Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-spmm/wasmsimd.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/spmm.h> |
| 15 | |
| 16 | |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame^] | 17 | void xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_x2( |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 18 | uint32_t batch_size, |
| 19 | uint32_t output_channels, |
| 20 | const float*restrict input, |
| 21 | const float*restrict weights, |
| 22 | const int32_t*restrict widx_dmap, |
| 23 | const uint32_t*restrict nidx_nnzmap, |
| 24 | float*restrict output, |
| 25 | const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 26 | { |
| 27 | assert(batch_size != 0); |
| 28 | |
| 29 | const v128_t vmin = wasm_v32x4_load_splat(¶ms->scalar.min); |
| 30 | const v128_t vmax = wasm_v32x4_load_splat(¶ms->scalar.max); |
Frank Barchard | fea2680 | 2020-10-22 10:35:40 -0700 | [diff] [blame] | 31 | const v128_t vzero = wasm_f32x4_splat(0.0f); |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 32 | size_t n = batch_size; |
| 33 | while XNN_LIKELY(n >= 16) { |
| 34 | const float*restrict w = weights; |
| 35 | const int32_t* dmap = widx_dmap; |
| 36 | const uint32_t* nnzmap = nidx_nnzmap; |
| 37 | size_t c = output_channels; |
| 38 | do { |
| 39 | uint32_t nnz = *nnzmap++; |
| 40 | v128_t vacc0123x0 = wasm_v32x4_load_splat(w); |
| 41 | w += 1; |
Frank Barchard | fea2680 | 2020-10-22 10:35:40 -0700 | [diff] [blame] | 42 | v128_t vacc0123x1 = vzero; |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 43 | v128_t vacc4567x0 = vacc0123x0; |
Frank Barchard | fea2680 | 2020-10-22 10:35:40 -0700 | [diff] [blame] | 44 | v128_t vacc4567x1 = vzero; |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 45 | v128_t vacc89ABx0 = vacc0123x0; |
Frank Barchard | fea2680 | 2020-10-22 10:35:40 -0700 | [diff] [blame] | 46 | v128_t vacc89ABx1 = vzero; |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 47 | v128_t vaccCDEFx0 = vacc0123x0; |
Frank Barchard | fea2680 | 2020-10-22 10:35:40 -0700 | [diff] [blame] | 48 | v128_t vaccCDEFx1 = vzero; |
Frank Barchard | c451e8a | 2020-10-21 17:13:12 -0700 | [diff] [blame] | 49 | for (; nnz >= 2; nnz -= 2) { |
| 50 | const intptr_t diff0 = dmap[0]; |
| 51 | const intptr_t diff1 = dmap[1]; |
| 52 | dmap += 2; |
| 53 | const v128_t vi0123x0 = wasm_v128_load(input); |
| 54 | const v128_t vi4567x0 = wasm_v128_load(input + 4); |
| 55 | const v128_t vi89ABx0 = wasm_v128_load(input + 8); |
| 56 | const v128_t viCDEFx0 = wasm_v128_load(input + 12); |
| 57 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff0); |
| 58 | const v128_t vw0 = wasm_v32x4_load_splat(w); |
| 59 | w += 1; |
| 60 | vacc0123x0 = wasm_f32x4_add(vacc0123x0, wasm_f32x4_mul(vi0123x0, vw0)); |
| 61 | vacc4567x0 = wasm_f32x4_add(vacc4567x0, wasm_f32x4_mul(vi4567x0, vw0)); |
| 62 | vacc89ABx0 = wasm_f32x4_add(vacc89ABx0, wasm_f32x4_mul(vi89ABx0, vw0)); |
| 63 | vaccCDEFx0 = wasm_f32x4_add(vaccCDEFx0, wasm_f32x4_mul(viCDEFx0, vw0)); |
| 64 | const v128_t vi0123x1 = wasm_v128_load(input); |
| 65 | const v128_t vi4567x1 = wasm_v128_load(input + 4); |
| 66 | const v128_t vi89ABx1 = wasm_v128_load(input + 8); |
| 67 | const v128_t viCDEFx1 = wasm_v128_load(input + 12); |
| 68 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff1); |
| 69 | const v128_t vw1 = wasm_v32x4_load_splat(w); |
| 70 | w += 1; |
| 71 | vacc0123x1 = wasm_f32x4_add(vacc0123x1, wasm_f32x4_mul(vi0123x1, vw1)); |
| 72 | vacc4567x1 = wasm_f32x4_add(vacc4567x1, wasm_f32x4_mul(vi4567x1, vw1)); |
| 73 | vacc89ABx1 = wasm_f32x4_add(vacc89ABx1, wasm_f32x4_mul(vi89ABx1, vw1)); |
| 74 | vaccCDEFx1 = wasm_f32x4_add(vaccCDEFx1, wasm_f32x4_mul(viCDEFx1, vw1)); |
| 75 | } |
| 76 | v128_t vacc0123 = vacc0123x0; |
| 77 | v128_t vacc4567 = vacc4567x0; |
| 78 | v128_t vacc89AB = vacc89ABx0; |
| 79 | v128_t vaccCDEF = vaccCDEFx0; |
| 80 | vacc0123 = wasm_f32x4_add(vacc0123, vacc0123x1); |
| 81 | vacc4567 = wasm_f32x4_add(vacc4567, vacc4567x1); |
| 82 | vacc89AB = wasm_f32x4_add(vacc89AB, vacc89ABx1); |
| 83 | vaccCDEF = wasm_f32x4_add(vaccCDEF, vaccCDEFx1); |
| 84 | if XNN_LIKELY(nnz != 0) { |
| 85 | do { |
| 86 | const intptr_t diff = *dmap++; |
| 87 | const v128_t vi0123 = wasm_v128_load(input); |
| 88 | const v128_t vi4567 = wasm_v128_load(input + 4); |
| 89 | const v128_t vi89AB = wasm_v128_load(input + 8); |
| 90 | const v128_t viCDEF = wasm_v128_load(input + 12); |
| 91 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff); |
| 92 | const v128_t vw = wasm_v32x4_load_splat(w); w += 1; |
| 93 | vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw)); |
| 94 | vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw)); |
| 95 | vacc89AB = wasm_f32x4_add(vacc89AB, wasm_f32x4_mul(vi89AB, vw)); |
| 96 | vaccCDEF = wasm_f32x4_add(vaccCDEF, wasm_f32x4_mul(viCDEF, vw)); |
| 97 | } while (--nnz != 0); |
| 98 | } |
| 99 | v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); |
| 100 | v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); |
| 101 | v128_t vout89AB = wasm_v128_bitselect(vacc89AB, vmax, wasm_f32x4_le(vacc89AB, vmax)); |
| 102 | v128_t voutCDEF = wasm_v128_bitselect(vaccCDEF, vmax, wasm_f32x4_le(vaccCDEF, vmax)); |
| 103 | vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); |
| 104 | vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin)); |
| 105 | vout89AB = wasm_v128_bitselect(vmin, vout89AB, wasm_f32x4_lt(vout89AB, vmin)); |
| 106 | voutCDEF = wasm_v128_bitselect(vmin, voutCDEF, wasm_f32x4_lt(voutCDEF, vmin)); |
| 107 | wasm_v128_store(output, vout0123); |
| 108 | wasm_v128_store(output + 4, vout4567); |
| 109 | wasm_v128_store(output + 8, vout89AB); |
| 110 | wasm_v128_store(output + 12, voutCDEF); |
| 111 | output += 1 * batch_size; |
| 112 | } while (--c != 0); |
| 113 | output -= batch_size * output_channels; |
| 114 | output += 16; |
| 115 | input += 16; |
| 116 | n -= 16; |
| 117 | } |
| 118 | if XNN_UNLIKELY(n != 0) { |
| 119 | if (n & 8) { |
| 120 | const float*restrict w = weights; |
| 121 | const int32_t* dmap = widx_dmap; |
| 122 | const uint32_t* nnzmap = nidx_nnzmap; |
| 123 | size_t c = output_channels; |
| 124 | do { |
| 125 | uint32_t nnz = *nnzmap++; |
| 126 | v128_t vacc0123 = wasm_v32x4_load_splat(w); w += 1; |
| 127 | v128_t vacc4567 = vacc0123; |
| 128 | if XNN_LIKELY(nnz != 0) { |
| 129 | do { |
| 130 | const intptr_t diff = *dmap++; |
| 131 | const v128_t vi0123 = wasm_v128_load(input); |
| 132 | const v128_t vi4567 = wasm_v128_load(input + 4); |
| 133 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff); |
| 134 | const v128_t vw = wasm_v32x4_load_splat(w); w += 1; |
| 135 | vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw)); |
| 136 | vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw)); |
| 137 | } while (--nnz != 0); |
| 138 | } |
| 139 | v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); |
| 140 | v128_t vout4567 = wasm_v128_bitselect(vacc4567, vmax, wasm_f32x4_le(vacc4567, vmax)); |
| 141 | vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); |
| 142 | vout4567 = wasm_v128_bitselect(vmin, vout4567, wasm_f32x4_lt(vout4567, vmin)); |
| 143 | wasm_v128_store(output, vout0123); |
| 144 | |
| 145 | wasm_v128_store(output + 4, vout4567); |
| 146 | output += 1 * batch_size; |
| 147 | } while (--c != 0); |
| 148 | output -= batch_size * output_channels; |
| 149 | output += 8; |
| 150 | input += 8; |
| 151 | } |
| 152 | if (n & 4) { |
| 153 | const float*restrict w = weights; |
| 154 | const int32_t* dmap = widx_dmap; |
| 155 | const uint32_t* nnzmap = nidx_nnzmap; |
| 156 | size_t c = output_channels; |
| 157 | do { |
| 158 | uint32_t nnz = *nnzmap++; |
| 159 | v128_t vacc0123 = wasm_v32x4_load_splat(w); w += 1; |
| 160 | if XNN_LIKELY(nnz != 0) { |
| 161 | do { |
| 162 | const intptr_t diff = *dmap++; |
| 163 | const v128_t vi0123 = wasm_v128_load(input); |
| 164 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff); |
| 165 | const v128_t vw = wasm_v32x4_load_splat(w); w += 1; |
| 166 | vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw)); |
| 167 | } while (--nnz != 0); |
| 168 | } |
| 169 | v128_t vout0123 = wasm_v128_bitselect(vacc0123, vmax, wasm_f32x4_le(vacc0123, vmax)); |
| 170 | vout0123 = wasm_v128_bitselect(vmin, vout0123, wasm_f32x4_lt(vout0123, vmin)); |
| 171 | wasm_v128_store(output, vout0123); |
| 172 | |
| 173 | output += 1 * batch_size; |
| 174 | } while (--c != 0); |
| 175 | output -= batch_size * output_channels; |
| 176 | output += 4; |
| 177 | input += 4; |
| 178 | } |
| 179 | if (n & 2) { |
| 180 | const float*restrict w = weights; |
| 181 | const int32_t* dmap = widx_dmap; |
| 182 | const uint32_t* nnzmap = nidx_nnzmap; |
| 183 | size_t c = output_channels; |
| 184 | do { |
| 185 | uint32_t nnz = *nnzmap++; |
| 186 | v128_t vacc01 = wasm_v32x4_load_splat(w); w += 1; |
| 187 | if XNN_LIKELY(nnz != 0) { |
| 188 | do { |
| 189 | const intptr_t diff = *dmap++; |
| 190 | const v128_t vi01 = wasm_v64x2_load_splat(input); |
| 191 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff); |
| 192 | const v128_t vw = wasm_v32x4_load_splat(w); w += 1; |
| 193 | vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw)); |
| 194 | } while (--nnz != 0); |
| 195 | } |
| 196 | v128_t vout01 = wasm_v128_bitselect(vacc01, vmax, wasm_f32x4_le(vacc01, vmax)); |
| 197 | vout01 = wasm_v128_bitselect(vmin, vout01, wasm_f32x4_lt(vout01, vmin)); |
| 198 | *((double*) output) = wasm_f64x2_extract_lane(vout01, 0); |
| 199 | |
| 200 | output += 1 * batch_size; |
| 201 | } while (--c != 0); |
| 202 | output -= batch_size * output_channels; |
| 203 | output += 2; |
| 204 | input += 2; |
| 205 | } |
| 206 | if (n & 1) { |
| 207 | const float*restrict w = weights; |
| 208 | const int32_t* dmap = widx_dmap; |
| 209 | const uint32_t* nnzmap = nidx_nnzmap; |
| 210 | size_t c = output_channels; |
| 211 | do { |
| 212 | uint32_t nnz = *nnzmap++; |
| 213 | v128_t vacc0 = wasm_v32x4_load_splat(w); w += 1; |
| 214 | if XNN_LIKELY(nnz != 0) { |
| 215 | do { |
| 216 | const intptr_t diff = *dmap++; |
| 217 | const v128_t vi0 = wasm_v32x4_load_splat(input); |
| 218 | input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff); |
| 219 | const v128_t vw = wasm_v32x4_load_splat(w); w += 1; |
| 220 | vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw)); |
| 221 | } while (--nnz != 0); |
| 222 | } |
| 223 | v128_t vout0 = wasm_v128_bitselect(vacc0, vmax, wasm_f32x4_le(vacc0, vmax)); |
| 224 | vout0 = wasm_v128_bitselect(vmin, vout0, wasm_f32x4_lt(vout0, vmin)); |
| 225 | *output = wasm_f32x4_extract_lane(vout0, 0); |
| 226 | |
| 227 | output += 1 * batch_size; |
| 228 | } while (--c != 0); |
| 229 | output -= batch_size * output_channels; |
| 230 | output += 1; |
| 231 | input += 1; |
| 232 | } |
| 233 | } |
| 234 | } |