Frank Barchard | c5704bf | 2020-12-21 23:09:00 -0800 | [diff] [blame^] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-dwconv2d-chw/3x3s2p1-wasmsimd-loadsplat.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <wasm_simd128.h> |
| 13 | |
| 14 | #include <xnnpack/dwconv.h> |
| 15 | #include <xnnpack/math.h> |
| 16 | |
| 17 | |
| 18 | |
| 19 | void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2( |
| 20 | size_t input_height, |
| 21 | size_t input_width, |
| 22 | const float* input, |
| 23 | const float* weights, |
| 24 | const float* zero, |
| 25 | float* output, |
| 26 | uint32_t padding_top, |
| 27 | const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 28 | { |
| 29 | assert(input_height != 0); |
| 30 | assert(input_width != 0); |
| 31 | assert(input_width % sizeof(float) == 0); |
| 32 | assert(padding_top >= 0); |
| 33 | assert(padding_top <= 1); |
| 34 | |
| 35 | const v128_t vmask_even = wasm_v128_load(params->scalar.mask_even); |
| 36 | const v128_t vmask_odd = wasm_v128_load(params->scalar.mask_odd); |
| 37 | const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max); |
| 38 | const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); |
| 39 | |
| 40 | const v128_t vw0123 = wasm_v128_load(weights); |
| 41 | const v128_t vw4567 = wasm_v128_load(weights + 4); |
| 42 | const v128_t vw89 = wasm_v64x2_load_splat(weights + 8); |
| 43 | const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); |
| 44 | const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1); |
| 45 | const v128_t vk01 = wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2); |
| 46 | const v128_t vk02 = wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3); |
| 47 | const v128_t vk10 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0); |
| 48 | const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1); |
| 49 | const v128_t vk12 = wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2); |
| 50 | const v128_t vk20 = wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3); |
| 51 | const v128_t vk21 = wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0); |
| 52 | const v128_t vk22 = wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1); |
| 53 | |
| 54 | const v128_t vzero = wasm_f32x4_splat(0.0f); |
| 55 | |
| 56 | const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float)); |
| 57 | const size_t output_width = round_down_po2((input_width + (2 /* padding */ - 3 /* kernel size */ + 2 /* subsampling */) * sizeof(float)) / 2, sizeof(float)); |
| 58 | |
| 59 | const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width)); |
| 60 | const float* i1 = (const float*) ((uintptr_t) i0 + input_width); |
| 61 | if XNN_UNPREDICTABLE(padding_top != 0) { |
| 62 | i0 = zero; |
| 63 | } |
| 64 | const float* i2 = (const float*) ((uintptr_t) i1 + input_width); |
| 65 | const float* i3 = (const float*) ((uintptr_t) i2 + input_width); |
| 66 | const float* i4 = (const float*) ((uintptr_t) i3 + input_width); |
| 67 | |
| 68 | float* o0 = output; |
| 69 | float* o1 = (float*) ((uintptr_t) o0 + output_width); |
| 70 | |
| 71 | size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */; |
| 72 | size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2; |
| 73 | do { |
| 74 | if XNN_UNPREDICTABLE(padded_input_height < 4) { |
| 75 | i2 = zero; |
| 76 | } |
| 77 | if XNN_UNPREDICTABLE(padded_input_height < 5) { |
| 78 | i3 = zero; |
| 79 | o1 = o0; |
| 80 | } |
| 81 | if XNN_UNPREDICTABLE(padded_input_height < 6) { |
| 82 | i4 = zero; |
| 83 | } |
| 84 | |
| 85 | v128_t vi0x1357 = vzero; |
| 86 | v128_t vi1x1357 = vzero; |
| 87 | v128_t vi2x1357 = vzero; |
| 88 | v128_t vi3x1357 = vzero; |
| 89 | v128_t vi4x1357 = vzero; |
| 90 | |
| 91 | size_t w = input_width; |
| 92 | for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) { |
| 93 | v128_t vo0p0 = vbias; |
| 94 | v128_t vo1p0 = vbias; |
| 95 | |
| 96 | const v128_t vi0x89AB = wasm_v128_load(i0); |
| 97 | const v128_t vi0xCDEF = wasm_v128_load(i0 + 4); |
| 98 | i0 += 8; |
| 99 | const v128_t vi1x89AB = wasm_v128_load(i1); |
| 100 | const v128_t vi1xCDEF = wasm_v128_load(i1 + 4); |
| 101 | i1 += 8; |
| 102 | const v128_t vi2x89AB = wasm_v128_load(i2); |
| 103 | const v128_t vi2xCDEF = wasm_v128_load(i2 + 4); |
| 104 | i2 += 8; |
| 105 | const v128_t vi3x89AB = wasm_v128_load(i3); |
| 106 | const v128_t vi3xCDEF = wasm_v128_load(i3 + 4); |
| 107 | i3 += 8; |
| 108 | const v128_t vi4x89AB = wasm_v128_load(i4); |
| 109 | const v128_t vi4xCDEF = wasm_v128_load(i4 + 4); |
| 110 | i4 += 8; |
| 111 | |
| 112 | const v128_t vi0x8ACE = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6); |
| 113 | const v128_t vi0x9BDF = wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 1, 3, 5, 7); |
| 114 | const v128_t vi1x8ACE = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 0, 2, 4, 6); |
| 115 | const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); |
| 116 | const v128_t vi2x8ACE = wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 0, 2, 4, 6); |
| 117 | const v128_t vi2x9BDF = wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 1, 3, 5, 7); |
| 118 | const v128_t vi3x8ACE = wasm_v32x4_shuffle(vi3x89AB, vi3xCDEF, 0, 2, 4, 6); |
| 119 | const v128_t vi3x9BDF = wasm_v32x4_shuffle(vi3x89AB, vi3xCDEF, 1, 3, 5, 7); |
| 120 | const v128_t vi4x8ACE = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6); |
| 121 | const v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); |
| 122 | |
| 123 | v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk01); |
| 124 | v128_t vo1p1 = wasm_f32x4_mul(vi2x8ACE, vk01); |
| 125 | |
| 126 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, vk11)); |
| 127 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, vk11)); |
| 128 | |
| 129 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, vk21)); |
| 130 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x8ACE, vk21)); |
| 131 | |
| 132 | const v128_t vi0x7BDF = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); |
| 133 | vi0x1357 = vi0x9BDF; |
| 134 | const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); |
| 135 | vi1x1357 = vi1x9BDF; |
| 136 | const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); |
| 137 | vi2x1357 = vi2x9BDF; |
| 138 | const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); |
| 139 | vi3x1357 = vi3x9BDF; |
| 140 | const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); |
| 141 | vi4x1357 = vi4x9BDF; |
| 142 | |
| 143 | vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x7BDF, vk00)); |
| 144 | vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, vk00)); |
| 145 | |
| 146 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x7BDF, vk10)); |
| 147 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); |
| 148 | |
| 149 | vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); |
| 150 | vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, vk20)); |
| 151 | |
| 152 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, vk02)); |
| 153 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x9BDF, vk02)); |
| 154 | |
| 155 | vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); |
| 156 | vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x9BDF, vk12)); |
| 157 | |
| 158 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, vk22)); |
| 159 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x9BDF, vk22)); |
| 160 | |
| 161 | vo0p0 = wasm_f32x4_add(vo0p0, vo0p1); |
| 162 | vo1p0 = wasm_f32x4_add(vo1p0, vo1p1); |
| 163 | |
| 164 | v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin)); |
| 165 | v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin)); |
| 166 | vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax)); |
| 167 | vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax)); |
| 168 | |
| 169 | wasm_v128_store(o1, vo1); o1 += 4; |
| 170 | wasm_v128_store(o0, vo0); o0 += 4; |
| 171 | } |
| 172 | // Last block has 0-7 pixels to process. |
| 173 | assert(w < 8 * sizeof(float)); |
| 174 | if XNN_LIKELY(w != 0) { |
| 175 | v128_t vo0p0 = vbias; |
| 176 | v128_t vo1p0 = vbias; |
| 177 | |
| 178 | const v128_t vi0x89AB = wasm_v128_load(i0); |
| 179 | const v128_t vi0xCDEF = wasm_v128_load(i0 + 4); |
| 180 | const v128_t vi1x89AB = wasm_v128_load(i1); |
| 181 | const v128_t vi1xCDEF = wasm_v128_load(i1 + 4); |
| 182 | const v128_t vi2x89AB = wasm_v128_load(i2); |
| 183 | const v128_t vi2xCDEF = wasm_v128_load(i2 + 4); |
| 184 | const v128_t vi3x89AB = wasm_v128_load(i3); |
| 185 | const v128_t vi3xCDEF = wasm_v128_load(i3 + 4); |
| 186 | const v128_t vi4x89AB = wasm_v128_load(i4); |
| 187 | const v128_t vi4xCDEF = wasm_v128_load(i4 + 4); |
| 188 | |
| 189 | const v128_t vi0x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 0, 2, 4, 6)); |
| 190 | const v128_t vi0x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi0x89AB, vi0xCDEF, 1, 3, 5, 7)); |
| 191 | const v128_t vi1x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 0, 2, 4, 6)); |
| 192 | const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7)); |
| 193 | const v128_t vi2x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 0, 2, 4, 6)); |
| 194 | const v128_t vi2x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi2x89AB, vi2xCDEF, 1, 3, 5, 7)); |
| 195 | const v128_t vi3x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi3x89AB, vi3xCDEF, 0, 2, 4, 6)); |
| 196 | const v128_t vi3x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi3x89AB, vi3xCDEF, 1, 3, 5, 7)); |
| 197 | const v128_t vi4x8ACE = wasm_v128_and(vmask_even, wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 0, 2, 4, 6)); |
| 198 | const v128_t vi4x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7)); |
| 199 | |
| 200 | v128_t vo0p1 = wasm_f32x4_mul(vi0x8ACE, vk01); |
| 201 | v128_t vo1p1 = wasm_f32x4_mul(vi2x8ACE, vk01); |
| 202 | |
| 203 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, vk11)); |
| 204 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, vk11)); |
| 205 | |
| 206 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x8ACE, vk21)); |
| 207 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x8ACE, vk21)); |
| 208 | |
| 209 | const v128_t vi0x7BDF = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); |
| 210 | const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); |
| 211 | const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); |
| 212 | const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); |
| 213 | const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); |
| 214 | |
| 215 | vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x7BDF, vk00)); |
| 216 | vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, vk00)); |
| 217 | |
| 218 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x7BDF, vk10)); |
| 219 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); |
| 220 | |
| 221 | vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); |
| 222 | vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, vk20)); |
| 223 | |
| 224 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x9BDF, vk02)); |
| 225 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x9BDF, vk02)); |
| 226 | |
| 227 | vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); |
| 228 | vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x9BDF, vk12)); |
| 229 | |
| 230 | vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x9BDF, vk22)); |
| 231 | vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x9BDF, vk22)); |
| 232 | |
| 233 | vo0p0 = wasm_f32x4_add(vo0p0, vo0p1); |
| 234 | vo1p0 = wasm_f32x4_add(vo1p0, vo1p1); |
| 235 | |
| 236 | v128_t vo0 = wasm_v128_bitselect(vmin, vo0p0, wasm_f32x4_lt(vo0p0, vmin)); |
| 237 | v128_t vo1 = wasm_v128_bitselect(vmin, vo1p0, wasm_f32x4_lt(vo1p0, vmin)); |
| 238 | vo0 = wasm_v128_bitselect(vo0, vmax, wasm_f32x4_le(vo0, vmax)); |
| 239 | vo1 = wasm_v128_bitselect(vo1, vmax, wasm_f32x4_le(vo1, vmax)); |
| 240 | |
| 241 | w += 1 * sizeof(float); |
| 242 | if (w & (8 * sizeof(float))) { |
| 243 | wasm_v128_store(o1, vo1); o1 += 4; |
| 244 | wasm_v128_store(o0, vo0); o0 += 4; |
| 245 | } else { |
| 246 | if (w & (4 * sizeof(float))) { |
| 247 | *((double*) o1) = wasm_f64x2_extract_lane(vo1, 0); o1 += 2; |
| 248 | *((double*) o0) = wasm_f64x2_extract_lane(vo0, 0); o0 += 2; |
| 249 | |
| 250 | vo0 = wasm_v32x4_shuffle(vo0, vo0, 2, 3, 0, 1); |
| 251 | vo1 = wasm_v32x4_shuffle(vo1, vo1, 2, 3, 0, 1); |
| 252 | } |
| 253 | if (w & (2 * sizeof(float))) { |
| 254 | *o1 = wasm_f32x4_extract_lane(vo1, 0); o1 += 1; |
| 255 | *o0 = wasm_f32x4_extract_lane(vo0, 0); o0 += 1; |
| 256 | } |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | i0 = (const float*) ((uintptr_t) i4 - input_decrement); |
| 261 | i1 = (const float*) ((uintptr_t) i0 + input_width); |
| 262 | i2 = (const float*) ((uintptr_t) i1 + input_width); |
| 263 | i3 = (const float*) ((uintptr_t) i2 + input_width); |
| 264 | i4 = (const float*) ((uintptr_t) i3 + input_width); |
| 265 | |
| 266 | o0 = o1; |
| 267 | o1 = (float*) ((uintptr_t) o0 + output_width); |
| 268 | |
| 269 | output_height = doz(output_height, 2); |
| 270 | padded_input_height = doz(padded_input_height, 4); |
| 271 | } while (output_height != 0); |
| 272 | } |