1// Copyright 2019 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$assert CHANNEL_TILE % 8 == 0 7$assert KERNEL_TILE >= 2 8$assert ACCUMULATORS >= 1 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/dwconv.h> 15 16 17static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0}; 18 19$ISA = {0: "avx", 3: "fma3"}[FMA] 20void xnn_f32_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__${ISA}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}( 21 size_t channels, 22 size_t output_width, 23 const float** input, 24 const float* weights, 25 float* output, 26 size_t input_stride, 27 size_t output_increment, 28 size_t input_offset, 29 const float* zero, 30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 31{ 32 assert(channels != 0); 33 assert(output_width != 0); 34 35 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max); 36 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min); 37 do { 38 $for K in range(KERNEL_TILE): 39 const float* i${K} = input[${K}]; 40 assert(i${K} != NULL); 41 if XNN_UNPREDICTABLE(i${K} != zero) { 42 i${K} = (const float*) ((uintptr_t) i${K} + input_offset); 43 } 44 input = (const float**) ((uintptr_t) input + input_stride); 45 46 size_t c = channels; 47 const float* w = weights; 48 for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) { 49 __m256 vacc${ABC[0:8]}p0 = _mm256_load_ps(w); 50 $for C in range(8, CHANNEL_TILE, 8): 51 __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(w + ${C}); 52 53 $for K in range(KERNEL_TILE): 54 55 const __m256 vi${K}x${ABC[0:8]} = _mm256_loadu_ps(i${K}); 56 $for C in range(8, CHANNEL_TILE, 8): 57 const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K} + ${C}); 58 i${K} += ${CHANNEL_TILE}; 59 60 $for C in range(0, CHANNEL_TILE, 8): 61 const __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(w + ${(K + 1) * CHANNEL_TILE + C}); 62 $for C in range(0, CHANNEL_TILE, 8): 63 $if 1 <= K < ACCUMULATORS: 64 __m256 vacc${ABC[C:C+8]}p${K} = _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}); 65 $elif FMA == 3: 66 vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}, vacc${ABC[C:C+8]}p${K % ACCUMULATORS}); 67 $else: 68 vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_add_ps(vacc${ABC[C:C+8]}p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]})); 69 70 w += ${(KERNEL_TILE + 1) * CHANNEL_TILE}; 71 72 $if ACCUMULATORS > 1: 73 // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0 74 $ACC_SLICE = 1 75 $while ACC_SLICE < ACCUMULATORS: 76 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 77 $if A + ACC_SLICE < ACCUMULATORS: 78 $for C in range(0, CHANNEL_TILE, 8): 79 vacc${ABC[C:C+8]}p${A} = _mm256_add_ps(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_SLICE}); 80 $ACC_SLICE *= 2 81 82 $for C in range(0, CHANNEL_TILE, 8): 83 __m256 vacc${ABC[C:C+8]} = _mm256_max_ps(vacc${ABC[C:C+8]}p0, vmin); 84 $for C in range(0, CHANNEL_TILE, 8): 85 vacc${ABC[C:C+8]} = _mm256_min_ps(vacc${ABC[C:C+8]}, vmax); 86 87 _mm256_storeu_ps(output, vacc${ABC[0:8]}); 88 $for C in range(8, CHANNEL_TILE, 8): 89 _mm256_storeu_ps(output + ${C}, vacc${ABC[C:C+8]}); 90 output += ${CHANNEL_TILE}; 91 } 92 $if CHANNEL_TILE > 8: 93 for (; c >= 8; c -= 8) { 94 __m256 vacc01234567p0 = _mm256_load_ps(w); 95 $for K in range(KERNEL_TILE): 96 97 const __m256 vi${K}x01234567 = _mm256_loadu_ps(i${K}); 98 i${K} += 8; 99 100 const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K + 1) * CHANNEL_TILE}); 101 $if 1 <= K < ACCUMULATORS: 102 __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567); 103 $elif FMA == 3: 104 vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS}); 105 $else: 106 vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567)); 107 108 w += 8; 109 110 $if ACCUMULATORS > 1: 111 // Add up all accumulators to vacc${ABC[0:8]}p0 112 $ACC_SLICE = 1 113 $while ACC_SLICE < ACCUMULATORS: 114 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 115 $if A + ACC_SLICE < ACCUMULATORS: 116 vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE}); 117 $ACC_SLICE *= 2 118 119 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin); 120 vacc01234567 = _mm256_min_ps(vacc01234567, vmax); 121 122 _mm256_storeu_ps(output, vacc01234567); 123 output += 8; 124 } 125 if XNN_UNLIKELY(c != 0) { 126 assert(c >= 1); 127 assert(c <= 7); 128 __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]); 129 130 __m256 vacc01234567p0 = _mm256_load_ps(w); 131 $for K in range(KERNEL_TILE): 132 133 const __m256 vi${K}x01234567 = _mm256_maskload_ps(i${K}, vmask); 134 const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K + 1) * CHANNEL_TILE}); 135 $if 1 <= K < ACCUMULATORS: 136 __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567); 137 $elif FMA == 3: 138 vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS}); 139 $else: 140 vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567)); 141 142 $if ACCUMULATORS > 1: 143 // Add up all accumulators to vacc${ABC[0:8]}p0 144 $ACC_SLICE = 1 145 $while ACC_SLICE < ACCUMULATORS: 146 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 147 $if A + ACC_SLICE < ACCUMULATORS: 148 vacc01234567p${A} = _mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE}); 149 $ACC_SLICE *= 2 150 151 __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin); 152 vacc01234567 = _mm256_min_ps(vacc01234567, vmax); 153 154 // _mm256_maskstore_ps(output, vmask, vacc01234567); output += c; could be used here, but triggers msan failures (probably an msan bug). 155 __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567); 156 if (c & 4) { 157 _mm_storeu_ps(output, vacc0123); 158 vacc0123 = _mm256_extractf128_ps(vacc01234567, 1); 159 output += 4; 160 } 161 if (c & 2) { 162 _mm_storel_pi((__m64*) output, vacc0123); 163 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123); 164 output += 2; 165 } 166 if (c & 1) { 167 _mm_store_ss(output, vacc0123); 168 output += 1; 169 } 170 } 171 172 output = (float*) ((uintptr_t) output + output_increment); 173 } while (--output_width != 0); 174} 175