// Copyright 2019 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include #include void xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1( size_t input_height, size_t input_width, size_t output_y_start, size_t output_y_end, const float* input, const float* zero, const float* weights, float* output, size_t input_padding_top, size_t output_channels, size_t output_height_stride, size_t output_channel_stride, const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(input_width != 0); assert(output_y_end > output_y_start); assert(input_padding_top <= 1); assert(output_channels != 0); const size_t input_height_stride = input_width * 3 /* channels */ * sizeof(float); const size_t input_width_decrement = round_down_po2(input_width, 2) * 3 /* channels */ * sizeof(float); const size_t output_width = (input_width + 1) / 2; const size_t output_channel_increment = output_channel_stride * 4 - output_width * sizeof(float); // Adjustment for padding processed below const float* i0 = (const float*) ((uintptr_t) input + input_height_stride * (output_y_start * 2 - input_padding_top)); const float* i1 = (const float*) ((uintptr_t) i0 + input_height_stride); const float* i2 = (const float*) ((uintptr_t) i1 + input_height_stride); float* output0 = (float*) ((uintptr_t) output + output_height_stride * output_y_start); if XNN_UNPREDICTABLE(output_y_start < input_padding_top) { i0 = zero; } const __m128 vmin = _mm_load_ps(params->sse.min); const __m128 vmax = _mm_load_ps(params->sse.max); for (size_t output_y = output_y_start; output_y < output_y_end; output_y += 1) { const size_t input_y2 = output_y * 2 + 2 - input_padding_top; if XNN_UNPREDICTABLE(input_y2 >= input_height) { i2 = zero; } const float* w = weights; size_t c = output_channels; float* o0c0 = output0; float* o0c1 = (float*) ((uintptr_t) o0c0 + output_channel_stride); float* o0c2 = (float*) ((uintptr_t) o0c1 + output_channel_stride); float* o0c3 = (float*) ((uintptr_t) o0c2 + output_channel_stride); do { if XNN_UNPREDICTABLE(c < 2) { o0c1 = o0c0; } if XNN_UNPREDICTABLE(c <= 2) { o0c2 = o0c1; } if XNN_UNPREDICTABLE(c < 4) { o0c3 = o0c2; } // Left edge padding __m128 vi00c0 = _mm_setzero_ps(); __m128 vi00c1 = _mm_setzero_ps(); __m128 vi00c2 = _mm_setzero_ps(); __m128 vi10c0 = _mm_setzero_ps(); __m128 vi10c1 = _mm_setzero_ps(); __m128 vi10c2 = _mm_setzero_ps(); __m128 vi20c0 = _mm_setzero_ps(); __m128 vi20c1 = _mm_setzero_ps(); __m128 vi20c2 = _mm_setzero_ps(); size_t iw = input_width; for (; iw >= 2; iw -= 2) { __m128 voc0123 = _mm_loadu_ps(w); const __m128 vk00c0x0123 = _mm_load_ps(w + 4); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk00c0x0123, vi00c0)); const __m128 vk10c0x0123 = _mm_load_ps(w + 8); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c0x0123, vi10c0)); const __m128 vk20c0x0123 = _mm_load_ps(w + 12); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c0x0123, vi20c0)); const __m128 vk00c1x0123 = _mm_load_ps(w + 16); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk00c1x0123, vi00c1)); const __m128 vk10c1x0123 = _mm_load_ps(w + 20); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c1x0123, vi10c1)); const __m128 vk20c1x0123 = _mm_load_ps(w + 24); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c1x0123, vi20c1)); const __m128 vk00c2x0123 = _mm_load_ps(w + 28); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk00c2x0123, vi00c2)); const __m128 vk10c2x0123 = _mm_load_ps(w + 32); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c2x0123, vi10c2)); const __m128 vk20c2x0123 = _mm_load_ps(w + 36); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c2x0123, vi20c2)); const __m128 vk01c0x0123 = _mm_load_ps(w + 40); const __m128 vi01c0 = _mm_load1_ps(i0); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c0x0123, vi01c0)); const __m128 vk11c0x0123 = _mm_load_ps(w + 44); const __m128 vi11c0 = _mm_load1_ps(i1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c0x0123, vi11c0)); const __m128 vk21c0x0123 = _mm_load_ps(w + 48); const __m128 vi21c0 = _mm_load1_ps(i2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c0x0123, vi21c0)); const __m128 vk01c1x0123 = _mm_load_ps(w + 52); const __m128 vi01c1 = _mm_load1_ps(i0 + 1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c1x0123, vi01c1)); const __m128 vk11c1x0123 = _mm_load_ps(w + 56); const __m128 vi11c1 = _mm_load1_ps(i1 + 1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c1x0123, vi11c1)); const __m128 vk21c1x0123 = _mm_load_ps(w + 60); const __m128 vi21c1 = _mm_load1_ps(i2 + 1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c1x0123, vi21c1)); const __m128 vk01c2x0123 = _mm_load_ps(w + 64); const __m128 vi01c2 = _mm_load1_ps(i0 + 2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c2x0123, vi01c2)); const __m128 vk11c2x0123 = _mm_load_ps(w + 68); const __m128 vi11c2 = _mm_load1_ps(i1 + 2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c2x0123, vi11c2)); const __m128 vk21c2x0123 = _mm_load_ps(w + 72); const __m128 vi21c2 = _mm_load1_ps(i2 + 2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c2x0123, vi21c2)); const __m128 vk02c0x0123 = _mm_load_ps(w + 76); const __m128 vi02c0 = _mm_load1_ps(i0 + 3); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk02c0x0123, vi02c0)); const __m128 vk12c0x0123 = _mm_load_ps(w + 80); const __m128 vi12c0 = _mm_load1_ps(i1 + 3); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk12c0x0123, vi12c0)); const __m128 vk22c0x0123 = _mm_load_ps(w + 84); const __m128 vi22c0 = _mm_load1_ps(i2 + 3); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk22c0x0123, vi22c0)); vi00c0 = vi02c0; vi10c0 = vi12c0; vi20c0 = vi22c0; const __m128 vk02c1x0123 = _mm_load_ps(w + 88); const __m128 vi02c1 = _mm_load1_ps(i0 + 4); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk02c1x0123, vi02c1)); const __m128 vk12c1x0123 = _mm_load_ps(w + 92); const __m128 vi12c1 = _mm_load1_ps(i1 + 4); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk12c1x0123, vi12c1)); const __m128 vk22c1x0123 = _mm_load_ps(w + 96); const __m128 vi22c1 = _mm_load1_ps(i2 + 4); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk22c1x0123, vi22c1)); vi00c1 = vi02c1; vi10c1 = vi12c1; vi20c1 = vi22c1; const __m128 vk02c2x0123 = _mm_load_ps(w + 100); const __m128 vi02c2 = _mm_load1_ps(i0 + 5); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk02c2x0123, vi02c2)); const __m128 vk12c2x0123 = _mm_load_ps(w + 104); const __m128 vi12c2 = _mm_load1_ps(i1 + 5); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk12c2x0123, vi12c2)); const __m128 vk22c2x0123 = _mm_load_ps(w + 108); const __m128 vi22c2 = _mm_load1_ps(i2 + 5); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk22c2x0123, vi22c2)); vi00c2 = vi02c2; vi10c2 = vi12c2; vi20c2 = vi22c2; voc0123 = _mm_min_ps(voc0123, vmax); voc0123 = _mm_max_ps(voc0123, vmin); _mm_store_ss(o0c0, voc0123); o0c0++; _mm_store_ss(o0c1, _mm_shuffle_ps(voc0123, voc0123, 1)); o0c1++; _mm_store_ss(o0c2, _mm_shuffle_ps(voc0123, voc0123, 2)); o0c2++; _mm_store_ss(o0c3, _mm_shuffle_ps(voc0123, voc0123, 3)); o0c3++; i0 += 6; i1 += 6; i2 += 6; } assert(iw < 2); if XNN_UNLIKELY(iw != 0) { __m128 voc0123 = _mm_load_ps(w); const __m128 vk00c0x0123 = _mm_load_ps(w + 4); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk00c0x0123, vi00c0)); const __m128 vk10c0x0123 = _mm_load_ps(w + 8); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c0x0123, vi10c0)); const __m128 vk20c0x0123 = _mm_load_ps(w + 12); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c0x0123, vi20c0)); const __m128 vk00c1x0123 = _mm_load_ps(w + 16); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk00c1x0123, vi00c1)); const __m128 vk10c1x0123 = _mm_load_ps(w + 20); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c1x0123, vi10c1)); const __m128 vk20c1x0123 = _mm_load_ps(w + 24); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c1x0123, vi20c1)); const __m128 vk00c2x0123 = _mm_load_ps(w + 28); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk00c2x0123, vi00c2)); const __m128 vk10c2x0123 = _mm_load_ps(w + 32); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c2x0123, vi10c2)); const __m128 vk20c2x0123 = _mm_load_ps(w + 36); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c2x0123, vi20c2)); const __m128 vk01c0x0123 = _mm_load_ps(w + 40); const __m128 vi01c0 = _mm_load1_ps(i0); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c0x0123, vi01c0)); const __m128 vk11c0x0123 = _mm_load_ps(w + 44); const __m128 vi11c0 = _mm_load1_ps(i1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c0x0123, vi11c0)); const __m128 vk21c0x0123 = _mm_load_ps(w + 48); const __m128 vi21c0 = _mm_load1_ps(i2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c0x0123, vi21c0)); const __m128 vk01c1x0123 = _mm_load_ps(w + 52); const __m128 vi01c1 = _mm_load1_ps(i0 + 1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c1x0123, vi01c1)); const __m128 vk11c1x0123 = _mm_load_ps(w + 56); const __m128 vi11c1 = _mm_load1_ps(i1 + 1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c1x0123, vi11c1)); const __m128 vk21c1x0123 = _mm_load_ps(w + 60); const __m128 vi21c1 = _mm_load1_ps(i2 + 1); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c1x0123, vi21c1)); const __m128 vk01c2x0123 = _mm_load_ps(w + 64); const __m128 vi01c2 = _mm_load1_ps(i0 + 2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk01c2x0123, vi01c2)); const __m128 vk11c2x0123 = _mm_load_ps(w + 68); const __m128 vi11c2 = _mm_load1_ps(i1 + 2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk11c2x0123, vi11c2)); const __m128 vk21c2x0123 = _mm_load_ps(w + 72); const __m128 vi21c2 = _mm_load1_ps(i2 + 2); voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk21c2x0123, vi21c2)); voc0123 = _mm_min_ps(voc0123, vmax); voc0123 = _mm_max_ps(voc0123, vmin); _mm_store_ss(o0c0, voc0123); o0c0++; _mm_store_ss(o0c1, _mm_shuffle_ps(voc0123, voc0123, 1)); o0c1++; _mm_store_ss(o0c2, _mm_shuffle_ps(voc0123, voc0123, 2)); o0c2++; _mm_store_ss(o0c3, _mm_shuffle_ps(voc0123, voc0123, 3)); o0c3++; } // Move output pointers back to the position of the first pixel in a row, // and forward to the next block of output channels. o0c0 = (float*) ((uintptr_t) o0c0 + output_channel_increment); o0c1 = (float*) ((uintptr_t) o0c1 + output_channel_increment); o0c2 = (float*) ((uintptr_t) o0c2 + output_channel_increment); o0c3 = (float*) ((uintptr_t) o0c3 + output_channel_increment); // Revert input pointers to the position of the first pixel in a row i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); // Move to the block of weights for the next 4 output channels w += 112; c = doz(c, 4); } while (c != 0); // Move output pointers forward to the next row output0 = (float*) ((uintptr_t) output0 + output_height_stride); // Move input pointers forward to the next row i0 = i2; i1 = (const float*) ((uintptr_t) i0 + input_height_stride); i2 = (const float*) ((uintptr_t) i1 + input_height_stride); } }