1// Copyright 2020 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$assert ROW_TILE >= 1 7$assert ACCUMULATORS >= 1 8#include <assert.h> 9 10#include <xnnpack/dwconv.h> 11#include <xnnpack/math.h> 12 13 14void xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_${ROW_TILE}x1${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}( 15 size_t input_height, 16 size_t input_width, 17 const float* input, 18 const float* weights, 19 const float* zero, 20 float* output, 21 uint32_t padding_top, 22 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) 23{ 24 assert(input_height != 0); 25 assert(input_width != 0); 26 assert(input_width % sizeof(float) == 0); 27 assert(padding_top == 2); 28 29 const float vmin = params->scalar.min; 30 const float vmax = params->scalar.max; 31 32 const float vbias = weights[0]; 33 $for R in range(5): 34 $for S in range(5): 35 const float vk${R}${S} = weights[${R*5+S+1}]; 36 37 const float* i0 = zero; 38 const float* i1 = zero; 39 const float* i2 = input; 40 $for M in range(3, 4 + ROW_TILE): 41 const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width); 42 43 float* o0 = output; 44 $for M in range(1, ROW_TILE): 45 float* o${M} = (float*) ((uintptr_t) o${M-1} + input_width); 46 47 size_t output_height = input_height; 48 do { 49 $for M in range(2, 3 + ROW_TILE): 50 if XNN_UNPREDICTABLE(output_height < ${M}) { 51 i${M+1} = zero; 52 $if M <= ROW_TILE: 53 o${M-1} = o${M-2}; 54 } 55 56 $for M in range(4 + ROW_TILE): 57 float vi${M}x0 = 0.0f; 58 59 $for M in range(4 + ROW_TILE): 60 float vi${M}x1 = 0.0f; 61 62 $for M in range(4 + ROW_TILE): 63 float vi${M}x2 = *i${M}++; 64 65 size_t w = input_width; 66 if (w > 1 * sizeof(float)) { 67 $for M in range(4 + ROW_TILE): 68 float vi${M}x3 = *i${M}++; 69 70 for (; w > 2 * sizeof(float); w -= 1 * sizeof(float)) { 71 $for M in range(4 + ROW_TILE): 72 const float vi${M}x4 = *i${M}++; 73 74 $for K in range(5): 75 $for M in range(ROW_TILE): 76 $if K == 0: 77 float vo${M}p0 = vbias + vi${M+K}x0 * vk${K}0; 78 $elif K < ACCUMULATORS: 79 float vo${M}p${K} = vi${M+K}x0 * vk${K}0; 80 $else: 81 vo${M}p${K % ACCUMULATORS} += vi${M+K}x0 * vk${K}0; 82 83 $for M in range(4 + ROW_TILE): 84 vi${M}x0 = vi${M}x1; 85 86 $for K in range(5): 87 $for M in range(ROW_TILE): 88 $if K+5 < ACCUMULATORS: 89 float vo${M}p${K+5} = vi${M+K}x1 * vk${K}1; 90 $else: 91 vo${M}p${(K+5) % ACCUMULATORS} += vi${M+K}x1 * vk${K}1; 92 93 $for M in range(4 + ROW_TILE): 94 vi${M}x1 = vi${M}x2; 95 96 $for K in range(5): 97 $for M in range(ROW_TILE): 98 vo${M}p${(K+10) % ACCUMULATORS} += vi${M+K}x2 * vk${K}2; 99 100 $for M in range(4 + ROW_TILE): 101 vi${M}x2 = vi${M}x3; 102 103 $for K in range(5): 104 $for M in range(ROW_TILE): 105 vo${M}p${(K+15) % ACCUMULATORS} += vi${M+K}x3 * vk${K}3; 106 107 $for M in range(4 + ROW_TILE): 108 vi${M}x3 = vi${M}x4; 109 110 $for K in range(5): 111 $for M in range(ROW_TILE): 112 vo${M}p${(K+20) % ACCUMULATORS} += vi${M+K}x4 * vk${K}4; 113 114 $if ACCUMULATORS > 1: 115 $ACC_SLICE = 1 116 $while ACC_SLICE < ACCUMULATORS: 117 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 118 $if A + ACC_SLICE < ACCUMULATORS: 119 $for M in range(ROW_TILE): 120 vo${M}p${A} += vo${M}p${A + ACC_SLICE}; 121 $ACC_SLICE *= 2 122 123 $for M in range(ROW_TILE): 124 float vo${M} = math_max_f32(vo${M}p0, vmin); 125 126 $for M in range(ROW_TILE): 127 vo${M} = math_min_f32(vo${M}, vmax); 128 129 $for M in reversed(range(ROW_TILE)): 130 *o${M}++ = vo${M}; 131 } 132 assert(w == 2 * sizeof(float)); 133 { 134 $for K in range(5): 135 $for M in range(ROW_TILE): 136 $if K == 0: 137 float vo${M}p0 = vbias + vi${M+K}x0 * vk${K}0; 138 $elif K < ACCUMULATORS: 139 float vo${M}p${K} = vi${M+K}x0 * vk${K}0; 140 $else: 141 vo${M}p${K % ACCUMULATORS} += vi${M+K}x0 * vk${K}0; 142 143 $for M in range(4 + ROW_TILE): 144 vi${M}x0 = vi${M}x1; 145 146 $for K in range(5): 147 $for M in range(ROW_TILE): 148 $if K+5 < ACCUMULATORS: 149 float vo${M}p${K+5} = vi${M+K}x1 * vk${K}1; 150 $else: 151 vo${M}p${(K+5) % ACCUMULATORS} += vi${M+K}x1 * vk${K}1; 152 153 $for M in range(4 + ROW_TILE): 154 vi${M}x1 = vi${M}x2; 155 156 $for K in range(5): 157 $for M in range(ROW_TILE): 158 vo${M}p${(K+10) % ACCUMULATORS} += vi${M+K}x2 * vk${K}2; 159 160 $for M in range(4 + ROW_TILE): 161 vi${M}x2 = vi${M}x3; 162 163 $for K in range(5): 164 $for M in range(ROW_TILE): 165 vo${M}p${(K+15) % ACCUMULATORS} += vi${M+K}x3 * vk${K}3; 166 167 $if ACCUMULATORS > 1: 168 $ACC_SLICE = 1 169 $while ACC_SLICE < ACCUMULATORS: 170 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 171 $if A + ACC_SLICE < ACCUMULATORS: 172 $for M in range(ROW_TILE): 173 vo${M}p${A} += vo${M}p${A + ACC_SLICE}; 174 $ACC_SLICE *= 2 175 176 $for M in range(ROW_TILE): 177 float vo${M} = math_max_f32(vo${M}p0, vmin); 178 179 $for M in range(ROW_TILE): 180 vo${M} = math_min_f32(vo${M}, vmax); 181 182 $for M in reversed(range(ROW_TILE)): 183 *o${M}++ = vo${M}; 184 } 185 w -= 1 * sizeof(float); 186 } 187 assert(w == 1 * sizeof(float)); 188 { 189 $for K in range(5): 190 $for M in range(ROW_TILE): 191 $if K == 0: 192 float vo${M}p0 = vbias + vi${M+K}x0 * vk${K}0; 193 $elif K < ACCUMULATORS: 194 float vo${M}p${K} = vi${M+K}x0 * vk${K}0; 195 $else: 196 vo${M}p${K % ACCUMULATORS} += vi${M+K}x0 * vk${K}0; 197 198 $for K in range(5): 199 $for M in range(ROW_TILE): 200 $if K+5 < ACCUMULATORS: 201 float vo${M}p${K+5} = vi${M+K}x1 * vk${K}1; 202 $else: 203 vo${M}p${(K+5) % ACCUMULATORS} += vi${M+K}x1 * vk${K}1; 204 205 $for K in range(5): 206 $for M in range(ROW_TILE): 207 vo${M}p${(K+10) % ACCUMULATORS} += vi${M+K}x2 * vk${K}2; 208 209 $if ACCUMULATORS > 1: 210 $ACC_SLICE = 1 211 $while ACC_SLICE < ACCUMULATORS: 212 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 213 $if A + ACC_SLICE < ACCUMULATORS: 214 $for M in range(ROW_TILE): 215 vo${M}p${A} += vo${M}p${A + ACC_SLICE}; 216 $ACC_SLICE *= 2 217 218 $for M in range(ROW_TILE): 219 float vo${M} = math_max_f32(vo${M}p0, vmin); 220 221 $for M in range(ROW_TILE): 222 vo${M} = math_min_f32(vo${M}, vmax); 223 224 $for M in reversed(range(ROW_TILE)): 225 *o${M}++ = vo${M}; 226 } 227 228 i0 = (const float*) ((uintptr_t) i${ROW_TILE} - input_width); 229 i1 = (const float*) ((uintptr_t) i${ROW_TILE+1} - input_width); 230 $if ROW_TILE > 1: 231 i2 = i${ROW_TILE+1}; 232 i3 = i${ROW_TILE+2}; 233 i4 = i${ROW_TILE+3}; 234 $for M in range(5, 4 + ROW_TILE): 235 i${M} = (const float*) ((uintptr_t) i${M-1} + input_width); 236 237 $if ROW_TILE > 1: 238 o0 = o${ROW_TILE - 1}; 239 $for M in range(1, ROW_TILE): 240 o${M} = (float*) ((uintptr_t) o${M-1} + input_width); 241 242 $if ROW_TILE > 1: 243 output_height = doz(output_height, ${ROW_TILE}); 244 } while (${"--" if ROW_TILE == 1 else ""}output_height != 0); 245} 246