|
| 1 | +// Auto-generated file. Do not edit! |
| 2 | +// Template: src/f32-rdsum/wasm-simd.c.in |
| 3 | +// Generator: tools/xngen |
| 4 | +// |
| 5 | +// Copyright 2024 Google LLC // |
| 6 | +// This source code is licensed under the BSD-style license found in the |
| 7 | +// LICENSE file in the root directory of this source tree. |
| 8 | + |
| 9 | +#include <assert.h> |
| 10 | + |
| 11 | +#include <wasm_simd128.h> |
| 12 | + |
| 13 | +#include <xnnpack/common.h> |
| 14 | +#include <xnnpack/reduce.h> |
| 15 | +#include <xnnpack/math.h> |
| 16 | + |
| 17 | +void xnn_f32_rdsum_ukernel_7p7x__wasmsimd_c16( |
| 18 | + size_t rows, |
| 19 | + size_t channels, |
| 20 | + const float* input, |
| 21 | + size_t input_stride, |
| 22 | + const float* zero, |
| 23 | + float* output, |
| 24 | + const union xnn_f32_scale_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 25 | +{ |
| 26 | + assert(rows != 0); |
| 27 | + assert(channels != 0); |
| 28 | + assert(input != NULL); |
| 29 | + assert(output != NULL); |
| 30 | + |
| 31 | + const v128_t vscale = wasm_v128_load32_splat(¶ms->scalar.scale); |
| 32 | + |
| 33 | + size_t input_increment = 7 * input_stride; |
| 34 | + for (; channels >= 16; channels -= 16) { |
| 35 | + const float* i0 = input; |
| 36 | + const float* i1 = (const float*) ((uintptr_t) input + 1 * input_stride); |
| 37 | + const float* i2 = (const float*) ((uintptr_t) input + 2 * input_stride); |
| 38 | + const float* i3 = (const float*) ((uintptr_t) input + 3 * input_stride); |
| 39 | + const float* i4 = (const float*) ((uintptr_t) input + 4 * input_stride); |
| 40 | + const float* i5 = (const float*) ((uintptr_t) input + 5 * input_stride); |
| 41 | + const float* i6 = (const float*) ((uintptr_t) input + 6 * input_stride); |
| 42 | + |
| 43 | + v128_t vacc0 = wasm_i32x4_const_splat(0.f); |
| 44 | + v128_t vacc1 = wasm_i32x4_const_splat(0.f); |
| 45 | + v128_t vacc2 = wasm_i32x4_const_splat(0.f); |
| 46 | + v128_t vacc3 = wasm_i32x4_const_splat(0.f); |
| 47 | + |
| 48 | + for (int r = rows; r > 0; r -= 7) { |
| 49 | + if XNN_UNPREDICTABLE(r < 2) { |
| 50 | + i1 = zero; |
| 51 | + } |
| 52 | + if XNN_UNPREDICTABLE(r <= 2) { |
| 53 | + i2 = zero; |
| 54 | + } |
| 55 | + if XNN_UNPREDICTABLE(r < 4) { |
| 56 | + i3 = zero; |
| 57 | + } |
| 58 | + if XNN_UNPREDICTABLE(r <= 4) { |
| 59 | + i4 = zero; |
| 60 | + } |
| 61 | + if XNN_UNPREDICTABLE(r < 6) { |
| 62 | + i5 = zero; |
| 63 | + } |
| 64 | + if XNN_UNPREDICTABLE(r <= 6) { |
| 65 | + i6 = zero; |
| 66 | + } |
| 67 | + v128_t vin0; |
| 68 | + v128_t vin1; |
| 69 | + v128_t vin2; |
| 70 | + v128_t vin3; |
| 71 | + vin0 = wasm_v128_load(&i0[0]); |
| 72 | + vin1 = wasm_v128_load(&i0[4]); |
| 73 | + vin2 = wasm_v128_load(&i0[8]); |
| 74 | + vin3 = wasm_v128_load(&i0[12]); |
| 75 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 76 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 77 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 78 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 79 | + vin0 = wasm_v128_load(&i1[0]); |
| 80 | + vin1 = wasm_v128_load(&i1[4]); |
| 81 | + vin2 = wasm_v128_load(&i1[8]); |
| 82 | + vin3 = wasm_v128_load(&i1[12]); |
| 83 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 84 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 85 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 86 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 87 | + vin0 = wasm_v128_load(&i2[0]); |
| 88 | + vin1 = wasm_v128_load(&i2[4]); |
| 89 | + vin2 = wasm_v128_load(&i2[8]); |
| 90 | + vin3 = wasm_v128_load(&i2[12]); |
| 91 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 92 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 93 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 94 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 95 | + vin0 = wasm_v128_load(&i3[0]); |
| 96 | + vin1 = wasm_v128_load(&i3[4]); |
| 97 | + vin2 = wasm_v128_load(&i3[8]); |
| 98 | + vin3 = wasm_v128_load(&i3[12]); |
| 99 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 100 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 101 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 102 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 103 | + vin0 = wasm_v128_load(&i4[0]); |
| 104 | + vin1 = wasm_v128_load(&i4[4]); |
| 105 | + vin2 = wasm_v128_load(&i4[8]); |
| 106 | + vin3 = wasm_v128_load(&i4[12]); |
| 107 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 108 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 109 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 110 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 111 | + vin0 = wasm_v128_load(&i5[0]); |
| 112 | + vin1 = wasm_v128_load(&i5[4]); |
| 113 | + vin2 = wasm_v128_load(&i5[8]); |
| 114 | + vin3 = wasm_v128_load(&i5[12]); |
| 115 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 116 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 117 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 118 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 119 | + vin0 = wasm_v128_load(&i6[0]); |
| 120 | + vin1 = wasm_v128_load(&i6[4]); |
| 121 | + vin2 = wasm_v128_load(&i6[8]); |
| 122 | + vin3 = wasm_v128_load(&i6[12]); |
| 123 | + vacc0 = wasm_f32x4_add(vin0, vacc0); |
| 124 | + vacc1 = wasm_f32x4_add(vin1, vacc1); |
| 125 | + vacc2 = wasm_f32x4_add(vin2, vacc2); |
| 126 | + vacc3 = wasm_f32x4_add(vin3, vacc3); |
| 127 | + i0 = (const float*) ((uintptr_t) i0 + input_increment); |
| 128 | + i1 = (const float*) ((uintptr_t) i1 + input_increment); |
| 129 | + i2 = (const float*) ((uintptr_t) i2 + input_increment); |
| 130 | + i3 = (const float*) ((uintptr_t) i3 + input_increment); |
| 131 | + i4 = (const float*) ((uintptr_t) i4 + input_increment); |
| 132 | + i5 = (const float*) ((uintptr_t) i5 + input_increment); |
| 133 | + i6 = (const float*) ((uintptr_t) i6 + input_increment); |
| 134 | + } |
| 135 | + vacc0 = wasm_f32x4_mul(vacc0, vscale); |
| 136 | + vacc1 = wasm_f32x4_mul(vacc1, vscale); |
| 137 | + vacc2 = wasm_f32x4_mul(vacc2, vscale); |
| 138 | + vacc3 = wasm_f32x4_mul(vacc3, vscale); |
| 139 | + |
| 140 | + const float* o = output; |
| 141 | + v128_t vo0 = wasm_v128_load(o); o += 4; |
| 142 | + v128_t vo1 = wasm_v128_load(o); o += 4; |
| 143 | + v128_t vo2 = wasm_v128_load(o); o += 4; |
| 144 | + v128_t vo3 = wasm_v128_load(o); o += 4; |
| 145 | + vacc0 = wasm_f32x4_add(vo0, vacc0); |
| 146 | + vacc1 = wasm_f32x4_add(vo1, vacc1); |
| 147 | + vacc2 = wasm_f32x4_add(vo2, vacc2); |
| 148 | + vacc3 = wasm_f32x4_add(vo3, vacc3); |
| 149 | + wasm_v128_store(output, vacc0); output += 4; |
| 150 | + wasm_v128_store(output, vacc1); output += 4; |
| 151 | + wasm_v128_store(output, vacc2); output += 4; |
| 152 | + wasm_v128_store(output, vacc3); output += 4; |
| 153 | + |
| 154 | + input = (const float*) ((uintptr_t) input + 16 * sizeof(float)); |
| 155 | + } |
| 156 | + if (channels != 0) { |
| 157 | + input_increment = 7 * input_stride; |
| 158 | + const float* i0 = input; |
| 159 | + const float* i1 = (const float*) ((uintptr_t) input + 1 * input_stride); |
| 160 | + const float* i2 = (const float*) ((uintptr_t) input + 2 * input_stride); |
| 161 | + const float* i3 = (const float*) ((uintptr_t) input + 3 * input_stride); |
| 162 | + const float* i4 = (const float*) ((uintptr_t) input + 4 * input_stride); |
| 163 | + const float* i5 = (const float*) ((uintptr_t) input + 5 * input_stride); |
| 164 | + const float* i6 = (const float*) ((uintptr_t) input + 6 * input_stride); |
| 165 | + v128_t vacc[4]; |
| 166 | + vacc[0] = wasm_i32x4_const_splat(0.f); |
| 167 | + vacc[1] = wasm_i32x4_const_splat(0.f); |
| 168 | + vacc[2] = wasm_i32x4_const_splat(0.f); |
| 169 | + vacc[3] = wasm_i32x4_const_splat(0.f); |
| 170 | + |
| 171 | + const size_t num_chunks = round_up_po2(channels, 4) >> 2; |
| 172 | + for (int r = rows; r > 0; r -= 7) { |
| 173 | + if XNN_UNPREDICTABLE(r < 2) { |
| 174 | + i1 = zero; |
| 175 | + } |
| 176 | + if XNN_UNPREDICTABLE(r <= 2) { |
| 177 | + i2 = zero; |
| 178 | + } |
| 179 | + if XNN_UNPREDICTABLE(r < 4) { |
| 180 | + i3 = zero; |
| 181 | + } |
| 182 | + if XNN_UNPREDICTABLE(r <= 4) { |
| 183 | + i4 = zero; |
| 184 | + } |
| 185 | + if XNN_UNPREDICTABLE(r < 6) { |
| 186 | + i5 = zero; |
| 187 | + } |
| 188 | + if XNN_UNPREDICTABLE(r <= 6) { |
| 189 | + i6 = zero; |
| 190 | + } |
| 191 | + for (int i = 0; i < num_chunks; ++i) { |
| 192 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i0[i*4]), vacc[i]); |
| 193 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i1[i*4]), vacc[i]); |
| 194 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i2[i*4]), vacc[i]); |
| 195 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i3[i*4]), vacc[i]); |
| 196 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i4[i*4]), vacc[i]); |
| 197 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i5[i*4]), vacc[i]); |
| 198 | + vacc[i] = wasm_f32x4_add(wasm_v128_load(&i6[i*4]), vacc[i]); |
| 199 | + } |
| 200 | + i0 = (const float*) ((uintptr_t) i0 + input_increment); |
| 201 | + i1 = (const float*) ((uintptr_t) i1 + input_increment); |
| 202 | + i2 = (const float*) ((uintptr_t) i2 + input_increment); |
| 203 | + i3 = (const float*) ((uintptr_t) i3 + input_increment); |
| 204 | + i4 = (const float*) ((uintptr_t) i4 + input_increment); |
| 205 | + i5 = (const float*) ((uintptr_t) i5 + input_increment); |
| 206 | + i6 = (const float*) ((uintptr_t) i6 + input_increment); |
| 207 | + } |
| 208 | + for (int i = 0; i < num_chunks; ++i) { |
| 209 | + vacc[i] = wasm_f32x4_mul(vacc[i], vscale); |
| 210 | + } |
| 211 | + |
| 212 | + v128_t vo[4]; |
| 213 | + const float* o = output; |
| 214 | + for (int i = 0; i < channels >> 2; ++i) { |
| 215 | + vo[i] = wasm_v128_load(o); o += 4; |
| 216 | + } |
| 217 | + for (int i = 0; i < channels >> 2; ++i) { |
| 218 | + vacc[i] = wasm_f32x4_add(vo[i], vacc[i]); |
| 219 | + } |
| 220 | + for (int i = 0; i < channels >> 2; ++i) { |
| 221 | + wasm_v128_store(output, vacc[i]); output += 4; |
| 222 | + } |
| 223 | + const size_t pos = channels / 4; |
| 224 | + v128_t vout = vacc[pos]; |
| 225 | + if (channels & 2) { |
| 226 | + v128_t vo = wasm_f32x4_make(output[0], output[1], 0.f, 0.f); |
| 227 | + wasm_v128_store64_lane(output, wasm_f32x4_add(vo, vout), 0); |
| 228 | + vout = wasm_v64x2_shuffle(vout, vout, 1, 1); |
| 229 | + output += 2; |
| 230 | + } |
| 231 | + if (channels & 1) { |
| 232 | + *output += wasm_f32x4_extract_lane(vout, 0); |
| 233 | + } |
| 234 | + } |
| 235 | +} |
0 commit comments