blob.elempack; __fp16* outptr = top_blob.row<__fp16>(p / out_elempack); for (int j = 0; j < outw; j++) { float16x4_t _sum0 = vdup_n_f16(0.f); float16x4_t _sum1 = vdup_n_f16(0.f); float16x4_t _sum2 = vdup_n_f16(0.f); float16x4_t _sum3 = vdup_n_f16(0.f); if (bias_data_ptr) { _sum0 = vld1_f16(bias_data_ptr + p); } const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4); int q = 0; for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x4_t _r0; float16x4_t _r1; if (elempack == 8) { float16x8_t _r01 = vld1q_f16(r0); _r0 = vget_low_f16(_r01); _r1 = vget_high_f16(_r01); r0 += dilation_w * 8; } else if (elempack == 4) { _r0 = vld1_f16(r0); _r1 = vld1_f16(r0 + N); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x4_t(); _r1 = float16x4_t(); _r0 = vset_lane_f16(r0[0], _r0, 0); _r0 = vset_lane_f16(r0[N], _r0, 1); _r0 = vset_lane_f16(r0[N * 2], _r0, 2); _r0 = vset_lane_f16(r0[N * 3], _r0, 3); _r1 = vset_lane_f16(r0[N * 4], _r1, 0); _r1 = vset_lane_f16(r0[N * 5], _r1, 1); _r1 = vset_lane_f16(r0[N * 6], _r1, 2); _r1 = vset_lane_f16(r0[N * 7], _r1, 3); r0 += dilation_w; } float16x4_t _w0 = vld1_f16(kptr); float16x4_t _w1 = vld1_f16(kptr + 4); float16x4_t _w2 = vld1_f16(kptr + 8); float16x4_t _w3 = vld1_f16(kptr + 12); float16x4_t _w4 = vld1_f16(kptr + 16); float16x4_t _w5 = vld1_f16(kptr + 20); float16x4_t _w6 = vld1_f16(kptr + 24); float16x4_t _w7 = vld1_f16(kptr + 28); _sum0 = vfma_lane_f16(_sum0, _w0, _r0, 0); _sum1 = vfma_lane_f16(_sum1, _w1, _r0, 1); _sum2 = vfma_lane_f16(_sum2, _w2, _r0, 2); _sum3 = vfma_lane_f16(_sum3, _w3, _r0, 3); _sum0 = vfma_lane_f16(_sum0, _w4, _r1, 0); _sum1 = vfma_lane_f16(_sum1, _w5, _r1, 1); _sum2 = vfma_lane_f16(_sum2, _w6, _r1, 2); _sum3 = vfma_lane_f16(_sum3, _w7, _r1, 3); kptr += 32; } } for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x4_t _r0; if (elempack == 4) { _r0 = vld1_f16(r0); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x4_t(); _r0 = vset_lane_f16(r0[0], _r0, 0); _r0 = vset_lane_f16(r0[N], _r0, 1); _r0 = vset_lane_f16(r0[N * 2], _r0, 2); _r0 = vset_lane_f16(r0[N * 3], _r0, 3); r0 += dilation_w; } float16x4_t _w0 = vld1_f16(kptr); float16x4_t _w1 = vld1_f16(kptr + 4); float16x4_t _w2 = vld1_f16(kptr + 8); float16x4_t _w3 = vld1_f16(kptr + 12); _sum0 = vfma_lane_f16(_sum0, _w0, _r0, 0); _sum1 = vfma_lane_f16(_sum1, _w1, _r0, 1); _sum2 = vfma_lane_f16(_sum2, _w2, _r0, 2); _sum3 = vfma_lane_f16(_sum3, _w3, _r0, 3); kptr += 16; } } for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { __fp16 val0; __fp16 val1; // if (elempack == 1) { val0 = r0[0]; val1 = r0[N]; r0 += dilation_w; } float16x4_t _w0 = vld1_f16(kptr); float16x4_t _w1 = vld1_f16(kptr + 4); _sum0 = vfma_n_f16(_sum0, _w0, val0); _sum1 = vfma_n_f16(_sum1, _w1, val1); kptr += 8; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { float16x4_t _val; // if (elempack == 1) { _val = vdup_n_f16(r0[0]); r0 += dilation_w; } float16x4_t _w = vld1_f16(kptr); _sum0 = vfma_f16(_sum0, _val, _w); kptr += 4; } } _sum0 = vadd_f16(_sum0, _sum1); _sum2 = vadd_f16(_sum2, _sum3); _sum0 = vadd_f16(_sum0, _sum2); _sum0 = activation_ps(_sum0, activation_type, activation_params); if (out_elempack == 4) { vst1_f16(outptr, _sum0); outptr += 4; } else // if (out_elempack == 1) { outptr[0] = vget_lane_f16(_sum0, 0); outptr[M] = vget_lane_f16(_sum0, 1); outptr[M * 2] = vget_lane_f16(_sum0, 2); outptr[M * 3] = vget_lane_f16(_sum0, 3); outptr += 1; } } } remain_outh_start += nn_outh * 4; nn_outh = (outh - remain_outh_start) / 2; for (int pp = 0; pp < nn_outh; pp++) { const int p = remain_outh_start + pp * 2; // shadowed variable for less openmp task args const int elempack = bottom_blob.elempack; const int inh = bottom_blob.h * elempack; const int outw = top_blob.w; __fp16* outptr0 = top_blob.row<__fp16>(p); __fp16* outptr1 = top_blob.row<__fp16>(p + 1); for (int j = 0; j < outw; j++) { __fp16 sum0 = 0.f; __fp16 sum1 = 0.f; if (bias_data_ptr) { sum0 = bias_data_ptr[p]; sum1 = bias_data_ptr[p + 1]; } const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2); int q = 0; float16x8_t _sum0 = vdupq_n_f16(0.f); float16x8_t _sum1 = vdupq_n_f16(0.f); for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x8_t _r0; if (elempack == 8) { _r0 = vld1q_f16(r0); r0 += dilation_w * 8; } else if (elempack == 4) { _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x8_t(); _r0 = vsetq_lane_f16(r0[0], _r0, 0); _r0 = vsetq_lane_f16(r0[N], _r0, 1); _r0 = vsetq_lane_f16(r0[N * 2], _r0, 2); _r0 = vsetq_lane_f16(r0[N * 3], _r0, 3); _r0 = vsetq_lane_f16(r0[N * 4], _r0, 4); _r0 = vsetq_lane_f16(r0[N * 5], _r0, 5); _r0 = vsetq_lane_f16(r0[N * 6], _r0, 6); _r0 = vsetq_lane_f16(r0[N * 7], _r0, 7); r0 += dilation_w; } float16x8_t _w0 = vld1q_f16(kptr); float16x8_t _w1 = vld1q_f16(kptr + 8); _sum0 = vfmaq_f16(_sum0, _r0, _w0); _sum1 = vfmaq_f16(_sum1, _r0, _w1); kptr += 16; } } for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x4_t _r0; if (elempack == 4) { _r0 = vld1_f16(r0); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x4_t(); _r0 = vset_lane_f16(r0[0], _r0, 0); _r0 = vset_lane_f16(r0[N], _r0, 1); _r0 = vset_lane_f16(r0[N * 2], _r0, 2); _r0 = vset_lane_f16(r0[N * 3], _r0, 3); r0 += dilation_w; } float16x4_t _w0 = vld1_f16(kptr); float16x4_t _w1 = vld1_f16(kptr + 4); _sum0 = vcombine_f16(vfma_f16(vget_low_f16(_sum0), _r0, _w0), vget_high_f16(_sum0)); _sum1 = vcombine_f16(vfma_f16(vget_low_f16(_sum1), _r0, _w1), vget_high_f16(_sum1)); kptr += 8; } } for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { __fp16 val0; __fp16 val1; // if (elempack == 1) { val0 = r0[0]; val1 = r0[N]; r0 += dilation_w; } sum0 += val0 * kptr[0]; sum1 += val0 * kptr[1]; sum0 += val1 * kptr[2]; sum1 += val1 * kptr[3]; kptr += 4; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { __fp16 val; // if (elempack == 1) { val = r0[0]; r0 += dilation_w; } sum0 += val * kptr[0]; sum1 += val * kptr[1]; kptr += 2; } } float16x4_t _ss0 = vadd_f16(vget_low_f16(_sum0), vget_high_f16(_sum0)); float16x4_t _ss1 = vadd_f16(vget_low_f16(_sum1), vget_high_f16(_sum1)); float16x4_t _ss = vpadd_f16(_ss0, _ss1); _ss = vpadd_f16(_ss, _ss); sum0 += vget_lane_f16(_ss, 0); sum1 += vget_lane_f16(_ss, 1); sum0 = activation_ss(sum0, activation_type, activation_params); sum1 = activation_ss(sum1, activation_type, activation_params); outptr0[0] = sum0; outptr1[0] = sum1; outptr0 += 1; outptr1 += 1; } } remain_outh_start += nn_outh * 2; for (int p = remain_outh_start; p < outh; p++) { __fp16* outptr = top_blob.row<__fp16>(p); for (int j = 0; j < outw; j++) { __fp16 sum = 0.f; if (bias_data_ptr) { sum = bias_data_ptr[p]; } const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2); int q = 0; float16x8_t _sum = vdupq_n_f16(0.f); for (; q + 7 < inh; q += 8) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x8_t _r0; if (elempack == 8) { _r0 = vld1q_f16(r0); r0 += dilation_w * 8; } else if (elempack == 4) { _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r0 + N)); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x8_t(); _r0 = vsetq_lane_f16(r0[0], _r0, 0); _r0 = vsetq_lane_f16(r0[N], _r0, 1); _r0 = vsetq_lane_f16(r0[N * 2], _r0, 2); _r0 = vsetq_lane_f16(r0[N * 3], _r0, 3); _r0 = vsetq_lane_f16(r0[N * 4], _r0, 4); _r0 = vsetq_lane_f16(r0[N * 5], _r0, 5); _r0 = vsetq_lane_f16(r0[N * 6], _r0, 6); _r0 = vsetq_lane_f16(r0[N * 7], _r0, 7); r0 += dilation_w; } float16x8_t _w0 = vld1q_f16(kptr); _sum = vfmaq_f16(_sum, _r0, _w0); kptr += 8; } } for (; q + 3 < inh; q += 4) { const __fp16* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack; for (int k = 0; k < kernel_w; k++) { float16x4_t _r0; if (elempack == 4) { _r0 = vld1_f16(r0); r0 += dilation_w * 4; } else // if (elempack == 1) { _r0 = float16x4_t(); _r0 = vset_lane_f16(r0[0], _r0, 0); _r0 = vset_lane_f16(r0[N], _r0, 1); _r0 = vset_lane_f16(r0[N * 2], _r0, 2); _r0 = vset_lane_f16(r0[N * 3], _r0, 3); r0 += dilation_w; } float16x4_t _w = vld1_f16(kptr); _sum = vcombine_f16(vfma_f16(vget_low_f16(_sum), _r0, _w), vget_high_f16(_sum)); kptr += 4; } } for (; q + 1 < inh; q += 2) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { __fp16 val0; __fp16 val1; // if (elempack == 1) { val0 = r0[0]; val1 = r0[N]; r0 += dilation_w; } sum += val0 * kptr[0]; sum += val1 * kptr[1]; kptr += 2; } } for (; q < inh; q++) { const __fp16* r0 = bottom_blob.row(q) + j * stride_w; for (int k = 0; k < kernel_w; k++) { __fp16 val; // if (elempack == 1) { val = r0[0]; r0 += dilation_w; } sum += val * kptr[0]; kptr += 1; } } float16x4_t _ss = vadd_f16(vget_low_f16(_sum), vget_high_f16(_sum)); _ss = vpadd_f16(_ss, _ss); _ss = vpadd_f16(_ss, _ss); sum += vget_lane_f16(_ss, 0); sum = activation_ss(sum, activation_type, activation_params); outptr[0] = sum; outptr += 1; } } }