Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1 | // Copyright (C) 2020 The Android Open Source Project |
| 2 | // |
| 3 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | // you may not use this file except in compliance with the License. |
| 5 | // You may obtain a copy of the License at |
| 6 | // |
| 7 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | // |
| 9 | // Unless required by applicable law or agreed to in writing, software |
| 10 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | // See the License for the specific language governing permissions and |
| 13 | // limitations under the License. |
| 14 | |
| 15 | OPERATOR_SRCS = [ |
| 16 | "src/add-nc.c", |
| 17 | "src/argmax-pooling-nhwc.c", |
| 18 | "src/average-pooling-nhwc.c", |
| 19 | "src/binary-elementwise-nd.c", |
| 20 | "src/channel-pad-nc.c", |
| 21 | "src/channel-shuffle-nc.c", |
| 22 | "src/clamp-nc.c", |
| 23 | "src/convolution-nchw.c", |
| 24 | "src/convolution-nhwc.c", |
| 25 | "src/deconvolution-nhwc.c", |
| 26 | "src/fully-connected-nc.c", |
| 27 | "src/global-average-pooling-ncw.c", |
| 28 | "src/global-average-pooling-nwc.c", |
| 29 | "src/hardswish-nc.c", |
| 30 | "src/leaky-relu-nc.c", |
| 31 | "src/max-pooling-nhwc.c", |
| 32 | "src/prelu-nc.c", |
| 33 | "src/resize-bilinear-nhwc.c", |
| 34 | "src/sigmoid-nc.c", |
| 35 | "src/softmax-nc.c", |
| 36 | "src/unpooling-nhwc.c", |
| 37 | ] |
| 38 | |
| 39 | TABLE_SRCS = [ |
| 40 | "src/tables/exp2-k-over-64.c", |
| 41 | "src/tables/exp2-k-over-2048.c", |
| 42 | ] |
| 43 | |
| 44 | SCALAR_UKERNELS = [ |
| 45 | "src/f32-argmaxpool/4x-scalar-c1.c", |
| 46 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
| 47 | "src/f32-argmaxpool/9x-scalar-c1.c", |
| 48 | "src/f32-avgpool/mp9p8q-scalar.c", |
| 49 | "src/f32-avgpool/up9-scalar.c", |
| 50 | "src/f32-bilinear/gen/scalar-c1.c", |
| 51 | "src/f32-bilinear/gen/scalar-c2.c", |
| 52 | "src/f32-bilinear/gen/scalar-c4.c", |
| 53 | "src/f32-clamp/scalar.c", |
| 54 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
| 55 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
| 56 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c", |
| 57 | "src/f32-dwconv-spchw/3x3p1-scalar.c", |
| 58 | "src/f32-dwconv-spchw/3x3s2p1-scalar.c", |
| 59 | "src/f32-dwconv-spchw/5x5p2-scalar.c", |
| 60 | "src/f32-dwconv-spchw/5x5s2p2-scalar.c", |
| 61 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 62 | "src/f32-dwconv/gen/up1x25-scalar.c", |
| 63 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 64 | "src/f32-dwconv/gen/up1x4-scalar.c", |
| 65 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 66 | "src/f32-dwconv/gen/up1x9-scalar.c", |
| 67 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 68 | "src/f32-dwconv/gen/up2x25-scalar.c", |
| 69 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 70 | "src/f32-dwconv/gen/up2x4-scalar.c", |
| 71 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 72 | "src/f32-dwconv/gen/up2x9-scalar.c", |
| 73 | "src/f32-gavgpool-spchw/scalar-x1.c", |
| 74 | "src/f32-gavgpool/mp7p7q-scalar.c", |
| 75 | "src/f32-gavgpool/up7-scalar.c", |
| 76 | "src/f32-gemm/gen-inc/1x4-scalar.c", |
| 77 | "src/f32-gemm/gen-inc/2x4-scalar.c", |
| 78 | "src/f32-gemm/gen-inc/4x4-scalar.c", |
| 79 | "src/f32-gemm/gen/1x4-scalar.c", |
| 80 | "src/f32-gemm/gen/2x4-scalar.c", |
| 81 | "src/f32-gemm/gen/4x2-scalar.c", |
| 82 | "src/f32-gemm/gen/4x4-scalar.c", |
| 83 | "src/f32-hswish/gen/scalar-x1.c", |
| 84 | "src/f32-hswish/gen/scalar-x2.c", |
| 85 | "src/f32-hswish/gen/scalar-x4.c", |
| 86 | "src/f32-igemm/gen/1x4-scalar.c", |
| 87 | "src/f32-igemm/gen/2x4-scalar.c", |
| 88 | "src/f32-igemm/gen/4x2-scalar.c", |
| 89 | "src/f32-igemm/gen/4x4-scalar.c", |
| 90 | "src/f32-maxpool/9p8x-scalar-c1.c", |
| 91 | "src/f32-pavgpool/mp9p8q-scalar.c", |
| 92 | "src/f32-pavgpool/up9-scalar.c", |
| 93 | "src/f32-ppmm/gen/2x4-scalar.c", |
| 94 | "src/f32-ppmm/gen/3x3-scalar.c", |
| 95 | "src/f32-ppmm/gen/4x2-scalar.c", |
| 96 | "src/f32-ppmm/gen/4x4-scalar.c", |
| 97 | "src/f32-prelu/gen/scalar-2x1.c", |
| 98 | "src/f32-prelu/gen/scalar-2x4.c", |
| 99 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
| 100 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
| 101 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
| 102 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
| 103 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 104 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
| 105 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
| 106 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
| 107 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
| 108 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
| 109 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 110 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
| 111 | "src/f32-rmax/scalar.c", |
| 112 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 113 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 114 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
| 115 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 116 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 117 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
| 118 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 119 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 120 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
| 121 | "src/f32-spmm/gen/1x1-scalar-pipelined.c", |
| 122 | "src/f32-spmm/gen/1x1-scalar.c", |
| 123 | "src/f32-spmm/gen/2x1-scalar-pipelined.c", |
| 124 | "src/f32-spmm/gen/2x1-scalar.c", |
| 125 | "src/f32-spmm/gen/4x1-scalar-pipelined.c", |
| 126 | "src/f32-spmm/gen/4x1-scalar.c", |
| 127 | "src/f32-spmm/gen/8x1-scalar-pipelined.c", |
| 128 | "src/f32-spmm/gen/8x1-scalar.c", |
| 129 | "src/f32-spmm/gen/8x2-scalar.c", |
| 130 | "src/f32-spmm/gen/8x4-scalar.c", |
| 131 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 132 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 133 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
| 134 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 135 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 136 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
| 137 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 138 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 139 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
| 140 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 141 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 142 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
| 143 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 144 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 145 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| 146 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 147 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 148 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| 149 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 150 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 151 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| 152 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 153 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 154 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
| 155 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 156 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 157 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
| 158 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 159 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 160 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
| 161 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 162 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 163 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
| 164 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 165 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 166 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
| 167 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 168 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 169 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
| 170 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 171 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 172 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
| 173 | "src/f32-vmulcaddc/gen/c1-scalar-2x.c", |
| 174 | "src/f32-vmulcaddc/gen/c2-scalar-2x.c", |
| 175 | "src/f32-vmulcaddc/gen/c4-scalar-2x.c", |
| 176 | "src/math/expminus-scalar-lut2048-p1.c", |
| 177 | "src/math/expminus-scalar-lut64-p2.c", |
| 178 | "src/math/expminus-scalar-p5.c", |
| 179 | "src/math/sigmoid-scalar-lut2048-p1-div.c", |
| 180 | "src/math/sigmoid-scalar-lut64-p2-div.c", |
| 181 | "src/math/sigmoid-scalar-p5-div.c", |
| 182 | "src/q8-avgpool/mp9p8q-scalar.c", |
| 183 | "src/q8-avgpool/up9-scalar.c", |
| 184 | "src/q8-dwconv/up1x9-scalar.c", |
| 185 | "src/q8-gavgpool/mp7p7q-scalar.c", |
| 186 | "src/q8-gavgpool/up7-scalar.c", |
| 187 | "src/q8-gemm/2x2-scalar.c", |
| 188 | "src/q8-igemm/2x2-scalar.c", |
| 189 | "src/q8-vadd/scalar.c", |
| 190 | "src/u8-clamp/scalar.c", |
| 191 | "src/u8-lut32norm/scalar.c", |
| 192 | "src/u8-maxpool/9p8x-scalar-c1.c", |
| 193 | "src/u8-rmax/scalar.c", |
| 194 | "src/x32-packx/x2-scalar.c", |
| 195 | "src/x32-packx/x3-scalar.c", |
| 196 | "src/x32-packx/x4-scalar.c", |
| 197 | "src/x32-pad/x2-scalar.c", |
| 198 | "src/x32-unpool/scalar.c", |
| 199 | "src/x32-zip/x2-scalar.c", |
| 200 | "src/x32-zip/x3-scalar.c", |
| 201 | "src/x32-zip/x4-scalar.c", |
| 202 | "src/x32-zip/xm-scalar.c", |
| 203 | "src/x8-lut/scalar.c", |
| 204 | "src/x8-zip/x2-scalar.c", |
| 205 | "src/x8-zip/x3-scalar.c", |
| 206 | "src/x8-zip/x4-scalar.c", |
| 207 | "src/x8-zip/xm-scalar.c", |
| 208 | ] |
| 209 | |
| 210 | PSIMD_FASTMATH_UKERNELS = [ |
| 211 | "src/f32-argmaxpool/4x-psimd-c4.c", |
| 212 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
| 213 | "src/f32-argmaxpool/9x-psimd-c4.c", |
| 214 | "src/f32-avgpool/mp9p8q-psimd.c", |
| 215 | "src/f32-avgpool/up9-psimd.c", |
| 216 | "src/f32-bilinear/gen/psimd-c4.c", |
| 217 | "src/f32-bilinear/gen/psimd-c8.c", |
| 218 | "src/f32-clamp/psimd.c", |
| 219 | "src/f32-dwconv/gen/up4x25-psimd-acc2.c", |
| 220 | "src/f32-dwconv/gen/up4x25-psimd.c", |
| 221 | "src/f32-dwconv/gen/up4x4-psimd-acc2.c", |
| 222 | "src/f32-dwconv/gen/up4x4-psimd.c", |
| 223 | "src/f32-dwconv/gen/up4x9-psimd-acc2.c", |
| 224 | "src/f32-dwconv/gen/up4x9-psimd.c", |
| 225 | "src/f32-dwconv/gen/up8x25-psimd-acc2.c", |
| 226 | "src/f32-dwconv/gen/up8x25-psimd.c", |
| 227 | "src/f32-dwconv/gen/up8x4-psimd-acc2.c", |
| 228 | "src/f32-dwconv/gen/up8x4-psimd.c", |
| 229 | "src/f32-dwconv/gen/up8x9-psimd-acc2.c", |
| 230 | "src/f32-dwconv/gen/up8x9-psimd.c", |
| 231 | "src/f32-gavgpool/mp7p7q-psimd.c", |
| 232 | "src/f32-gavgpool/up7-psimd.c", |
| 233 | "src/f32-gemm/gen/1x8-psimd-loadsplat.c", |
| 234 | "src/f32-gemm/gen/1x8-psimd-splat.c", |
| 235 | "src/f32-gemm/gen/1x8s4-psimd.c", |
| 236 | "src/f32-gemm/gen/4x8-psimd-loadsplat.c", |
| 237 | "src/f32-gemm/gen/4x8-psimd-splat.c", |
| 238 | "src/f32-gemm/gen/4x8s4-psimd.c", |
| 239 | "src/f32-gemm/gen/6x8-psimd-loadsplat.c", |
| 240 | "src/f32-gemm/gen/6x8-psimd-splat.c", |
| 241 | "src/f32-gemm/gen/6x8s4-psimd.c", |
| 242 | "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c", |
| 243 | "src/f32-gemm/gen-inc/1x8-psimd-splat.c", |
| 244 | "src/f32-gemm/gen-inc/1x8s4-psimd.c", |
| 245 | "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c", |
| 246 | "src/f32-gemm/gen-inc/4x8-psimd-splat.c", |
| 247 | "src/f32-gemm/gen-inc/4x8s4-psimd.c", |
| 248 | "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c", |
| 249 | "src/f32-gemm/gen-inc/6x8-psimd-splat.c", |
| 250 | "src/f32-gemm/gen-inc/6x8s4-psimd.c", |
| 251 | "src/f32-hswish/gen/psimd-x4.c", |
| 252 | "src/f32-hswish/gen/psimd-x8.c", |
| 253 | "src/f32-igemm/gen/1x8-psimd-loadsplat.c", |
| 254 | "src/f32-igemm/gen/1x8-psimd-splat.c", |
| 255 | "src/f32-igemm/gen/1x8s4-psimd.c", |
| 256 | "src/f32-igemm/gen/4x2c4-psimd.c", |
| 257 | "src/f32-igemm/gen/4x8-psimd-loadsplat.c", |
| 258 | "src/f32-igemm/gen/4x8-psimd-splat.c", |
| 259 | "src/f32-igemm/gen/4x8s4-psimd.c", |
| 260 | "src/f32-igemm/gen/6x8-psimd-loadsplat.c", |
| 261 | "src/f32-igemm/gen/6x8-psimd-splat.c", |
| 262 | "src/f32-igemm/gen/6x8s4-psimd.c", |
| 263 | "src/f32-maxpool/9p8x-psimd-c4.c", |
| 264 | "src/f32-pavgpool/mp9p8q-psimd.c", |
| 265 | "src/f32-pavgpool/up9-psimd.c", |
| 266 | "src/f32-ppmm/gen/4x8-psimd.c", |
| 267 | "src/f32-prelu/gen/psimd-2x4.c", |
| 268 | "src/f32-prelu/gen/psimd-2x8.c", |
| 269 | "src/f32-rmax/psimd.c", |
| 270 | "src/f32-vbinary/gen/vadd-psimd-x4.c", |
| 271 | "src/f32-vbinary/gen/vadd-psimd-x8.c", |
| 272 | "src/f32-vbinary/gen/vaddc-psimd-x4.c", |
| 273 | "src/f32-vbinary/gen/vaddc-psimd-x8.c", |
| 274 | "src/f32-vbinary/gen/vdiv-psimd-x4.c", |
| 275 | "src/f32-vbinary/gen/vdiv-psimd-x8.c", |
| 276 | "src/f32-vbinary/gen/vdivc-psimd-x4.c", |
| 277 | "src/f32-vbinary/gen/vdivc-psimd-x8.c", |
| 278 | "src/f32-vbinary/gen/vmax-psimd-x4.c", |
| 279 | "src/f32-vbinary/gen/vmax-psimd-x8.c", |
| 280 | "src/f32-vbinary/gen/vmaxc-psimd-x4.c", |
| 281 | "src/f32-vbinary/gen/vmaxc-psimd-x8.c", |
| 282 | "src/f32-vbinary/gen/vmin-psimd-x4.c", |
| 283 | "src/f32-vbinary/gen/vmin-psimd-x8.c", |
| 284 | "src/f32-vbinary/gen/vminc-psimd-x4.c", |
| 285 | "src/f32-vbinary/gen/vminc-psimd-x8.c", |
| 286 | "src/f32-vbinary/gen/vmul-psimd-x4.c", |
| 287 | "src/f32-vbinary/gen/vmul-psimd-x8.c", |
| 288 | "src/f32-vbinary/gen/vmulc-psimd-x4.c", |
| 289 | "src/f32-vbinary/gen/vmulc-psimd-x8.c", |
| 290 | "src/f32-vbinary/gen/vrdivc-psimd-x4.c", |
| 291 | "src/f32-vbinary/gen/vrdivc-psimd-x8.c", |
| 292 | "src/f32-vbinary/gen/vrsubc-psimd-x4.c", |
| 293 | "src/f32-vbinary/gen/vrsubc-psimd-x8.c", |
| 294 | "src/f32-vbinary/gen/vsub-psimd-x4.c", |
| 295 | "src/f32-vbinary/gen/vsub-psimd-x8.c", |
| 296 | "src/f32-vbinary/gen/vsubc-psimd-x4.c", |
| 297 | "src/f32-vbinary/gen/vsubc-psimd-x8.c", |
| 298 | "src/f32-vmulcaddc/gen/c4-psimd-2x.c", |
| 299 | "src/f32-vmulcaddc/gen/c8-psimd-2x.c", |
| 300 | "src/x32-packx/x4-psimd.c", |
| 301 | "src/x32-pad/x2-psimd.c", |
| 302 | "src/x32-unpool/psimd.c", |
| 303 | "src/x32-zip/x2-psimd.c", |
| 304 | "src/x32-zip/x3-psimd.c", |
| 305 | "src/x32-zip/x4-psimd.c", |
| 306 | "src/x32-zip/xm-psimd.c", |
| 307 | ] |
| 308 | |
| 309 | PSIMD_ACCMATH_UKERNELS = [ |
| 310 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c", |
| 311 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c", |
| 312 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c", |
| 313 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c", |
| 314 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c", |
| 315 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c", |
| 316 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c", |
| 317 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c", |
| 318 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c", |
| 319 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c", |
| 320 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c", |
| 321 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c", |
| 322 | "src/f32-sigmoid/gen/psimd-p5-div-x4.c", |
| 323 | "src/f32-sigmoid/gen/psimd-p5-div-x8.c", |
| 324 | "src/f32-sigmoid/gen/psimd-p5-div-x12.c", |
| 325 | "src/f32-sigmoid/gen/psimd-p5-div-x16.c", |
| 326 | "src/f32-sigmoid/gen/psimd-p5-div-x20.c", |
| 327 | "src/f32-sigmoid/gen/psimd-p5-div-x24.c", |
| 328 | "src/math/sigmoid-psimd-p5-div.c", |
| 329 | ] |
| 330 | |
| 331 | // ISA-specific micro-kernels |
| 332 | NEON_UKERNELS = [ |
| 333 | "src/f32-avgpool/mp9p8q-neon.c", |
| 334 | "src/f32-avgpool/up9-neon.c", |
| 335 | "src/f32-bilinear/gen/neon-c4.c", |
| 336 | "src/f32-bilinear/gen/neon-c8.c", |
| 337 | "src/f32-clamp/neon.c", |
| 338 | "src/f32-dwconv/gen/up4x9-neon.c", |
| 339 | "src/f32-dwconv/gen/up4x9-neon-acc2.c", |
| 340 | "src/f32-dwconv/gen/up8x9-neon.c", |
| 341 | "src/f32-dwconv/gen/up8x9-neon-acc2.c", |
| 342 | "src/f32-gavgpool-spchw/neon-x4.c", |
| 343 | "src/f32-gavgpool/mp7p7q-neon.c", |
| 344 | "src/f32-gavgpool/up7-neon.c", |
| 345 | "src/f32-gemm/gen/1x8-neon-lane-ld64.c", |
| 346 | "src/f32-gemm/gen/4x2-neon-lane-ld64.c", |
| 347 | "src/f32-gemm/gen/4x8-neon-lane-ld128.c", |
| 348 | "src/f32-gemm/gen/4x8-neon-lane-ld64.c", |
| 349 | "src/f32-gemm/gen/5x8-neon-lane-ld64.c", |
| 350 | "src/f32-gemm/gen/6x8-neon-lane-ld64.c", |
| 351 | "src/f32-gemm/gen/6x8-neon-lane-ld128.c", |
| 352 | "src/f32-gemm/gen/1x8-neon-dup-ld64.c", |
| 353 | "src/f32-gemm/gen/4x8-neon-dup-ld128.c", |
| 354 | "src/f32-gemm/gen/4x8-neon-dup-ld64.c", |
| 355 | "src/f32-gemm/gen/6x8-neon-dup-ld64.c", |
| 356 | "src/f32-gemm/gen/6x8-neon-dup-ld128.c", |
| 357 | "src/f32-gemm/gen/1x8s4-neon.c", |
| 358 | "src/f32-gemm/gen/4x8s4-neon.c", |
| 359 | "src/f32-gemm/gen/6x8s4-neon.c", |
| 360 | "src/f32-gemm/gen/8x8s4-neon.c", |
| 361 | "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c", |
| 362 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c", |
| 363 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c", |
| 364 | "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c", |
| 365 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c", |
| 366 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c", |
| 367 | "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c", |
| 368 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c", |
| 369 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c", |
| 370 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c", |
| 371 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c", |
| 372 | "src/f32-gemm/gen-inc/1x8s4-neon.c", |
| 373 | "src/f32-gemm/gen-inc/4x8s4-neon.c", |
| 374 | "src/f32-gemm/gen-inc/6x8s4-neon.c", |
| 375 | "src/f32-gemm/gen-inc/8x8s4-neon.c", |
| 376 | "src/f32-hswish/gen/neon-x4.c", |
| 377 | "src/f32-hswish/gen/neon-x8.c", |
| 378 | "src/f32-igemm/gen/1x8-neon-lane-ld64.c", |
| 379 | "src/f32-igemm/gen/4x2-neon-lane-ld64.c", |
| 380 | "src/f32-igemm/gen/4x4-neon-lane-ld64.c", |
| 381 | "src/f32-igemm/gen/4x8-neon-lane-ld128.c", |
| 382 | "src/f32-igemm/gen/4x8-neon-lane-ld64.c", |
| 383 | "src/f32-igemm/gen/6x8-neon-lane-ld64.c", |
| 384 | "src/f32-igemm/gen/6x8-neon-lane-ld128.c", |
| 385 | "src/f32-igemm/gen/1x8-neon-dup-ld64.c", |
| 386 | "src/f32-igemm/gen/4x8-neon-dup-ld128.c", |
| 387 | "src/f32-igemm/gen/4x8-neon-dup-ld64.c", |
| 388 | "src/f32-igemm/gen/6x8-neon-dup-ld64.c", |
| 389 | "src/f32-igemm/gen/6x8-neon-dup-ld128.c", |
| 390 | "src/f32-igemm/gen/1x8s4-neon.c", |
| 391 | "src/f32-igemm/gen/4x8s4-neon.c", |
| 392 | "src/f32-igemm/gen/6x8s4-neon.c", |
| 393 | "src/f32-igemm/gen/8x8s4-neon.c", |
| 394 | "src/f32-pavgpool/mp9p8q-neon.c", |
| 395 | "src/f32-pavgpool/up9-neon.c", |
| 396 | "src/f32-ppmm/gen/4x8-neon.c", |
| 397 | "src/f32-ppmm/gen/8x8-neon.c", |
| 398 | "src/f32-prelu/gen/neon-2x4.c", |
| 399 | "src/f32-prelu/gen/neon-2x8.c", |
| 400 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 401 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 402 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 403 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 404 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 405 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 406 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 407 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 408 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 409 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| 410 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 411 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 412 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| 413 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| 414 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| 415 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| 416 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 417 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| 418 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| 419 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 420 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| 421 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 422 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 423 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
| 424 | "src/f32-rmax/neon.c", |
| 425 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
| 426 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 427 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 428 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 429 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 430 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 431 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| 432 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 433 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 434 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 435 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 436 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 437 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 438 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 439 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 440 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 441 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 442 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 443 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
| 444 | "src/f32-vbinary/gen/vadd-neon-x4.c", |
| 445 | "src/f32-vbinary/gen/vadd-neon-x8.c", |
| 446 | "src/f32-vbinary/gen/vaddc-neon-x4.c", |
| 447 | "src/f32-vbinary/gen/vaddc-neon-x8.c", |
| 448 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 449 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 450 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 451 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 452 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 453 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 454 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 455 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
| 456 | "src/f32-vbinary/gen/vmul-neon-x4.c", |
| 457 | "src/f32-vbinary/gen/vmul-neon-x8.c", |
| 458 | "src/f32-vbinary/gen/vmulc-neon-x4.c", |
| 459 | "src/f32-vbinary/gen/vmulc-neon-x8.c", |
| 460 | "src/f32-vbinary/gen/vrsubc-neon-x4.c", |
| 461 | "src/f32-vbinary/gen/vrsubc-neon-x8.c", |
| 462 | "src/f32-vbinary/gen/vsub-neon-x4.c", |
| 463 | "src/f32-vbinary/gen/vsub-neon-x8.c", |
| 464 | "src/f32-vbinary/gen/vsubc-neon-x4.c", |
| 465 | "src/f32-vbinary/gen/vsubc-neon-x8.c", |
| 466 | "src/f32-vmulcaddc/gen/c4-neon-2x.c", |
| 467 | "src/f32-vmulcaddc/gen/c8-neon-2x.c", |
| 468 | "src/q8-avgpool/mp9p8q-neon.c", |
| 469 | "src/q8-avgpool/up9-neon.c", |
| 470 | "src/q8-dwconv/up8x9-neon.c", |
| 471 | "src/q8-gavgpool/mp7p7q-neon.c", |
| 472 | "src/q8-gavgpool/up7-neon.c", |
| 473 | "src/q8-gemm/4x8-neon.c", |
| 474 | "src/q8-gemm/8x8-neon.c", |
| 475 | "src/q8-igemm/4x8-neon.c", |
| 476 | "src/q8-igemm/8x8-neon.c", |
| 477 | "src/q8-vadd/neon.c", |
| 478 | "src/u8-clamp/neon.c", |
| 479 | "src/u8-maxpool/9p8x-neon-c16.c", |
| 480 | "src/u8-rmax/neon.c", |
| 481 | "src/x32-packx/x4-neon-st4.c", |
| 482 | "src/x32-pad/x2-neon.c", |
| 483 | "src/x32-zip/x2-neon.c", |
| 484 | "src/x32-zip/x3-neon.c", |
| 485 | "src/x32-zip/x4-neon.c", |
| 486 | "src/x32-zip/xm-neon.c", |
| 487 | "src/x8-zip/x2-neon.c", |
| 488 | "src/x8-zip/x3-neon.c", |
| 489 | "src/x8-zip/x4-neon.c", |
| 490 | "src/x8-zip/xm-neon.c", |
| 491 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
| 492 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 493 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 494 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 495 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 496 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 497 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
| 498 | ] |
| 499 | |
| 500 | NEONFMA_UKERNELS = [ |
| 501 | "src/f32-bilinear/gen/neonfma-c4.c", |
| 502 | "src/f32-bilinear/gen/neonfma-c8.c", |
| 503 | "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c", |
| 504 | "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c", |
| 505 | "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c", |
| 506 | "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c", |
| 507 | "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c", |
| 508 | "src/f32-igemm/gen/1x8s4-neonfma.c", |
| 509 | "src/f32-igemm/gen/4x8s4-neonfma.c", |
| 510 | "src/f32-igemm/gen/6x8s4-neonfma.c", |
| 511 | "src/f32-igemm/gen/8x8s4-neonfma.c", |
| 512 | "src/f32-dwconv/gen/up4x9-neonfma.c", |
| 513 | "src/f32-dwconv/gen/up4x9-neonfma-acc2.c", |
| 514 | "src/f32-dwconv/gen/up8x9-neonfma.c", |
| 515 | "src/f32-dwconv/gen/up8x9-neonfma-acc2.c", |
| 516 | "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c", |
| 517 | "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c", |
| 518 | "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c", |
| 519 | "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c", |
| 520 | "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c", |
| 521 | "src/f32-gemm/gen/1x8s4-neonfma.c", |
| 522 | "src/f32-gemm/gen/4x8s4-neonfma.c", |
| 523 | "src/f32-gemm/gen/6x8s4-neonfma.c", |
| 524 | "src/f32-gemm/gen/8x8s4-neonfma.c", |
| 525 | "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c", |
| 526 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c", |
| 527 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c", |
| 528 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c", |
| 529 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c", |
| 530 | "src/f32-gemm/gen-inc/1x8s4-neonfma.c", |
| 531 | "src/f32-gemm/gen-inc/4x8s4-neonfma.c", |
| 532 | "src/f32-gemm/gen-inc/6x8s4-neonfma.c", |
| 533 | "src/f32-gemm/gen-inc/8x8s4-neonfma.c", |
| 534 | "src/f32-hswish/gen/neonfma-x4.c", |
| 535 | "src/f32-hswish/gen/neonfma-x8.c", |
| 536 | "src/f32-ppmm/gen/4x8-neonfma.c", |
| 537 | "src/f32-ppmm/gen/8x8-neonfma.c", |
| 538 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 539 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 540 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 541 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 542 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 543 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 544 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 545 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 546 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 547 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| 548 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 549 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 550 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| 551 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| 552 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| 553 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| 554 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 555 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| 556 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| 557 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 558 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| 559 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 560 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 561 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
| 562 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 563 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 564 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 565 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 566 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 567 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 568 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 569 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 570 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 571 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 572 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 573 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 574 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 575 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 576 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 577 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 578 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 579 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| 580 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 581 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 582 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 583 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 584 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 585 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| 586 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 587 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 588 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 589 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 590 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 591 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| 592 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 593 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 594 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 595 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 596 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 597 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| 598 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 599 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 600 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 601 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 602 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 603 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| 604 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 605 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 606 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 607 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 608 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 609 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| 610 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 611 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 612 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 613 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 614 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 615 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
| 616 | "src/f32-vmulcaddc/gen/c4-neonfma-2x.c", |
| 617 | "src/f32-vmulcaddc/gen/c8-neonfma-2x.c", |
| 618 | "src/math/exp-neonfma-lut64-p2.c", |
| 619 | "src/math/exp-neonfma-p5.c", |
| 620 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 621 | "src/math/expminus-neonfma-lut64-p2.c", |
| 622 | "src/math/expminus-neonfma-p5.c", |
| 623 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 624 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 625 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| 626 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 627 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 628 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| 629 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 630 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 631 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| 632 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 633 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 634 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| 635 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 636 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 637 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| 638 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 639 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 640 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
| 641 | ] |
| 642 | |
| 643 | AARCH64_NEONFMA_UKERNELS = [ |
| 644 | "src/f32-vbinary/gen/vdiv-neon-x4.c", |
| 645 | "src/f32-vbinary/gen/vdiv-neon-x8.c", |
| 646 | "src/f32-vbinary/gen/vdivc-neon-x4.c", |
| 647 | "src/f32-vbinary/gen/vdivc-neon-x8.c", |
| 648 | "src/f32-vbinary/gen/vrdivc-neon-x4.c", |
| 649 | "src/f32-vbinary/gen/vrdivc-neon-x8.c", |
| 650 | "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c", |
| 651 | "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c", |
| 652 | "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c", |
| 653 | "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c", |
| 654 | "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c", |
| 655 | "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c", |
| 656 | "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c", |
| 657 | "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c", |
| 658 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c", |
| 659 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c", |
| 660 | "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c", |
| 661 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c", |
| 662 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c", |
| 663 | "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c", |
| 664 | "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c", |
| 665 | "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c", |
| 666 | "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c", |
| 667 | "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c", |
| 668 | "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c", |
| 669 | "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c", |
| 670 | "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c", |
| 671 | "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c", |
| 672 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c", |
| 673 | "src/f32-dwconv-spchw/3x3p1-neonfma.c", |
| 674 | "src/f32-dwconv-spchw/5x5p2-neonfma.c", |
| 675 | "src/f32-dwconv-spchw/3x3s2p1-neonfma.c", |
| 676 | "src/f32-dwconv-spchw/5x5s2p2-neonfma.c", |
| 677 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 678 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 679 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 680 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 681 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 682 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| 683 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 684 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 685 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 686 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 687 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 688 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 689 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 690 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 691 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 692 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 693 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 694 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
| 695 | "src/f32-spmm/gen/12x1-neonfma.c", |
| 696 | "src/f32-spmm/gen/12x2-neonfma.c", |
| 697 | "src/f32-spmm/gen/12x4-neonfma.c", |
| 698 | "src/f32-spmm/gen/16x1-neonfma-pipelined.c", |
| 699 | "src/f32-spmm/gen/16x1-neonfma-unroll2.c", |
| 700 | "src/f32-spmm/gen/16x1-neonfma.c", |
| 701 | "src/f32-spmm/gen/16x2-neonfma.c", |
| 702 | "src/f32-spmm/gen/16x4-neonfma.c", |
| 703 | "src/f32-spmm/gen/4x1-neonfma-pipelined.c", |
| 704 | "src/f32-spmm/gen/4x1-neonfma-unroll2.c", |
| 705 | "src/f32-spmm/gen/4x1-neonfma.c", |
| 706 | "src/f32-spmm/gen/4x2-neonfma.c", |
| 707 | "src/f32-spmm/gen/4x4-neonfma.c", |
| 708 | "src/f32-spmm/gen/8x1-neonfma-pipelined.c", |
| 709 | "src/f32-spmm/gen/8x1-neonfma-unroll2.c", |
| 710 | "src/f32-spmm/gen/8x1-neonfma.c", |
| 711 | "src/f32-spmm/gen/8x2-neonfma.c", |
| 712 | "src/f32-spmm/gen/8x4-neonfma.c", |
| 713 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| 714 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| 715 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| 716 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| 717 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| 718 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
| 719 | ] |
| 720 | |
| 721 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
| 722 | "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c", |
| 723 | "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c", |
| 724 | "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c", |
| 725 | ] |
| 726 | |
| 727 | SSE_UKERNELS = [ |
| 728 | "src/f32-avgpool/mp9p8q-sse.c", |
| 729 | "src/f32-avgpool/up9-sse.c", |
| 730 | "src/f32-bilinear/gen/sse-c4.c", |
| 731 | "src/f32-bilinear/gen/sse-c8.c", |
| 732 | "src/f32-clamp/sse.c", |
| 733 | "src/f32-dwconv-spchw/3x3p1-sse.c", |
| 734 | "src/f32-dwconv-spchw/3x3s2p1-sse.c", |
| 735 | "src/f32-dwconv/gen/up4x25-sse-acc2.c", |
| 736 | "src/f32-dwconv/gen/up4x25-sse.c", |
| 737 | "src/f32-dwconv/gen/up4x4-sse-acc2.c", |
| 738 | "src/f32-dwconv/gen/up4x4-sse.c", |
| 739 | "src/f32-dwconv/gen/up4x9-sse-acc2.c", |
| 740 | "src/f32-dwconv/gen/up4x9-sse.c", |
| 741 | "src/f32-dwconv/gen/up8x25-sse-acc2.c", |
| 742 | "src/f32-dwconv/gen/up8x25-sse.c", |
| 743 | "src/f32-dwconv/gen/up8x4-sse-acc2.c", |
| 744 | "src/f32-dwconv/gen/up8x4-sse.c", |
| 745 | "src/f32-dwconv/gen/up8x9-sse-acc2.c", |
| 746 | "src/f32-dwconv/gen/up8x9-sse.c", |
| 747 | "src/f32-gavgpool-spchw/sse-x4.c", |
| 748 | "src/f32-gavgpool/mp7p7q-sse.c", |
| 749 | "src/f32-gavgpool/up7-sse.c", |
| 750 | "src/f32-gemm/gen/1x8-sse-dup.c", |
| 751 | "src/f32-gemm/gen/1x8-sse-load1.c", |
| 752 | "src/f32-gemm/gen/1x8s4-sse.c", |
| 753 | "src/f32-gemm/gen/4x8-sse-dup.c", |
| 754 | "src/f32-gemm/gen/4x8-sse-load1.c", |
| 755 | "src/f32-gemm/gen/4x8s4-sse.c", |
| 756 | "src/f32-gemm/gen-inc/1x8-sse-dup.c", |
| 757 | "src/f32-gemm/gen-inc/1x8-sse-load1.c", |
| 758 | "src/f32-gemm/gen-inc/1x8s4-sse.c", |
| 759 | "src/f32-gemm/gen-inc/4x8-sse-dup.c", |
| 760 | "src/f32-gemm/gen-inc/4x8-sse-load1.c", |
| 761 | "src/f32-gemm/gen-inc/4x8s4-sse.c", |
| 762 | "src/f32-hswish/gen/sse-x4.c", |
| 763 | "src/f32-hswish/gen/sse-x8.c", |
| 764 | "src/f32-igemm/gen/1x8-sse-dup.c", |
| 765 | "src/f32-igemm/gen/1x8-sse-load1.c", |
| 766 | "src/f32-igemm/gen/1x8s4-sse.c", |
| 767 | "src/f32-igemm/gen/4x2c4-sse.c", |
| 768 | "src/f32-igemm/gen/4x8-sse-dup.c", |
| 769 | "src/f32-igemm/gen/4x8-sse-load1.c", |
| 770 | "src/f32-igemm/gen/4x8s4-sse.c", |
| 771 | "src/f32-maxpool/9p8x-sse-c4.c", |
| 772 | "src/f32-pavgpool/mp9p8q-sse.c", |
| 773 | "src/f32-pavgpool/up9-sse.c", |
| 774 | "src/f32-ppmm/gen/4x8-sse.c", |
| 775 | "src/f32-rmax/sse.c", |
| 776 | "src/f32-spmm/gen/4x1-sse.c", |
| 777 | "src/f32-spmm/gen/8x1-sse.c", |
| 778 | "src/f32-vbinary/gen/vadd-sse-x4.c", |
| 779 | "src/f32-vbinary/gen/vadd-sse-x8.c", |
| 780 | "src/f32-vbinary/gen/vaddc-sse-x4.c", |
| 781 | "src/f32-vbinary/gen/vaddc-sse-x8.c", |
| 782 | "src/f32-vbinary/gen/vdiv-sse-x4.c", |
| 783 | "src/f32-vbinary/gen/vdiv-sse-x8.c", |
| 784 | "src/f32-vbinary/gen/vdivc-sse-x4.c", |
| 785 | "src/f32-vbinary/gen/vdivc-sse-x8.c", |
| 786 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 787 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 788 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 789 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 790 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 791 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 792 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 793 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
| 794 | "src/f32-vbinary/gen/vmul-sse-x4.c", |
| 795 | "src/f32-vbinary/gen/vmul-sse-x8.c", |
| 796 | "src/f32-vbinary/gen/vmulc-sse-x4.c", |
| 797 | "src/f32-vbinary/gen/vmulc-sse-x8.c", |
| 798 | "src/f32-vbinary/gen/vrdivc-sse-x4.c", |
| 799 | "src/f32-vbinary/gen/vrdivc-sse-x8.c", |
| 800 | "src/f32-vbinary/gen/vrsubc-sse-x4.c", |
| 801 | "src/f32-vbinary/gen/vrsubc-sse-x8.c", |
| 802 | "src/f32-vbinary/gen/vsub-sse-x4.c", |
| 803 | "src/f32-vbinary/gen/vsub-sse-x8.c", |
| 804 | "src/f32-vbinary/gen/vsubc-sse-x4.c", |
| 805 | "src/f32-vbinary/gen/vsubc-sse-x8.c", |
| 806 | "src/f32-vmulcaddc/gen/c4-sse-2x.c", |
| 807 | "src/f32-vmulcaddc/gen/c8-sse-2x.c", |
| 808 | "src/x32-packx/x4-sse.c", |
| 809 | ] |
| 810 | |
| 811 | SSE2_UKERNELS = [ |
| 812 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 813 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 814 | "src/f32-argmaxpool/9x-sse2-c4.c", |
| 815 | "src/f32-prelu/gen/sse2-2x4.c", |
| 816 | "src/f32-prelu/gen/sse2-2x8.c", |
| 817 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| 818 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| 819 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| 820 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| 821 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 822 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| 823 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| 824 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 825 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| 826 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 827 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 828 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
| 829 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
| 830 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
| 831 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
| 832 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
| 833 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 834 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
| 835 | "src/q8-avgpool/mp9p8q-sse2.c", |
| 836 | "src/q8-avgpool/up9-sse2.c", |
| 837 | "src/q8-igemm/4x4c2-sse2.c", |
| 838 | "src/q8-dwconv/up8x9-sse2.c", |
| 839 | "src/q8-gavgpool/mp7p7q-sse2.c", |
| 840 | "src/q8-gavgpool/up7-sse2.c", |
| 841 | "src/q8-gemm/2x4c8-sse2.c", |
| 842 | "src/q8-gemm/4x4c2-sse2.c", |
| 843 | "src/q8-vadd/sse2.c", |
| 844 | "src/u8-clamp/sse2.c", |
| 845 | "src/u8-maxpool/9p8x-sse2-c16.c", |
| 846 | "src/u8-rmax/sse2.c", |
| 847 | "src/x32-pad/x2-sse2.c", |
| 848 | "src/x32-zip/x2-sse2.c", |
| 849 | "src/x32-zip/x3-sse2.c", |
| 850 | "src/x32-zip/x4-sse2.c", |
| 851 | "src/x32-zip/xm-sse2.c", |
| 852 | "src/x8-zip/x2-sse2.c", |
| 853 | "src/x8-zip/x3-sse2.c", |
| 854 | "src/x8-zip/x4-sse2.c", |
| 855 | "src/x8-zip/xm-sse2.c", |
| 856 | "src/math/exp-sse2-p5.c", |
| 857 | "src/math/expminus-sse2-p5.c", |
| 858 | "src/math/sigmoid-sse2-p5-div.c", |
| 859 | ] |
| 860 | |
| 861 | SSE41_UKERNELS = [ |
| 862 | "src/f32-prelu/gen/sse41-2x4.c", |
| 863 | "src/f32-prelu/gen/sse41-2x8.c", |
| 864 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
| 865 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
| 866 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
| 867 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
| 868 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 869 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
| 870 | ] |
| 871 | |
| 872 | AVX_UKERNELS = [ |
| 873 | "src/f32-clamp/avx.c", |
| 874 | "src/f32-dwconv/gen/up16x4-avx-acc2.c", |
| 875 | "src/f32-dwconv/gen/up16x4-avx.c", |
| 876 | "src/f32-dwconv/gen/up8x4-avx-acc2.c", |
| 877 | "src/f32-dwconv/gen/up8x4-avx.c", |
| 878 | "src/f32-dwconv/gen/up16x9-avx-acc2.c", |
| 879 | "src/f32-dwconv/gen/up16x9-avx.c", |
| 880 | "src/f32-dwconv/gen/up8x9-avx-acc2.c", |
| 881 | "src/f32-dwconv/gen/up8x9-avx.c", |
| 882 | "src/f32-dwconv/gen/up16x25-avx-acc2.c", |
| 883 | "src/f32-dwconv/gen/up16x25-avx.c", |
| 884 | "src/f32-dwconv/gen/up8x25-avx-acc2.c", |
| 885 | "src/f32-dwconv/gen/up8x25-avx.c", |
| 886 | "src/f32-gemm/gen/1x8-avx-broadcast.c", |
| 887 | "src/f32-gemm/gen/4x8-avx-broadcast.c", |
| 888 | "src/f32-gemm/gen/5x8-avx-broadcast.c", |
| 889 | "src/f32-gemm/gen/6x8-avx-broadcast.c", |
| 890 | "src/f32-gemm/gen/7x8-avx-broadcast.c", |
| 891 | "src/f32-gemm/gen/1x16-avx-broadcast.c", |
| 892 | "src/f32-gemm/gen/3x16-avx-broadcast.c", |
| 893 | "src/f32-gemm/gen/4x16-avx-broadcast.c", |
| 894 | "src/f32-gemm/gen/5x16-avx-broadcast.c", |
| 895 | "src/f32-gemm/gen-inc/1x8-avx-broadcast.c", |
| 896 | "src/f32-gemm/gen-inc/4x8-avx-broadcast.c", |
| 897 | "src/f32-gemm/gen-inc/5x8-avx-broadcast.c", |
| 898 | "src/f32-gemm/gen-inc/6x8-avx-broadcast.c", |
| 899 | "src/f32-gemm/gen-inc/7x8-avx-broadcast.c", |
| 900 | "src/f32-gemm/gen-inc/1x16-avx-broadcast.c", |
| 901 | "src/f32-gemm/gen-inc/3x16-avx-broadcast.c", |
| 902 | "src/f32-gemm/gen-inc/4x16-avx-broadcast.c", |
| 903 | "src/f32-gemm/gen-inc/5x16-avx-broadcast.c", |
| 904 | "src/f32-hswish/gen/avx-x8.c", |
| 905 | "src/f32-hswish/gen/avx-x16.c", |
| 906 | "src/f32-igemm/gen/1x8-avx-broadcast.c", |
| 907 | "src/f32-igemm/gen/4x8-avx-broadcast.c", |
| 908 | "src/f32-igemm/gen/5x8-avx-broadcast.c", |
| 909 | "src/f32-igemm/gen/6x8-avx-broadcast.c", |
| 910 | "src/f32-igemm/gen/7x8-avx-broadcast.c", |
| 911 | "src/f32-igemm/gen/1x16-avx-broadcast.c", |
| 912 | "src/f32-igemm/gen/3x16-avx-broadcast.c", |
| 913 | "src/f32-igemm/gen/4x16-avx-broadcast.c", |
| 914 | "src/f32-igemm/gen/5x16-avx-broadcast.c", |
| 915 | "src/f32-rmax/avx.c", |
| 916 | "src/f32-vbinary/gen/vadd-avx-x8.c", |
| 917 | "src/f32-vbinary/gen/vadd-avx-x16.c", |
| 918 | "src/f32-vbinary/gen/vaddc-avx-x8.c", |
| 919 | "src/f32-vbinary/gen/vaddc-avx-x16.c", |
| 920 | "src/f32-vbinary/gen/vdiv-avx-x8.c", |
| 921 | "src/f32-vbinary/gen/vdiv-avx-x16.c", |
| 922 | "src/f32-vbinary/gen/vdivc-avx-x8.c", |
| 923 | "src/f32-vbinary/gen/vdivc-avx-x16.c", |
| 924 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 925 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 926 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 927 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 928 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 929 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 930 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 931 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
| 932 | "src/f32-vbinary/gen/vmul-avx-x8.c", |
| 933 | "src/f32-vbinary/gen/vmul-avx-x16.c", |
| 934 | "src/f32-vbinary/gen/vmulc-avx-x8.c", |
| 935 | "src/f32-vbinary/gen/vmulc-avx-x16.c", |
| 936 | "src/f32-vbinary/gen/vrdivc-avx-x8.c", |
| 937 | "src/f32-vbinary/gen/vrdivc-avx-x16.c", |
| 938 | "src/f32-vbinary/gen/vrsubc-avx-x8.c", |
| 939 | "src/f32-vbinary/gen/vrsubc-avx-x16.c", |
| 940 | "src/f32-vbinary/gen/vsub-avx-x8.c", |
| 941 | "src/f32-vbinary/gen/vsub-avx-x16.c", |
| 942 | "src/f32-vbinary/gen/vsubc-avx-x8.c", |
| 943 | "src/f32-vbinary/gen/vsubc-avx-x16.c", |
| 944 | "src/f32-vscale/avx-unroll32.c", |
| 945 | ] |
| 946 | |
| 947 | FMA3_UKERNELS = [ |
| 948 | "src/f32-dwconv/gen/up16x4-fma3-acc2.c", |
| 949 | "src/f32-dwconv/gen/up16x4-fma3.c", |
| 950 | "src/f32-dwconv/gen/up8x4-fma3-acc2.c", |
| 951 | "src/f32-dwconv/gen/up8x4-fma3.c", |
| 952 | "src/f32-dwconv/gen/up16x9-fma3-acc2.c", |
| 953 | "src/f32-dwconv/gen/up16x9-fma3.c", |
| 954 | "src/f32-dwconv/gen/up8x9-fma3-acc2.c", |
| 955 | "src/f32-dwconv/gen/up8x9-fma3.c", |
| 956 | "src/f32-dwconv/gen/up16x25-fma3-acc2.c", |
| 957 | "src/f32-dwconv/gen/up16x25-fma3.c", |
| 958 | "src/f32-dwconv/gen/up8x25-fma3-acc2.c", |
| 959 | "src/f32-dwconv/gen/up8x25-fma3.c", |
| 960 | "src/f32-gemm/gen/1x8-fma3-broadcast.c", |
| 961 | "src/f32-gemm/gen/4x8-fma3-broadcast.c", |
| 962 | "src/f32-gemm/gen/5x8-fma3-broadcast.c", |
| 963 | "src/f32-gemm/gen/6x8-fma3-broadcast.c", |
| 964 | "src/f32-gemm/gen/7x8-fma3-broadcast.c", |
| 965 | "src/f32-gemm/gen/8x8-fma3-broadcast.c", |
| 966 | "src/f32-gemm/gen/1x16-fma3-broadcast.c", |
| 967 | "src/f32-gemm/gen/3x16-fma3-broadcast.c", |
| 968 | "src/f32-gemm/gen/4x16-fma3-broadcast.c", |
| 969 | "src/f32-gemm/gen/5x16-fma3-broadcast.c", |
| 970 | "src/f32-gemm/gen/1x16s4-fma3-broadcast.c", |
| 971 | "src/f32-gemm/gen/3x16s4-fma3-broadcast.c", |
| 972 | "src/f32-gemm/gen/4x16s4-fma3-broadcast.c", |
| 973 | "src/f32-gemm/gen/5x16s4-fma3-broadcast.c", |
| 974 | "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c", |
| 975 | "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c", |
| 976 | "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c", |
| 977 | "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c", |
| 978 | "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c", |
| 979 | "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c", |
| 980 | "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c", |
| 981 | "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c", |
| 982 | "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c", |
| 983 | "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c", |
| 984 | "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c", |
| 985 | "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c", |
| 986 | "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c", |
| 987 | "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c", |
| 988 | "src/f32-hswish/gen/fma3-x8.c", |
| 989 | "src/f32-hswish/gen/fma3-x16.c", |
| 990 | "src/f32-igemm/gen/1x8-fma3-broadcast.c", |
| 991 | "src/f32-igemm/gen/4x8-fma3-broadcast.c", |
| 992 | "src/f32-igemm/gen/5x8-fma3-broadcast.c", |
| 993 | "src/f32-igemm/gen/6x8-fma3-broadcast.c", |
| 994 | "src/f32-igemm/gen/7x8-fma3-broadcast.c", |
| 995 | "src/f32-igemm/gen/8x8-fma3-broadcast.c", |
| 996 | "src/f32-igemm/gen/1x16-fma3-broadcast.c", |
| 997 | "src/f32-igemm/gen/3x16-fma3-broadcast.c", |
| 998 | "src/f32-igemm/gen/4x16-fma3-broadcast.c", |
| 999 | "src/f32-igemm/gen/5x16-fma3-broadcast.c", |
| 1000 | "src/f32-igemm/gen/1x16s4-fma3-broadcast.c", |
| 1001 | "src/f32-igemm/gen/3x16s4-fma3-broadcast.c", |
| 1002 | "src/f32-igemm/gen/4x16s4-fma3-broadcast.c", |
| 1003 | "src/f32-igemm/gen/5x16s4-fma3-broadcast.c", |
| 1004 | ] |
| 1005 | |
| 1006 | AVX2_UKERNELS = [ |
| 1007 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 1008 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1009 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1010 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 1011 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1012 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 1013 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1014 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1015 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 1016 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1017 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1018 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 1019 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 1020 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 1021 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 1022 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 1023 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 1024 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 1025 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 1026 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 1027 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 1028 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 1029 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 1030 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 1031 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 1032 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1033 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1034 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 1035 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1036 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 1037 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1038 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1039 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 1040 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1041 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1042 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 1043 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 1044 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 1045 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 1046 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 1047 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 1048 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 1049 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 1050 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 1051 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 1052 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 1053 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 1054 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 1055 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 1056 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 1057 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 1058 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 1059 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 1060 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 1061 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 1062 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 1063 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 1064 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 1065 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 1066 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 1067 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 1068 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 1069 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 1070 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 1071 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 1072 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
| 1073 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 1074 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 1075 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 1076 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 1077 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 1078 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 1079 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 1080 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 1081 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 1082 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 1083 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 1084 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 1085 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 1086 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 1087 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 1088 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 1089 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 1090 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 1091 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 1092 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 1093 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 1094 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 1095 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 1096 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
| 1097 | "src/math/exp-avx2-p5.c", |
| 1098 | "src/math/exp-avx2-perm-p3.c", |
| 1099 | "src/math/exp-avx2-perm-p4.c", |
| 1100 | "src/math/expminus-avx2-p5.c", |
| 1101 | "src/math/extexp-avx2-p5.c", |
| 1102 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 1103 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 1104 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| 1105 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 1106 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 1107 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
| 1108 | ] |
| 1109 | |
| 1110 | AVX512F_UKERNELS = [ |
| 1111 | "src/f32-clamp/avx512f.c", |
| 1112 | "src/f32-dwconv/gen/up32x4-avx512f-acc2.c", |
| 1113 | "src/f32-dwconv/gen/up32x4-avx512f.c", |
| 1114 | "src/f32-dwconv/gen/up16x4-avx512f-acc2.c", |
| 1115 | "src/f32-dwconv/gen/up16x4-avx512f.c", |
| 1116 | "src/f32-dwconv/gen/up32x9-avx512f-acc2.c", |
| 1117 | "src/f32-dwconv/gen/up32x9-avx512f.c", |
| 1118 | "src/f32-dwconv/gen/up16x9-avx512f-acc2.c", |
| 1119 | "src/f32-dwconv/gen/up16x9-avx512f.c", |
| 1120 | "src/f32-dwconv/gen/up32x25-avx512f-acc2.c", |
| 1121 | "src/f32-dwconv/gen/up32x25-avx512f.c", |
| 1122 | "src/f32-dwconv/gen/up16x25-avx512f-acc2.c", |
| 1123 | "src/f32-dwconv/gen/up16x25-avx512f.c", |
| 1124 | "src/f32-gemm/gen/1x16-avx512f-broadcast.c", |
| 1125 | "src/f32-gemm/gen/4x16-avx512f-broadcast.c", |
| 1126 | "src/f32-gemm/gen/5x16-avx512f-broadcast.c", |
| 1127 | "src/f32-gemm/gen/6x16-avx512f-broadcast.c", |
| 1128 | "src/f32-gemm/gen/7x16-avx512f-broadcast.c", |
| 1129 | "src/f32-gemm/gen/8x16-avx512f-broadcast.c", |
| 1130 | "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c", |
| 1131 | "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c", |
| 1132 | "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c", |
| 1133 | "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c", |
| 1134 | "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c", |
| 1135 | "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c", |
| 1136 | "src/f32-hswish/gen/avx512f-x16.c", |
| 1137 | "src/f32-hswish/gen/avx512f-x32.c", |
| 1138 | "src/f32-igemm/gen/1x16-avx512f-broadcast.c", |
| 1139 | "src/f32-igemm/gen/4x16-avx512f-broadcast.c", |
| 1140 | "src/f32-igemm/gen/5x16-avx512f-broadcast.c", |
| 1141 | "src/f32-igemm/gen/6x16-avx512f-broadcast.c", |
| 1142 | "src/f32-igemm/gen/7x16-avx512f-broadcast.c", |
| 1143 | "src/f32-igemm/gen/8x16-avx512f-broadcast.c", |
| 1144 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1145 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1146 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1147 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1148 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1149 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1150 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1151 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1152 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1153 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1154 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1155 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1156 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 1157 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1158 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1159 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 1160 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1161 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 1162 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1163 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1164 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 1165 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1166 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1167 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1168 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1169 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1170 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1171 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1172 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1173 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1174 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1175 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1176 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1177 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1178 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1179 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1180 | "src/f32-rmax/avx512f.c", |
| 1181 | "src/f32-vbinary/gen/vadd-avx512f-x16.c", |
| 1182 | "src/f32-vbinary/gen/vadd-avx512f-x32.c", |
| 1183 | "src/f32-vbinary/gen/vaddc-avx512f-x16.c", |
| 1184 | "src/f32-vbinary/gen/vaddc-avx512f-x32.c", |
| 1185 | "src/f32-vbinary/gen/vdiv-avx512f-x16.c", |
| 1186 | "src/f32-vbinary/gen/vdiv-avx512f-x32.c", |
| 1187 | "src/f32-vbinary/gen/vdivc-avx512f-x16.c", |
| 1188 | "src/f32-vbinary/gen/vdivc-avx512f-x32.c", |
| 1189 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 1190 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 1191 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 1192 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 1193 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 1194 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 1195 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 1196 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
| 1197 | "src/f32-vbinary/gen/vmul-avx512f-x16.c", |
| 1198 | "src/f32-vbinary/gen/vmul-avx512f-x32.c", |
| 1199 | "src/f32-vbinary/gen/vmulc-avx512f-x16.c", |
| 1200 | "src/f32-vbinary/gen/vmulc-avx512f-x32.c", |
| 1201 | "src/f32-vbinary/gen/vrdivc-avx512f-x16.c", |
| 1202 | "src/f32-vbinary/gen/vrdivc-avx512f-x32.c", |
| 1203 | "src/f32-vbinary/gen/vrsubc-avx512f-x16.c", |
| 1204 | "src/f32-vbinary/gen/vrsubc-avx512f-x32.c", |
| 1205 | "src/f32-vbinary/gen/vsub-avx512f-x16.c", |
| 1206 | "src/f32-vbinary/gen/vsub-avx512f-x32.c", |
| 1207 | "src/f32-vbinary/gen/vsubc-avx512f-x16.c", |
| 1208 | "src/f32-vbinary/gen/vsubc-avx512f-x32.c", |
| 1209 | "src/f32-vscale/avx512f-unroll64.c", |
| 1210 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 1211 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 1212 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 1213 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 1214 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 1215 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 1216 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 1217 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1218 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1219 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1220 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 1221 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1222 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 1223 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 1224 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 1225 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 1226 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 1227 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 1228 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 1229 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 1230 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 1231 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 1232 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 1233 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
| 1234 | "src/math/exp-avx512f-p5-scalef.c", |
| 1235 | "src/math/exp-avx512f-p5.c", |
| 1236 | "src/math/exp-avx512f-perm-p3.c", |
| 1237 | "src/math/exp-avx512f-perm2-p2.c", |
| 1238 | "src/math/extexp-avx512f-p5.c", |
| 1239 | ] |
| 1240 | |
| 1241 | AARCH32_ASM_UKERNELS = [ |
| 1242 | "src/q8-dwconv/up8x9-aarch32-neon.S", |
| 1243 | "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S", |
| 1244 | "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S", |
| 1245 | "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S", |
| 1246 | "src/f32-gemm/4x8-aarch32-neon-ld64.S", |
Miao Wang | 7fc14e3 | 2020-02-17 14:56:46 -0800 | [diff] [blame] | 1247 | "src/f32-igemm/4x8-aarch32-neon-ld64.S", |
| 1248 | "src/f32-igemm/gen/4x8-aarch32-neon-cortex-a75.S", |
| 1249 | "src/f32-igemm/gen/4x8-aarch32-neon-pld-cortex-a75.S", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1250 | ] |
| 1251 | |
| 1252 | AARCH64_ASM_UKERNELS = [ |
| 1253 | "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S", |
| 1254 | "src/f32-dwconv/up4x9-aarch64-neonfma.S", |
| 1255 | "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S", |
| 1256 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S", |
| 1257 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1258 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
| 1259 | "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S", |
| 1260 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S", |
| 1261 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1262 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1263 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S", |
| 1264 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S", |
| 1265 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S", |
| 1266 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
| 1267 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S", |
| 1268 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S", |
| 1269 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
| 1270 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
| 1271 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S", |
| 1272 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S", |
| 1273 | "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S", |
| 1274 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S", |
| 1275 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S", |
| 1276 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S", |
| 1277 | "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S", |
| 1278 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S", |
| 1279 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S", |
| 1280 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S", |
| 1281 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S", |
| 1282 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S", |
| 1283 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S", |
| 1284 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S", |
| 1285 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S", |
| 1286 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S", |
| 1287 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S", |
| 1288 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S", |
| 1289 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S", |
| 1290 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S", |
| 1291 | "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S", |
| 1292 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S", |
| 1293 | "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1294 | "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
| 1295 | "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S", |
| 1296 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S", |
| 1297 | "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1298 | "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1299 | "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S", |
| 1300 | "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
| 1301 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S", |
| 1302 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S", |
| 1303 | "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
| 1304 | "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
| 1305 | ] |
| 1306 | |
| 1307 | cc_defaults { |
| 1308 | name: "xnnpack_internal_default", |
| 1309 | vendor_available: true, |
| 1310 | sdk_version: "current", |
| 1311 | local_include_dirs: [ |
| 1312 | "include", |
| 1313 | "src", |
| 1314 | ], |
| 1315 | cflags: [ |
| 1316 | "-std=c99", |
| 1317 | "-DXNN_LOG_LEVEL=2", |
| 1318 | "-Wno-unused-parameter", |
| 1319 | "-Wno-missing-field-initializers", |
| 1320 | "-Wno-pointer-arith", |
| 1321 | ], |
| 1322 | stl: "libc++_static", |
| 1323 | } |
| 1324 | |
| 1325 | cc_library_static { |
| 1326 | name: "xnnpack_tables", |
| 1327 | defaults: ["xnnpack_internal_default"], |
| 1328 | srcs: TABLE_SRCS, |
| 1329 | } |
| 1330 | |
| 1331 | cc_library_static { |
| 1332 | name: "xnnpack_im2col", |
| 1333 | defaults: ["xnnpack_internal_default"], |
| 1334 | srcs: [ |
| 1335 | "src/im2col.c", |
| 1336 | ], |
| 1337 | } |
| 1338 | |
| 1339 | cc_library_static { |
| 1340 | name: "xnnpack_indirection", |
| 1341 | defaults: ["xnnpack_internal_default"], |
| 1342 | srcs: [ |
| 1343 | "src/indirection.c", |
| 1344 | ], |
| 1345 | header_libs: [ |
| 1346 | "fp16_headers", |
| 1347 | "fxdiv_headers", |
| 1348 | ], |
| 1349 | static_libs: [ |
| 1350 | "libpthreadpool", |
| 1351 | ], |
| 1352 | } |
| 1353 | |
| 1354 | cc_library_static { |
| 1355 | name: "xnnpack_operator_run", |
| 1356 | defaults: ["xnnpack_internal_default"], |
| 1357 | srcs: [ |
| 1358 | "src/operator-run.c", |
| 1359 | ], |
| 1360 | cflags: [ |
| 1361 | "-Wno-vla", |
| 1362 | ], |
| 1363 | header_libs: [ |
| 1364 | "fp16_headers", |
| 1365 | "fxdiv_headers", |
| 1366 | ], |
| 1367 | static_libs: [ |
| 1368 | "libclog", |
| 1369 | "libpthreadpool", |
| 1370 | ], |
| 1371 | } |
| 1372 | |
| 1373 | cc_library_static { |
| 1374 | name: "xnnpack_operators", |
| 1375 | defaults: ["xnnpack_internal_default"], |
| 1376 | srcs: OPERATOR_SRCS + [ |
| 1377 | "src/memory.c", |
| 1378 | "src/operator-delete.c", |
| 1379 | ], |
| 1380 | header_libs: [ |
| 1381 | "fp16_headers", |
| 1382 | "fxdiv_headers", |
| 1383 | ], |
| 1384 | static_libs: [ |
| 1385 | "libclog", |
| 1386 | "libpthreadpool", |
| 1387 | ], |
| 1388 | whole_static_libs: [ |
| 1389 | "xnnpack_indirection", |
| 1390 | ], |
| 1391 | } |
| 1392 | |
| 1393 | cc_library_static { |
| 1394 | name: "xnnpack_scalar_ukernels", |
| 1395 | defaults: ["xnnpack_internal_default"], |
| 1396 | srcs: SCALAR_UKERNELS, |
| 1397 | header_libs: [ |
| 1398 | "fp16_headers", |
| 1399 | "fxdiv_headers", |
| 1400 | ], |
| 1401 | static_libs: [ |
| 1402 | "libpthreadpool", |
| 1403 | "xnnpack_tables", |
| 1404 | ], |
| 1405 | } |
| 1406 | |
| 1407 | cc_library_static { |
| 1408 | name: "xnnpack_psimd_fastmath_ukernels", |
| 1409 | defaults: ["xnnpack_internal_default"], |
| 1410 | srcs: PSIMD_FASTMATH_UKERNELS, |
| 1411 | cflags: [ |
| 1412 | "-O3", |
| 1413 | "-ffast-math", |
| 1414 | ], |
| 1415 | arch: { |
| 1416 | arm: { |
| 1417 | cflags: [ |
| 1418 | "-marm", |
| 1419 | "-mfpu=neon", |
| 1420 | ], |
| 1421 | }, |
| 1422 | }, |
| 1423 | header_libs: [ |
| 1424 | "fp16_headers", |
| 1425 | "psimd_headers", |
| 1426 | ], |
| 1427 | static_libs: [ |
| 1428 | "libpthreadpool", |
| 1429 | "xnnpack_tables", |
| 1430 | ], |
| 1431 | } |
| 1432 | |
| 1433 | cc_library_static { |
| 1434 | name: "xnnpack_psimd_accmath_ukernels", |
| 1435 | defaults: ["xnnpack_internal_default"], |
| 1436 | srcs: PSIMD_ACCMATH_UKERNELS, |
| 1437 | cflags: [ |
| 1438 | "-O3", |
| 1439 | ], |
| 1440 | arch: { |
| 1441 | arm: { |
| 1442 | cflags: [ |
| 1443 | "-marm", |
| 1444 | "-mfpu=neon", |
| 1445 | ], |
| 1446 | }, |
| 1447 | }, |
| 1448 | header_libs: [ |
| 1449 | "fp16_headers", |
| 1450 | "psimd_headers", |
| 1451 | ], |
| 1452 | static_libs: [ |
| 1453 | "libpthreadpool", |
| 1454 | "xnnpack_tables", |
| 1455 | ], |
| 1456 | } |
| 1457 | |
| 1458 | cc_library_static { |
| 1459 | name: "xnnpack_neon_ukernels", |
| 1460 | defaults: ["xnnpack_internal_default"], |
| 1461 | arch: { |
| 1462 | arm: { |
| 1463 | srcs: NEON_UKERNELS, |
| 1464 | cflags: [ |
| 1465 | "-marm", |
| 1466 | "-mfpu=neon", |
| 1467 | ], |
| 1468 | }, |
| 1469 | arm64: { |
| 1470 | srcs: NEON_UKERNELS, |
| 1471 | }, |
| 1472 | x86: { enabled: false, }, |
| 1473 | x86_64: { enabled: false, }, |
| 1474 | }, |
| 1475 | header_libs: [ |
| 1476 | "fp16_headers", |
| 1477 | ], |
| 1478 | static_libs: [ |
| 1479 | "libpthreadpool", |
| 1480 | "xnnpack_tables", |
| 1481 | ], |
| 1482 | } |
| 1483 | |
| 1484 | cc_library_static { |
| 1485 | name: "xnnpack_neonfma_ukernels", |
| 1486 | defaults: ["xnnpack_internal_default"], |
| 1487 | arch: { |
| 1488 | arm: { |
| 1489 | srcs: NEONFMA_UKERNELS, |
| 1490 | cflags: [ |
| 1491 | "-marm", |
| 1492 | "-mfpu=neon-vfpv4", |
| 1493 | ], |
| 1494 | }, |
| 1495 | arm64: { |
| 1496 | srcs: NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 1497 | }, |
| 1498 | x86: { enabled: false, }, |
| 1499 | x86_64: { enabled: false, }, |
| 1500 | }, |
| 1501 | header_libs: [ |
| 1502 | "fp16_headers", |
| 1503 | ], |
| 1504 | static_libs: [ |
| 1505 | "libpthreadpool", |
| 1506 | "xnnpack_tables", |
| 1507 | ], |
| 1508 | } |
| 1509 | |
| 1510 | cc_library_static { |
| 1511 | name: "xnnpack_neonfp16arith_ukernels", |
| 1512 | defaults: ["xnnpack_internal_default"], |
| 1513 | arch: { |
| 1514 | arm: { enabled: false, }, |
| 1515 | arm64: { |
| 1516 | srcs: AARCH64_NEONFP16ARITH_UKERNELS, |
| 1517 | cflags: [ |
| 1518 | "-march=armv8.2-a+fp16", |
| 1519 | ], |
| 1520 | }, |
| 1521 | x86: { enabled: false, }, |
| 1522 | x86_64: { enabled: false, }, |
| 1523 | }, |
| 1524 | header_libs: [ |
| 1525 | "fp16_headers", |
| 1526 | ], |
| 1527 | static_libs: [ |
| 1528 | "libpthreadpool", |
| 1529 | "xnnpack_tables", |
| 1530 | ], |
| 1531 | } |
| 1532 | |
| 1533 | cc_library_static { |
| 1534 | name: "xnnpack_asm_ukernels", |
| 1535 | defaults: ["xnnpack_internal_default"], |
| 1536 | arch: { |
| 1537 | arm: { |
| 1538 | srcs: AARCH32_ASM_UKERNELS, |
| 1539 | }, |
| 1540 | arm64: { |
| 1541 | srcs: AARCH64_ASM_UKERNELS, |
| 1542 | }, |
| 1543 | x86: { enabled: false, }, |
| 1544 | x86_64: { enabled: false, }, |
| 1545 | }, |
| 1546 | } |
| 1547 | |
| 1548 | cc_library_static { |
| 1549 | name: "xnnpack_sse2_ukernels", |
| 1550 | defaults: ["xnnpack_internal_default"], |
| 1551 | arch: { |
| 1552 | arm: { enabled: false, }, |
| 1553 | arm64: { enabled: false, }, |
| 1554 | x86: { |
| 1555 | srcs: SSE_UKERNELS + SSE2_UKERNELS, |
| 1556 | cflags: [ |
| 1557 | "-msse2", |
| 1558 | ], |
| 1559 | }, |
| 1560 | x86_64: { |
| 1561 | srcs: SSE_UKERNELS + SSE2_UKERNELS, |
| 1562 | cflags: [ |
| 1563 | "-msse2", |
| 1564 | ], |
| 1565 | }, |
| 1566 | }, |
| 1567 | header_libs: [ |
| 1568 | "fp16_headers", |
| 1569 | ], |
| 1570 | static_libs: [ |
| 1571 | "libpthreadpool", |
| 1572 | "xnnpack_tables", |
| 1573 | ], |
| 1574 | } |
| 1575 | |
| 1576 | cc_library_static { |
| 1577 | name: "xnnpack_sse41_ukernels", |
| 1578 | defaults: ["xnnpack_internal_default"], |
| 1579 | arch: { |
| 1580 | arm: { enabled: false, }, |
| 1581 | arm64: { enabled: false, }, |
| 1582 | x86: { |
| 1583 | srcs: SSE41_UKERNELS, |
| 1584 | cflags: [ |
| 1585 | "-msse4.1", |
| 1586 | ], |
| 1587 | }, |
| 1588 | x86_64: { |
| 1589 | srcs: SSE41_UKERNELS, |
| 1590 | cflags: [ |
| 1591 | "-msse4.1", |
| 1592 | ], |
| 1593 | }, |
| 1594 | }, |
| 1595 | header_libs: [ |
| 1596 | "fp16_headers", |
| 1597 | ], |
| 1598 | static_libs: [ |
| 1599 | "libpthreadpool", |
| 1600 | "xnnpack_tables", |
| 1601 | ], |
| 1602 | } |
| 1603 | |
| 1604 | cc_library_static { |
| 1605 | name: "xnnpack_avx_ukernels", |
| 1606 | defaults: ["xnnpack_internal_default"], |
| 1607 | arch: { |
| 1608 | arm: { enabled: false, }, |
| 1609 | arm64: { enabled: false, }, |
| 1610 | x86: { |
| 1611 | srcs: AVX_UKERNELS, |
| 1612 | cflags: [ |
| 1613 | "-mavx", |
| 1614 | ], |
| 1615 | }, |
| 1616 | x86_64: { |
| 1617 | srcs: AVX_UKERNELS, |
| 1618 | cflags: [ |
| 1619 | "-mavx", |
| 1620 | ], |
| 1621 | }, |
| 1622 | }, |
| 1623 | header_libs: [ |
| 1624 | "fp16_headers", |
| 1625 | ], |
| 1626 | static_libs: [ |
| 1627 | "libpthreadpool", |
| 1628 | "xnnpack_tables", |
| 1629 | ], |
| 1630 | } |
| 1631 | |
| 1632 | cc_library_static { |
| 1633 | name: "xnnpack_fma3_ukernels", |
| 1634 | defaults: ["xnnpack_internal_default"], |
| 1635 | arch: { |
| 1636 | arm: { enabled: false, }, |
| 1637 | arm64: { enabled: false, }, |
| 1638 | x86: { |
| 1639 | srcs: FMA3_UKERNELS, |
| 1640 | cflags: [ |
| 1641 | "-mfma", |
| 1642 | ], |
| 1643 | }, |
| 1644 | x86_64: { |
| 1645 | srcs: FMA3_UKERNELS, |
| 1646 | cflags: [ |
| 1647 | "-mfma", |
| 1648 | ], |
| 1649 | }, |
| 1650 | }, |
| 1651 | header_libs: [ |
| 1652 | "fp16_headers", |
| 1653 | ], |
| 1654 | static_libs: [ |
| 1655 | "libpthreadpool", |
| 1656 | "xnnpack_tables", |
| 1657 | ], |
| 1658 | } |
| 1659 | |
| 1660 | cc_library_static { |
| 1661 | name: "xnnpack_avx2_ukernels", |
| 1662 | defaults: ["xnnpack_internal_default"], |
| 1663 | arch: { |
| 1664 | arm: { enabled: false, }, |
| 1665 | arm64: { enabled: false, }, |
| 1666 | x86: { |
| 1667 | srcs: AVX2_UKERNELS, |
| 1668 | cflags: [ |
| 1669 | "-mfma", |
| 1670 | "-mavx2", |
| 1671 | ], |
| 1672 | }, |
| 1673 | x86_64: { |
| 1674 | srcs: AVX2_UKERNELS, |
| 1675 | cflags: [ |
| 1676 | "-mfma", |
| 1677 | "-mavx2", |
| 1678 | ], |
| 1679 | }, |
| 1680 | }, |
| 1681 | header_libs: [ |
| 1682 | "fp16_headers", |
| 1683 | ], |
| 1684 | static_libs: [ |
| 1685 | "libpthreadpool", |
| 1686 | "xnnpack_tables", |
| 1687 | ], |
| 1688 | } |
| 1689 | |
| 1690 | cc_library_static { |
| 1691 | name: "xnnpack_avx512f_ukernels", |
| 1692 | defaults: ["xnnpack_internal_default"], |
| 1693 | arch: { |
| 1694 | arm: { enabled: false, }, |
| 1695 | arm64: { enabled: false, }, |
| 1696 | x86: { |
| 1697 | srcs: AVX512F_UKERNELS, |
| 1698 | cflags: [ |
| 1699 | "-mavx512f", |
| 1700 | ], |
| 1701 | }, |
| 1702 | x86_64: { |
| 1703 | srcs: AVX512F_UKERNELS, |
| 1704 | cflags: [ |
| 1705 | "-mavx512f", |
| 1706 | ], |
| 1707 | }, |
| 1708 | }, |
| 1709 | header_libs: [ |
| 1710 | "fp16_headers", |
| 1711 | ], |
| 1712 | static_libs: [ |
| 1713 | "libpthreadpool", |
| 1714 | "xnnpack_tables", |
| 1715 | ], |
| 1716 | } |
| 1717 | |
| 1718 | cc_library_static { |
| 1719 | name: "xnnpack_ukernels", |
| 1720 | defaults: ["xnnpack_internal_default"], |
| 1721 | arch: { |
| 1722 | arm: { |
| 1723 | whole_static_libs: [ |
| 1724 | "xnnpack_psimd_fastmath_ukernels", |
| 1725 | "xnnpack_psimd_accmath_ukernels", |
| 1726 | "xnnpack_neon_ukernels", |
| 1727 | "xnnpack_neonfma_ukernels", |
| 1728 | "xnnpack_asm_ukernels", |
| 1729 | ], |
| 1730 | }, |
| 1731 | arm64: { |
| 1732 | whole_static_libs: [ |
| 1733 | "xnnpack_psimd_fastmath_ukernels", |
| 1734 | "xnnpack_psimd_accmath_ukernels", |
| 1735 | "xnnpack_neon_ukernels", |
| 1736 | "xnnpack_neonfma_ukernels", |
| 1737 | "xnnpack_neonfp16arith_ukernels", |
| 1738 | "xnnpack_asm_ukernels", |
| 1739 | ], |
| 1740 | }, |
| 1741 | x86: { |
| 1742 | whole_static_libs: [ |
| 1743 | "xnnpack_psimd_fastmath_ukernels", |
| 1744 | "xnnpack_psimd_accmath_ukernels", |
| 1745 | "xnnpack_sse2_ukernels", |
| 1746 | "xnnpack_sse41_ukernels", |
| 1747 | "xnnpack_avx_ukernels", |
| 1748 | "xnnpack_fma3_ukernels", |
| 1749 | "xnnpack_avx2_ukernels", |
| 1750 | "xnnpack_avx512f_ukernels", |
| 1751 | ], |
| 1752 | }, |
| 1753 | x86_64: { |
| 1754 | whole_static_libs: [ |
| 1755 | "xnnpack_psimd_fastmath_ukernels", |
| 1756 | "xnnpack_psimd_accmath_ukernels", |
| 1757 | "xnnpack_sse2_ukernels", |
| 1758 | "xnnpack_sse41_ukernels", |
| 1759 | "xnnpack_avx_ukernels", |
| 1760 | "xnnpack_fma3_ukernels", |
| 1761 | "xnnpack_avx2_ukernels", |
| 1762 | "xnnpack_avx512f_ukernels", |
| 1763 | ], |
| 1764 | }, |
| 1765 | }, |
| 1766 | whole_static_libs: [ |
| 1767 | "xnnpack_scalar_ukernels", |
| 1768 | "xnnpack_tables", |
| 1769 | ], |
| 1770 | } |
| 1771 | |
| 1772 | cc_library_static { |
| 1773 | name: "libXNNPACK", |
| 1774 | defaults: ["xnnpack_internal_default"], |
| 1775 | export_include_dirs: ["include"], |
| 1776 | srcs: [ |
| 1777 | "src/init.c", |
| 1778 | "src/runtime.c", |
| 1779 | "src/subgraph.c", |
| 1780 | "src/tensor.c", |
| 1781 | ], |
| 1782 | whole_static_libs: [ |
| 1783 | "libclog", |
| 1784 | "libcpuinfo", |
| 1785 | "libpthreadpool", |
| 1786 | "xnnpack_ukernels", |
| 1787 | "xnnpack_operator_run", |
| 1788 | "xnnpack_operators", |
| 1789 | ], |
| 1790 | } |
| 1791 | |
| 1792 | // Tests and benchmarks |
| 1793 | cc_defaults { |
| 1794 | name: "xnnpack_tests_default", |
| 1795 | vendor_available: true, |
| 1796 | stl: "libc++_static", |
| 1797 | local_include_dirs: [ |
| 1798 | "bench", |
| 1799 | "models", |
| 1800 | "test", |
| 1801 | "src", |
| 1802 | ], |
| 1803 | cflags: [ |
| 1804 | "-Wno-unused-function" |
| 1805 | ], |
| 1806 | header_libs: [ |
| 1807 | "fp16_headers", |
| 1808 | ], |
| 1809 | static_libs: [ |
| 1810 | "libXNNPACK", |
| 1811 | "libpthreadpool", |
| 1812 | "libgmock", |
| 1813 | ], |
| 1814 | shared_libs: [ |
| 1815 | "liblog", |
| 1816 | ], |
| 1817 | } |
| 1818 | |
| 1819 | cc_library_static { |
| 1820 | name: "xnnpack_mobilenet_v1", |
| 1821 | defaults: ["xnnpack_tests_default"], |
| 1822 | srcs: [ |
| 1823 | "models/mobilenet-v1.cc", |
| 1824 | ], |
| 1825 | } |
| 1826 | |
| 1827 | cc_library_static { |
| 1828 | name: "xnnpack_mobilenet_v2", |
| 1829 | defaults: ["xnnpack_tests_default"], |
| 1830 | srcs: [ |
| 1831 | "models/mobilenet-v2.cc", |
| 1832 | ], |
| 1833 | } |
| 1834 | |
| 1835 | cc_library_static { |
| 1836 | name: "xnnpack_mobilenet_v3_large", |
| 1837 | defaults: ["xnnpack_tests_default"], |
| 1838 | srcs: [ |
| 1839 | "models/mobilenet-v3-large.cc", |
| 1840 | ], |
| 1841 | } |
| 1842 | |
| 1843 | cc_library_static { |
| 1844 | name: "xnnpack_mobilenet_v3_small", |
| 1845 | defaults: ["xnnpack_tests_default"], |
| 1846 | srcs: [ |
| 1847 | "models/mobilenet-v3-small.cc", |
| 1848 | ], |
| 1849 | } |
| 1850 | |
| 1851 | cc_benchmark { |
| 1852 | name: "xnnpack_end2end_bench", |
| 1853 | defaults: ["xnnpack_tests_default"], |
| 1854 | srcs: [ |
| 1855 | "bench/end2end.cc", |
| 1856 | "bench/utils.cc", |
| 1857 | ], |
| 1858 | cflags: [ |
| 1859 | "-Wno-unused-result" |
| 1860 | ], |
| 1861 | static_libs: [ |
| 1862 | "libcpuinfo", |
| 1863 | "libgoogle-benchmark", |
| 1864 | "xnnpack_mobilenet_v1", |
| 1865 | "xnnpack_mobilenet_v2", |
| 1866 | "xnnpack_mobilenet_v3_large", |
| 1867 | "xnnpack_mobilenet_v3_small", |
| 1868 | ], |
| 1869 | } |
| 1870 | |
| 1871 | cc_test { |
| 1872 | name: "xnnpack_add_nc_test", |
| 1873 | defaults: ["xnnpack_tests_default"], |
| 1874 | srcs: [ |
| 1875 | "test/add-nc.cc", |
| 1876 | ], |
| 1877 | test_suites: [ |
| 1878 | "general-tests", |
| 1879 | ], |
| 1880 | } |
| 1881 | |
| 1882 | cc_test { |
| 1883 | name: "xnnpack_add_nd_test", |
| 1884 | defaults: ["xnnpack_tests_default"], |
| 1885 | srcs: [ |
| 1886 | "test/add-nd.cc", |
| 1887 | ], |
| 1888 | test_suites: [ |
| 1889 | "general-tests", |
| 1890 | ], |
| 1891 | } |
| 1892 | |
| 1893 | cc_test { |
| 1894 | name: "xnnpack_argmax_pooling_nhwc_test", |
| 1895 | defaults: ["xnnpack_tests_default"], |
| 1896 | srcs: [ |
| 1897 | "test/argmax-pooling-nhwc.cc", |
| 1898 | ], |
| 1899 | test_suites: [ |
| 1900 | "general-tests", |
| 1901 | ], |
| 1902 | } |
| 1903 | |
| 1904 | cc_test { |
| 1905 | name: "xnnpack_average_pooling_nhwc_test", |
| 1906 | defaults: ["xnnpack_tests_default"], |
| 1907 | srcs: [ |
| 1908 | "test/average-pooling-nhwc.cc", |
| 1909 | ], |
| 1910 | test_suites: [ |
| 1911 | "general-tests", |
| 1912 | ], |
| 1913 | } |
| 1914 | |
| 1915 | cc_test { |
| 1916 | name: "xnnpack_channel_pad_nc_test", |
| 1917 | defaults: ["xnnpack_tests_default"], |
| 1918 | srcs: [ |
| 1919 | "test/channel-pad-nc.cc", |
| 1920 | ], |
| 1921 | test_suites: [ |
| 1922 | "general-tests", |
| 1923 | ], |
| 1924 | } |
| 1925 | |
| 1926 | cc_test { |
| 1927 | name: "xnnpack_channel_shuffle_nc_test", |
| 1928 | defaults: ["xnnpack_tests_default"], |
| 1929 | srcs: [ |
| 1930 | "test/channel-shuffle-nc.cc", |
| 1931 | ], |
| 1932 | test_suites: [ |
| 1933 | "general-tests", |
| 1934 | ], |
| 1935 | } |
| 1936 | |
| 1937 | cc_test { |
| 1938 | name: "xnnpack_clamp_nc_test", |
| 1939 | defaults: ["xnnpack_tests_default"], |
| 1940 | srcs: [ |
| 1941 | "test/clamp-nc.cc", |
| 1942 | ], |
| 1943 | test_suites: [ |
| 1944 | "general-tests", |
| 1945 | ], |
| 1946 | } |
| 1947 | |
| 1948 | cc_test { |
| 1949 | name: "xnnpack_convolution_nhwc_test", |
| 1950 | defaults: ["xnnpack_tests_default"], |
| 1951 | srcs: [ |
| 1952 | "test/convolution-nhwc.cc", |
| 1953 | ], |
| 1954 | test_suites: [ |
| 1955 | "general-tests", |
| 1956 | ], |
| 1957 | } |
| 1958 | |
| 1959 | cc_test { |
| 1960 | name: "xnnpack_convolution_nchw_test", |
| 1961 | defaults: ["xnnpack_tests_default"], |
| 1962 | srcs: [ |
| 1963 | "test/convolution-nchw.cc", |
| 1964 | ], |
| 1965 | test_suites: [ |
| 1966 | "general-tests", |
| 1967 | ], |
| 1968 | } |
| 1969 | |
| 1970 | cc_test { |
| 1971 | name: "xnnpack_deconvolution_nhwc_test", |
| 1972 | defaults: ["xnnpack_tests_default"], |
| 1973 | srcs: [ |
| 1974 | "test/deconvolution-nhwc.cc", |
| 1975 | ], |
| 1976 | test_suites: [ |
| 1977 | "general-tests", |
| 1978 | ], |
| 1979 | } |
| 1980 | |
| 1981 | cc_test { |
| 1982 | name: "xnnpack_divide_nd_test", |
| 1983 | defaults: ["xnnpack_tests_default"], |
| 1984 | srcs: [ |
| 1985 | "test/divide-nd.cc", |
| 1986 | ], |
| 1987 | test_suites: [ |
| 1988 | "general-tests", |
| 1989 | ], |
| 1990 | } |
| 1991 | |
| 1992 | cc_test { |
| 1993 | name: "xnnpack_fully_connected_nc_test", |
| 1994 | defaults: ["xnnpack_tests_default"], |
| 1995 | srcs: [ |
| 1996 | "test/fully-connected-nc.cc", |
| 1997 | ], |
| 1998 | test_suites: [ |
| 1999 | "general-tests", |
| 2000 | ], |
| 2001 | } |
| 2002 | |
| 2003 | cc_test { |
| 2004 | name: "xnnpack_global_average_pooling_nwc_test", |
| 2005 | defaults: ["xnnpack_tests_default"], |
| 2006 | srcs: [ |
| 2007 | "test/global-average-pooling-nwc.cc", |
| 2008 | ], |
| 2009 | test_suites: [ |
| 2010 | "general-tests", |
| 2011 | ], |
| 2012 | } |
| 2013 | |
| 2014 | cc_test { |
| 2015 | name: "xnnpack_global_average_pooling_ncw_test", |
| 2016 | defaults: ["xnnpack_tests_default"], |
| 2017 | srcs: [ |
| 2018 | "test/global-average-pooling-ncw.cc", |
| 2019 | ], |
| 2020 | test_suites: [ |
| 2021 | "general-tests", |
| 2022 | ], |
| 2023 | } |
| 2024 | |
| 2025 | cc_test { |
| 2026 | name: "xnnpack_hardswish_nc_test", |
| 2027 | defaults: ["xnnpack_tests_default"], |
| 2028 | srcs: [ |
| 2029 | "test/hardswish-nc.cc", |
| 2030 | ], |
| 2031 | test_suites: [ |
| 2032 | "general-tests", |
| 2033 | ], |
| 2034 | } |
| 2035 | |
| 2036 | cc_test { |
| 2037 | name: "xnnpack_leaky_relu_nc_test", |
| 2038 | defaults: ["xnnpack_tests_default"], |
| 2039 | srcs: [ |
| 2040 | "test/leaky-relu-nc.cc", |
| 2041 | ], |
| 2042 | test_suites: [ |
| 2043 | "general-tests", |
| 2044 | ], |
| 2045 | } |
| 2046 | |
| 2047 | cc_test { |
| 2048 | name: "xnnpack_max_pooling_nhwc_test", |
| 2049 | defaults: ["xnnpack_tests_default"], |
| 2050 | srcs: [ |
| 2051 | "test/max-pooling-nhwc.cc", |
| 2052 | ], |
| 2053 | test_suites: [ |
| 2054 | "general-tests", |
| 2055 | ], |
| 2056 | } |
| 2057 | |
| 2058 | cc_test { |
| 2059 | name: "xnnpack_maximum_nd_test", |
| 2060 | defaults: ["xnnpack_tests_default"], |
| 2061 | srcs: [ |
| 2062 | "test/maximum-nd.cc", |
| 2063 | ], |
| 2064 | test_suites: [ |
| 2065 | "general-tests", |
| 2066 | ], |
| 2067 | } |
| 2068 | |
| 2069 | cc_test { |
| 2070 | name: "xnnpack_minimum_nd_test", |
| 2071 | defaults: ["xnnpack_tests_default"], |
| 2072 | srcs: [ |
| 2073 | "test/minimum-nd.cc", |
| 2074 | ], |
| 2075 | test_suites: [ |
| 2076 | "general-tests", |
| 2077 | ], |
| 2078 | } |
| 2079 | |
| 2080 | cc_test { |
| 2081 | name: "xnnpack_multiply_nd_test", |
| 2082 | defaults: ["xnnpack_tests_default"], |
| 2083 | srcs: [ |
| 2084 | "test/multiply-nd.cc", |
| 2085 | ], |
| 2086 | test_suites: [ |
| 2087 | "general-tests", |
| 2088 | ], |
| 2089 | } |
| 2090 | |
| 2091 | cc_test { |
| 2092 | name: "xnnpack_prelu_nc_test", |
| 2093 | defaults: ["xnnpack_tests_default"], |
| 2094 | srcs: [ |
| 2095 | "test/prelu-nc.cc", |
| 2096 | ], |
| 2097 | test_suites: [ |
| 2098 | "general-tests", |
| 2099 | ], |
| 2100 | } |
| 2101 | |
| 2102 | cc_test { |
| 2103 | name: "xnnpack_resize_bilinear_nhwc_test", |
| 2104 | defaults: ["xnnpack_tests_default"], |
| 2105 | srcs: [ |
| 2106 | "test/resize-bilinear-nhwc.cc", |
| 2107 | ], |
| 2108 | test_suites: [ |
| 2109 | "general-tests", |
| 2110 | ], |
| 2111 | } |
| 2112 | |
| 2113 | cc_test { |
| 2114 | name: "xnnpack_sigmoid_nc_test", |
| 2115 | defaults: ["xnnpack_tests_default"], |
| 2116 | srcs: [ |
| 2117 | "test/sigmoid-nc.cc", |
| 2118 | ], |
| 2119 | test_suites: [ |
| 2120 | "general-tests", |
| 2121 | ], |
| 2122 | } |
| 2123 | |
| 2124 | cc_test { |
| 2125 | name: "xnnpack_softmax_nc_test", |
| 2126 | defaults: ["xnnpack_tests_default"], |
| 2127 | srcs: [ |
| 2128 | "test/softmax-nc.cc", |
| 2129 | ], |
| 2130 | test_suites: [ |
| 2131 | "general-tests", |
| 2132 | ], |
| 2133 | } |
| 2134 | |
| 2135 | cc_test { |
| 2136 | name: "xnnpack_subtract_nd_test", |
| 2137 | defaults: ["xnnpack_tests_default"], |
| 2138 | srcs: [ |
| 2139 | "test/subtract-nd.cc", |
| 2140 | ], |
| 2141 | test_suites: [ |
| 2142 | "general-tests", |
| 2143 | ], |
| 2144 | } |
| 2145 | |
| 2146 | cc_test { |
| 2147 | name: "xnnpack_unpooling_nhwc_test", |
| 2148 | defaults: ["xnnpack_tests_default"], |
| 2149 | srcs: [ |
| 2150 | "test/unpooling-nhwc.cc", |
| 2151 | ], |
| 2152 | test_suites: [ |
| 2153 | "general-tests", |
| 2154 | ], |
| 2155 | } |