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 <wasm_simd128.h> 11 12#include <xnnpack/dwconv.h> 13#include <xnnpack/math.h> 14 15 16$ARCH_SUFFIX = "_x86" if X86 else "_arm" 17 18void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd${ARCH_SUFFIX}_splat_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}( 19 size_t input_height, 20 size_t input_width, 21 const float* input, 22 const float* weights, 23 const float* zero, 24 float* output, 25 uint32_t padding_top, 26 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 27{ 28 assert(input_height != 0); 29 assert(input_width != 0); 30 assert(input_width % sizeof(float) == 0); 31 assert(padding_top >= 0); 32 assert(padding_top <= 1); 33 34 const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even); 35 const v128_t vmask_odd = wasm_v128_load(params->scalar.mask_odd); 36 const v128_t vmax = wasm_v128_load32_splat(¶ms->scalar.max); 37 const v128_t vmin = wasm_v128_load32_splat(¶ms->scalar.min); 38 39 const v128_t vw0123 = wasm_v128_load(weights); 40 const v128_t vw4567 = wasm_v128_load(weights + 4); 41 const v128_t vw89 = wasm_v128_load64_splat(weights + 8); 42 43 const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float)); 44 $if ROW_TILE > 1: 45 const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float)); 46 47 const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width)); 48 const float* i1 = (const float*) ((uintptr_t) i0 + input_width); 49 if XNN_UNPREDICTABLE(padding_top != 0) { 50 i0 = zero; 51 } 52 $for M in range(2, 1 + 2 * ROW_TILE): 53 const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width); 54 55 float* o0 = output; 56 $for M in range(1, ROW_TILE): 57 float* o${M} = (float*) ((uintptr_t) o${M-1} + output_width); 58 59 size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */; 60 size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2; 61 do { 62 $for M in range(2, 1 + 2 * ROW_TILE): 63 if XNN_UNPREDICTABLE(padded_input_height < ${2 + M}) { 64 i${M} = zero; 65 $if M % 2 == 1: 66 o${(M - 1) // 2} = o${(M - 1) // 2 - 1}; 67 } 68 69 $for M in range(1 + 2 * ROW_TILE): 70 v128_t vi${M}x1357 = wasm_f32x4_const_splat(0.0f); 71 72 size_t w = input_width; 73 for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) { 74 $for M in range(ROW_TILE): 75 v128_t vo${M}p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); 76 77 $for M in range(1 + 2 * ROW_TILE): 78 const v128_t vi${M}x89AB = wasm_v128_load(i${M}); 79 const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4); 80 i${M} += 8; 81 82 $for M in range(1 + 2 * ROW_TILE): 83 const v128_t vi${M}x8ACE = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6); 84 const v128_t vi${M}x9BDF = wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7); 85 86 $for M in range(ROW_TILE): 87 $if ACCUMULATORS > 1: 88 v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)); 89 $else: 90 vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2))); 91 92 $for M in range(ROW_TILE): 93 $if ACCUMULATORS > 2: 94 v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); 95 $else: 96 vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1))); 97 98 $for M in range(ROW_TILE): 99 $if ACCUMULATORS > 3: 100 v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); 101 $else: 102 vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))); 103 104 $for M in range(1 + 2 * ROW_TILE): 105 const v128_t vi${M}x7BDF = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6); 106 vi${M}x1357 = vi${M}x9BDF; 107 108 $for M in range(ROW_TILE): 109 $if ACCUMULATORS > 4: 110 v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M}x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)); 111 $else: 112 vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1))); 113 114 $for M in range(ROW_TILE): 115 $if ACCUMULATORS > 5: 116 v128_t vo${M}p5 = wasm_f32x4_mul(vi${2*M+1}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)); 117 $else: 118 vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0))); 119 120 $for M in range(ROW_TILE): 121 $if ACCUMULATORS > 6: 122 v128_t vo${M}p6 = wasm_f32x4_mul(vi${2*M+2}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); 123 $else: 124 vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3))); 125 126 $for M in range(ROW_TILE): 127 vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3))); 128 129 $for M in range(ROW_TILE): 130 vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2))); 131 132 $for M in range(ROW_TILE): 133 vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))); 134 135 $if ACCUMULATORS > 1: 136 $ACC_SLICE = 1 137 $while ACC_SLICE < ACCUMULATORS: 138 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 139 $if A + ACC_SLICE < ACCUMULATORS: 140 $for M in range(ROW_TILE): 141 vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE}); 142 $ACC_SLICE *= 2 143 144 $if X86: 145 $for M in range(ROW_TILE): 146 v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0); 147 $for M in range(ROW_TILE): 148 vo${M} = wasm_f32x4_pmin(vmax, vo${M}); 149 $else: 150 $for M in range(ROW_TILE): 151 v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin); 152 $for M in range(ROW_TILE): 153 vo${M} = wasm_f32x4_min(vo${M}, vmax); 154 155 $for M in reversed(range(ROW_TILE)): 156 wasm_v128_store(o${M}, vo${M}); o${M} += 4; 157 } 158 // Last block has 0-7 pixels to process. 159 assert(w < 8 * sizeof(float)); 160 if XNN_LIKELY(w != 0) { 161 $for M in range(ROW_TILE): 162 v128_t vo${M}p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); 163 164 $for M in range(1 + 2 * ROW_TILE): 165 const v128_t vi${M}x89AB = wasm_v128_load(i${M}); 166 const v128_t vi${M}xCDEF = wasm_v128_load(i${M} + 4); 167 168 $for M in range(1 + 2 * ROW_TILE): 169 const v128_t vi${M}x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 0, 2, 4, 6)); 170 const v128_t vi${M}x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi${M}x89AB, vi${M}xCDEF, 1, 3, 5, 7)); 171 172 $for M in range(ROW_TILE): 173 $if ACCUMULATORS > 1: 174 v128_t vo${M}p1 = wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2)); 175 $else: 176 vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M}x8ACE, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2))); 177 178 $for M in range(ROW_TILE): 179 $if ACCUMULATORS > 2: 180 v128_t vo${M}p2 = wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); 181 $else: 182 vo${M}p0 = wasm_f32x4_add(vo${M}p0, wasm_f32x4_mul(vi${2*M+1}x8ACE, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1))); 183 184 $for M in range(ROW_TILE): 185 $if ACCUMULATORS > 3: 186 v128_t vo${M}p3 = wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0)); 187 $else: 188 vo${M}p${4 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${4 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x8ACE, wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0))); 189 190 $for M in range(1 + 2 * ROW_TILE): 191 const v128_t vi${M}x7BDF = wasm_v32x4_shuffle(vi${M}x1357, vi${M}x9BDF, 3, 4, 5, 6); 192 193 $for M in range(ROW_TILE): 194 $if ACCUMULATORS > 4: 195 v128_t vo${M}p4 = wasm_f32x4_mul(vi${2*M}x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1)); 196 $else: 197 vo${M}p${5 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${5 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1))); 198 199 $for M in range(ROW_TILE): 200 $if ACCUMULATORS > 5: 201 v128_t vo${M}p5 = wasm_f32x4_mul(vi${2*M+1}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0)); 202 $else: 203 vo${M}p${6 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${6 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0))); 204 205 $for M in range(ROW_TILE): 206 $if ACCUMULATORS > 6: 207 v128_t vo${M}p6 = wasm_f32x4_mul(vi${2*M+2}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1)); 208 $else: 209 vo${M}p${7 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${7 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3))); 210 211 $for M in range(ROW_TILE): 212 vo${M}p${8 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${8 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M}x9BDF, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3))); 213 214 $for M in range(ROW_TILE): 215 vo${M}p${9 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${9 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+1}x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2))); 216 217 $for M in range(ROW_TILE): 218 vo${M}p${10 % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${10 % ACCUMULATORS}, wasm_f32x4_mul(vi${2*M+2}x9BDF, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))); 219 220 $if ACCUMULATORS > 1: 221 $ACC_SLICE = 1 222 $while ACC_SLICE < ACCUMULATORS: 223 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 224 $if A + ACC_SLICE < ACCUMULATORS: 225 $for M in range(ROW_TILE): 226 vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE}); 227 $ACC_SLICE *= 2 228 229 $if X86: 230 $for M in range(ROW_TILE): 231 v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0); 232 $for M in range(ROW_TILE): 233 vo${M} = wasm_f32x4_pmin(vmax, vo${M}); 234 $else: 235 $for M in range(ROW_TILE): 236 v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin); 237 $for M in range(ROW_TILE): 238 vo${M} = wasm_f32x4_min(vo${M}, vmax); 239 240 w += 1 * sizeof(float); 241 if (w & (8 * sizeof(float))) { 242 $for M in reversed(range(ROW_TILE)): 243 wasm_v128_store(o${M}, vo${M}); o${M} += 4; 244 } else { 245 if (w & (4 * sizeof(float))) { 246 $for M in reversed(range(ROW_TILE)): 247 *((double*) o${M}) = wasm_f64x2_extract_lane(vo${M}, 0); o${M} += 2; 248 249 $for M in range(ROW_TILE): 250 vo${M} = wasm_v32x4_shuffle(vo${M}, vo${M}, 2, 3, 0, 1); 251 } 252 if (w & (2 * sizeof(float))) { 253 $for M in reversed(range(ROW_TILE)): 254 *o${M} = wasm_f32x4_extract_lane(vo${M}, 0); o${M} += 1; 255 } 256 } 257 } 258 259 i0 = (const float*) ((uintptr_t) i${2 * ROW_TILE} - input_decrement); 260 $for M in range(1, 1 + 2 * ROW_TILE): 261 i${M} = (const float*) ((uintptr_t) i${M-1} + input_width); 262 263 $if ROW_TILE > 1: 264 o0 = o${ROW_TILE - 1}; 265 $for M in range(1, ROW_TILE): 266 o${M} = (float*) ((uintptr_t) o${M-1} + output_width); 267 268 $if ROW_TILE > 1: 269 output_height = doz(output_height, ${ROW_TILE}); 270 padded_input_height = doz(padded_input_height, ${ROW_TILE * 2}); 271 $else: 272 output_height -= 1; 273 padded_input_height -= 2; 274 } while (output_height != 0); 275} 276