Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1 | # Copyright 2019 Google LLC |
| 2 | # |
| 3 | # This source code is licensed under the BSD-style license found in the |
| 4 | # LICENSE file in the root directory of this source tree. |
| 5 | # |
| 6 | # Description: |
| 7 | # XNNPACK - optimized floating-point neural network operators library |
| 8 | |
| 9 | licenses(["notice"]) |
| 10 | |
| 11 | exports_files(["LICENSE"]) |
| 12 | |
| 13 | load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_unit_test", "xnnpack_visibility") |
| 14 | |
| 15 | OPERATOR_BENCHMARK_DEPS = [ |
| 16 | ":XNNPACK", |
| 17 | ":bench_utils", |
| 18 | "@cpuinfo", |
| 19 | "@pthreadpool", |
| 20 | ] |
| 21 | |
| 22 | MICROKERNEL_BENCHMARK_DEPS = [ |
| 23 | ":ukernels", |
| 24 | ":bench_utils", |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 25 | ":enable_assembly", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 26 | "@cpuinfo", |
| 27 | "@FP16", |
| 28 | "@pthreadpool", |
| 29 | ] |
| 30 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 31 | ACCURACY_EVAL_DEPS = [ |
| 32 | ":XNNPACK", |
| 33 | ":ukernels", |
| 34 | "@FP16", |
| 35 | "@pthreadpool", |
| 36 | ] |
| 37 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 38 | MICROKERNEL_TEST_DEPS = [ |
| 39 | ":ukernels", |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 40 | ":enable_assembly", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 41 | "@cpuinfo", |
| 42 | "@FP16", |
| 43 | "@pthreadpool", |
| 44 | ] |
| 45 | |
| 46 | OPERATOR_TEST_DEPS = [ |
| 47 | ":XNNPACK", |
| 48 | "@pthreadpool", |
| 49 | "@FP16", |
| 50 | ] |
| 51 | |
| 52 | OPERATOR_SRCS = [ |
| 53 | "src/add.c", |
| 54 | "src/argmax-pooling.c", |
| 55 | "src/average-pooling.c", |
| 56 | "src/channel-pad.c", |
| 57 | "src/channel-shuffle.c", |
| 58 | "src/clamp.c", |
| 59 | "src/convolution-spnchw.c", |
| 60 | "src/convolution.c", |
| 61 | "src/deconvolution.c", |
| 62 | "src/fully-connected.c", |
| 63 | "src/global-average-pooling-spnchw.c", |
| 64 | "src/global-average-pooling.c", |
| 65 | "src/hardswish.c", |
| 66 | "src/leaky-relu.c", |
| 67 | "src/max-pooling.c", |
| 68 | "src/prelu.c", |
| 69 | "src/sigmoid.c", |
| 70 | "src/softargmax.c", |
| 71 | "src/unpooling.c", |
| 72 | ] |
| 73 | |
| 74 | SCALAR_UKERNELS = [ |
| 75 | "src/f32-argmaxpool/mp9p8q-scalar.c", |
| 76 | "src/f32-argmaxpool/up4-scalar.c", |
| 77 | "src/f32-argmaxpool/up9-scalar.c", |
| 78 | "src/f32-avgpool/mp9p8q-scalar.c", |
| 79 | "src/f32-avgpool/up9-scalar.c", |
| 80 | "src/f32-clamp/scalar.c", |
| 81 | "src/f32-igemm/1x4-scalar.c", |
| 82 | "src/f32-igemm/2x4-scalar.c", |
| 83 | "src/f32-igemm/4x2-scalar.c", |
| 84 | "src/f32-igemm/4x4-scalar.c", |
| 85 | "src/f32-dwconv/up1x25-scalar.c", |
| 86 | "src/f32-dwconv/up1x4-scalar.c", |
| 87 | "src/f32-dwconv/up1x9-scalar.c", |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 88 | "src/f32-dwconv-spchw/3x3p1-scalar.c", |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame^] | 89 | "src/f32-dwconv-spchw/3x3s2p1-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 90 | "src/f32-gavgpool/mp7p7q-scalar.c", |
| 91 | "src/f32-gavgpool/up7-scalar.c", |
| 92 | "src/f32-gemm/1x4-scalar.c", |
| 93 | "src/f32-gemm/2x4-scalar.c", |
| 94 | "src/f32-gemm/4x2-scalar.c", |
| 95 | "src/f32-gemm/4x4-scalar.c", |
| 96 | "src/f32-gemminc/1x4-scalar.c", |
| 97 | "src/f32-gemminc/2x4-scalar.c", |
| 98 | "src/f32-gemminc/4x4-scalar.c", |
| 99 | "src/f32-hswish/scalar.c", |
| 100 | "src/f32-maxpool/9p8q-scalar.c", |
| 101 | "src/f32-pavgpool/mp9p8q-scalar.c", |
| 102 | "src/f32-pavgpool/up9-scalar.c", |
| 103 | "src/f32-ppmm/2x4-scalar.c", |
| 104 | "src/f32-ppmm/3x3-scalar.c", |
| 105 | "src/f32-ppmm/4x2-scalar.c", |
| 106 | "src/f32-ppmm/4x4-scalar.c", |
| 107 | "src/f32-prelu/x4-scalar.c", |
| 108 | "src/f32-rmax/scalar.c", |
| 109 | "src/f32-spmm/1x1-scalar-pipelined.c", |
| 110 | "src/f32-spmm/1x1-scalar-unroll2.c", |
| 111 | "src/f32-spmm/1x1-scalar.c", |
| 112 | "src/f32-spmm/2x1-scalar-pipelined.c", |
| 113 | "src/f32-spmm/2x1-scalar-unroll2.c", |
| 114 | "src/f32-spmm/2x1-scalar.c", |
| 115 | "src/f32-spmm/4x1-scalar-pipelined.c", |
| 116 | "src/f32-spmm/4x1-scalar-unroll2.c", |
| 117 | "src/f32-spmm/4x1-scalar.c", |
| 118 | "src/f32-spmm/8x1-scalar-pipelined.c", |
| 119 | "src/f32-spmm/8x1-scalar-unroll2.c", |
| 120 | "src/f32-spmm/8x1-scalar.c", |
| 121 | "src/f32-vadd/scalar.c", |
| 122 | "src/f32-vmul/scalar.c", |
| 123 | "src/f32-vmulcaddc/c1-scalar-x2.c", |
| 124 | "src/f32-vsub/scalar.c", |
| 125 | "src/q8-avgpool/mp9p8q-scalar.c", |
| 126 | "src/q8-avgpool/up9-scalar.c", |
| 127 | "src/q8-igemm/2x2-scalar.c", |
| 128 | "src/q8-dwconv/up1x9-scalar.c", |
| 129 | "src/q8-gavgpool/mp7p7q-scalar.c", |
| 130 | "src/q8-gavgpool/up7-scalar.c", |
| 131 | "src/q8-gemm/2x2-scalar.c", |
| 132 | "src/q8-vadd/scalar.c", |
| 133 | "src/u8-clamp/scalar.c", |
| 134 | "src/u8-lut32norm/scalar.c", |
| 135 | "src/u8-maxpool/9p8q-scalar.c", |
| 136 | "src/u8-rmax/scalar.c", |
| 137 | "src/x32-packx/x2-scalar.c", |
| 138 | "src/x32-packx/x3-scalar.c", |
| 139 | "src/x32-packx/x4-scalar.c", |
| 140 | "src/x32-pad/x2-scalar.c", |
| 141 | "src/x32-unpool/scalar.c", |
| 142 | "src/x32-zip/x2-scalar.c", |
| 143 | "src/x32-zip/x3-scalar.c", |
| 144 | "src/x32-zip/x4-scalar.c", |
| 145 | "src/x32-zip/xm-scalar.c", |
| 146 | "src/x8-lut/scalar.c", |
| 147 | "src/x8-zip/x2-scalar.c", |
| 148 | "src/x8-zip/x3-scalar.c", |
| 149 | "src/x8-zip/x4-scalar.c", |
| 150 | "src/x8-zip/xm-scalar.c", |
| 151 | ] |
| 152 | |
| 153 | PSIMD_UKERNELS = [ |
| 154 | "src/f32-argmaxpool/mp9p8q-psimd.c", |
| 155 | "src/f32-argmaxpool/up4-psimd.c", |
| 156 | "src/f32-argmaxpool/up9-psimd.c", |
| 157 | "src/f32-avgpool/mp9p8q-psimd.c", |
| 158 | "src/f32-avgpool/up9-psimd.c", |
| 159 | "src/f32-clamp/psimd.c", |
| 160 | "src/f32-igemm/1x8-psimd-loadsplat.c", |
| 161 | "src/f32-igemm/1x8-psimd-splat.c", |
| 162 | "src/f32-igemm/1x8s4-psimd.c", |
| 163 | "src/f32-igemm/4x2c4-psimd.c", |
| 164 | "src/f32-igemm/4x8-psimd-loadsplat.c", |
| 165 | "src/f32-igemm/4x8-psimd-splat.c", |
| 166 | "src/f32-igemm/4x8s4-psimd.c", |
| 167 | "src/f32-igemm/6x8-psimd-loadsplat.c", |
| 168 | "src/f32-igemm/6x8-psimd-splat.c", |
| 169 | "src/f32-igemm/6x8s4-psimd.c", |
| 170 | "src/f32-dwconv/up4x25-psimd.c", |
| 171 | "src/f32-dwconv/up4x4-psimd.c", |
| 172 | "src/f32-dwconv/up4x9-psimd.c", |
| 173 | "src/f32-gavgpool/mp7p7q-psimd.c", |
| 174 | "src/f32-gavgpool/up7-psimd.c", |
| 175 | "src/f32-gemm/1x8-psimd-loadsplat.c", |
| 176 | "src/f32-gemm/1x8-psimd-splat.c", |
| 177 | "src/f32-gemm/1x8s4-psimd.c", |
| 178 | "src/f32-gemm/4x8-psimd-loadsplat.c", |
| 179 | "src/f32-gemm/4x8-psimd-splat.c", |
| 180 | "src/f32-gemm/4x8s4-psimd.c", |
| 181 | "src/f32-gemm/6x8-psimd-loadsplat.c", |
| 182 | "src/f32-gemm/6x8-psimd-splat.c", |
| 183 | "src/f32-gemm/6x8s4-psimd.c", |
| 184 | "src/f32-gemminc/1x8-psimd-loadsplat.c", |
| 185 | "src/f32-gemminc/1x8-psimd-splat.c", |
| 186 | "src/f32-gemminc/1x8s4-psimd.c", |
| 187 | "src/f32-gemminc/4x8-psimd-loadsplat.c", |
| 188 | "src/f32-gemminc/4x8-psimd-splat.c", |
| 189 | "src/f32-gemminc/4x8s4-psimd.c", |
| 190 | "src/f32-gemminc/6x8-psimd-loadsplat.c", |
| 191 | "src/f32-gemminc/6x8-psimd-splat.c", |
| 192 | "src/f32-gemminc/6x8s4-psimd.c", |
| 193 | "src/f32-hswish/psimd.c", |
| 194 | "src/f32-maxpool/9p8q-psimd.c", |
| 195 | "src/f32-pavgpool/mp9p8q-psimd.c", |
| 196 | "src/f32-pavgpool/up9-psimd.c", |
| 197 | "src/f32-ppmm/4x8-psimd.c", |
| 198 | "src/f32-prelu/x4-psimd.c", |
| 199 | "src/f32-vadd/psimd.c", |
| 200 | "src/f32-vmul/psimd.c", |
| 201 | "src/f32-vmulcaddc/c4-psimd-x2.c", |
| 202 | "src/f32-vsub/psimd.c", |
| 203 | "src/x32-packx/x4-psimd.c", |
| 204 | "src/x32-pad/x2-psimd.c", |
| 205 | "src/x32-unpool/psimd.c", |
| 206 | "src/x32-zip/x2-psimd.c", |
| 207 | "src/x32-zip/x3-psimd.c", |
| 208 | "src/x32-zip/x4-psimd.c", |
| 209 | "src/x32-zip/xm-psimd.c", |
| 210 | ] |
| 211 | |
| 212 | # ISA-specific micro-kernels |
| 213 | NEON_UKERNELS = [ |
| 214 | "src/f32-avgpool/mp9p8q-neon.c", |
| 215 | "src/f32-avgpool/up9-neon.c", |
| 216 | "src/f32-clamp/neon.c", |
| 217 | "src/f32-igemm/1x8-neon-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 218 | "src/f32-igemm/4x2-neon-ld64.c", |
| 219 | "src/f32-igemm/4x4-neon-ld64.c", |
| 220 | "src/f32-igemm/4x8-neon-ld128.c", |
| 221 | "src/f32-igemm/4x8-neon-ld64.c", |
| 222 | "src/f32-igemm/6x8-neon-ld64.c", |
| 223 | "src/f32-dwconv/up4x9-neon.c", |
| 224 | "src/f32-gavgpool-spchw/neon-x4.c", |
| 225 | "src/f32-gavgpool/mp7p7q-neon.c", |
| 226 | "src/f32-gavgpool/up7-neon.c", |
| 227 | "src/f32-gemm/1x8-neon-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 228 | "src/f32-gemm/4x2-neon-ld64.c", |
| 229 | "src/f32-gemm/4x8-neon-ld128.c", |
| 230 | "src/f32-gemm/4x8-neon-ld64.c", |
| 231 | "src/f32-gemm/5x8-neon-ld64.c", |
| 232 | "src/f32-gemm/6x8-neon-ld64.c", |
| 233 | "src/f32-gemminc/1x8-neon-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 234 | "src/f32-gemminc/4x8-neon-ld128.c", |
| 235 | "src/f32-gemminc/4x8-neon-ld64.c", |
| 236 | "src/f32-gemminc/5x8-neon-ld64.c", |
| 237 | "src/f32-gemminc/6x8-neon-ld64.c", |
| 238 | "src/f32-hswish/neon.c", |
| 239 | "src/f32-pavgpool/mp9p8q-neon.c", |
| 240 | "src/f32-pavgpool/up9-neon.c", |
| 241 | "src/f32-ppmm/4x8-neon.c", |
| 242 | "src/f32-ppmm/8x8-neon.c", |
| 243 | "src/f32-rmax/neon.c", |
| 244 | "src/f32-vmulcaddc/c4-neon-x2.c", |
| 245 | "src/q8-avgpool/mp9p8q-neon.c", |
| 246 | "src/q8-avgpool/up9-neon.c", |
| 247 | "src/q8-igemm/4x8-neon.c", |
| 248 | "src/q8-igemm/8x8-neon.c", |
| 249 | "src/q8-dwconv/up8x9-neon.c", |
| 250 | "src/q8-gavgpool/mp7p7q-neon.c", |
| 251 | "src/q8-gavgpool/up7-neon.c", |
| 252 | "src/q8-gemm/4x8-neon.c", |
| 253 | "src/q8-gemm/8x8-neon.c", |
| 254 | "src/q8-vadd/neon.c", |
| 255 | "src/u8-clamp/neon.c", |
| 256 | "src/u8-maxpool/9p8q-neon.c", |
| 257 | "src/u8-rmax/neon.c", |
| 258 | "src/x32-packx/x4-neon-st4.c", |
| 259 | "src/x32-pad/x2-neon.c", |
| 260 | "src/x32-zip/x2-neon.c", |
| 261 | "src/x32-zip/x3-neon.c", |
| 262 | "src/x32-zip/x4-neon.c", |
| 263 | "src/x32-zip/xm-neon.c", |
| 264 | "src/x8-zip/x2-neon.c", |
| 265 | "src/x8-zip/x3-neon.c", |
| 266 | "src/x8-zip/x4-neon.c", |
| 267 | "src/x8-zip/xm-neon.c", |
| 268 | ] |
| 269 | |
| 270 | NEONFMA_UKERNELS = [ |
Frank Barchard | db45b6a | 2019-10-09 16:42:45 -0700 | [diff] [blame] | 271 | "src/f32-igemm/1x8-neonfma-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 272 | "src/f32-igemm/4x2-neonfma-ld64.c", |
| 273 | "src/f32-igemm/4x4-neonfma-ld64.c", |
| 274 | "src/f32-igemm/4x8-neonfma-ld128.c", |
| 275 | "src/f32-igemm/4x8-neonfma-ld64.c", |
| 276 | "src/f32-igemm/6x8-neonfma-ld64.c", |
| 277 | "src/f32-dwconv/up4x9-neonfma.c", |
| 278 | "src/f32-dwconv/up8x9-neonfma.c", |
| 279 | "src/f32-gemm/1x8-neonfma-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 280 | "src/f32-gemm/4x2-neonfma-ld64.c", |
| 281 | "src/f32-gemm/4x8-neonfma-ld128.c", |
| 282 | "src/f32-gemm/4x8-neonfma-ld64.c", |
| 283 | "src/f32-gemm/5x8-neonfma-ld64.c", |
| 284 | "src/f32-gemm/6x8-neonfma-ld64.c", |
| 285 | "src/f32-gemminc/1x8-neonfma-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 286 | "src/f32-gemminc/4x8-neonfma-ld128.c", |
| 287 | "src/f32-gemminc/4x8-neonfma-ld64.c", |
| 288 | "src/f32-gemminc/5x8-neonfma-ld64.c", |
| 289 | "src/f32-gemminc/6x8-neonfma-ld64.c", |
| 290 | "src/f32-hswish/neonfma.c", |
| 291 | "src/f32-ppmm/4x8-neonfma.c", |
| 292 | "src/f32-ppmm/8x8-neonfma.c", |
| 293 | "src/f32-vmulcaddc/c4-neonfma-x2.c", |
| 294 | ] |
| 295 | |
| 296 | AARCH64_NEONFMA_UKERNELS = [ |
| 297 | "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c", |
| 298 | "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c", |
| 299 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c", |
| 300 | "src/f32-dwconv-spchw/3x3p1-neonfma.c", |
| 301 | "src/f32-dwconv-spchw/5x5p2-neonfma.c", |
| 302 | "src/f32-dwconv-spchw/3x3s2p1-neonfma.c", |
| 303 | "src/f32-dwconv-spchw/5x5s2p2-neonfma.c", |
| 304 | "src/f32-spmm/12x1-neonfma.c", |
| 305 | "src/f32-spmm/12x2-neonfma.c", |
| 306 | "src/f32-spmm/12x4-neonfma.c", |
| 307 | "src/f32-spmm/16x1-neonfma-pipelined.c", |
| 308 | "src/f32-spmm/16x1-neonfma-unroll2.c", |
| 309 | "src/f32-spmm/16x1-neonfma.c", |
| 310 | "src/f32-spmm/16x2-neonfma.c", |
| 311 | "src/f32-spmm/16x4-neonfma.c", |
| 312 | "src/f32-spmm/4x1-neonfma-pipelined.c", |
| 313 | "src/f32-spmm/4x1-neonfma-unroll2.c", |
| 314 | "src/f32-spmm/4x1-neonfma.c", |
| 315 | "src/f32-spmm/4x2-neonfma.c", |
| 316 | "src/f32-spmm/4x4-neonfma.c", |
| 317 | "src/f32-spmm/8x1-neonfma-pipelined.c", |
| 318 | "src/f32-spmm/8x1-neonfma-unroll2.c", |
| 319 | "src/f32-spmm/8x1-neonfma.c", |
| 320 | "src/f32-spmm/8x2-neonfma.c", |
| 321 | "src/f32-spmm/8x4-neonfma.c", |
| 322 | ] |
| 323 | |
| 324 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
| 325 | "src/f16-gemm/4x8-neonfp16arith-ld64.c", |
| 326 | "src/f16-gemm/6x8-neonfp16arith-ld64.c", |
| 327 | "src/f16-gemm/8x8-neonfp16arith-ld64.c", |
| 328 | ] |
| 329 | |
| 330 | SSE_UKERNELS = [ |
| 331 | "src/f32-avgpool/mp9p8q-sse.c", |
| 332 | "src/f32-avgpool/up9-sse.c", |
| 333 | "src/f32-clamp/sse.c", |
| 334 | "src/f32-igemm/1x8-sse-dup.c", |
| 335 | "src/f32-igemm/1x8-sse-load1.c", |
| 336 | "src/f32-igemm/1x8s4-sse.c", |
| 337 | "src/f32-igemm/4x2c4-sse.c", |
| 338 | "src/f32-igemm/4x8-sse-dup.c", |
| 339 | "src/f32-igemm/4x8-sse-load1.c", |
| 340 | "src/f32-igemm/4x8s4-sse.c", |
| 341 | "src/f32-dwconv/up4x25-sse.c", |
| 342 | "src/f32-dwconv/up4x4-sse.c", |
| 343 | "src/f32-dwconv/up4x9-sse.c", |
| 344 | "src/f32-gavgpool-spchw/sse-x4.c", |
| 345 | "src/f32-gavgpool/mp7p7q-sse.c", |
| 346 | "src/f32-gavgpool/up7-sse.c", |
| 347 | "src/f32-gemm/1x8-sse-dup.c", |
| 348 | "src/f32-gemm/1x8-sse-load1.c", |
| 349 | "src/f32-gemm/1x8s4-sse.c", |
| 350 | "src/f32-gemm/4x8-sse-dup.c", |
| 351 | "src/f32-gemm/4x8-sse-load1.c", |
| 352 | "src/f32-gemm/4x8s4-sse.c", |
| 353 | "src/f32-gemminc/1x8-sse-dup.c", |
| 354 | "src/f32-gemminc/1x8-sse-load1.c", |
| 355 | "src/f32-gemminc/1x8s4-sse.c", |
| 356 | "src/f32-gemminc/4x8-sse-dup.c", |
| 357 | "src/f32-gemminc/4x8-sse-load1.c", |
| 358 | "src/f32-gemminc/4x8s4-sse.c", |
| 359 | "src/f32-hswish/sse.c", |
| 360 | "src/f32-maxpool/9p8q-sse.c", |
| 361 | "src/f32-pavgpool/mp9p8q-sse.c", |
| 362 | "src/f32-pavgpool/up9-sse.c", |
| 363 | "src/f32-dwconv-spchw/3x3p1-sse.c", |
| 364 | "src/f32-dwconv-spchw/3x3s2p1-sse.c", |
| 365 | "src/f32-ppmm/4x8-sse.c", |
| 366 | "src/f32-prelu/x4-sse.c", |
| 367 | "src/f32-rmax/sse.c", |
| 368 | "src/f32-spmm/4x1-sse.c", |
| 369 | "src/f32-spmm/8x1-sse.c", |
| 370 | "src/f32-vadd/sse.c", |
| 371 | "src/f32-vmul/sse.c", |
| 372 | "src/f32-vmulcaddc/c4-sse-x2.c", |
| 373 | "src/f32-vsub/sse.c", |
| 374 | "src/x32-packx/x4-sse.c", |
| 375 | ] |
| 376 | |
| 377 | SSE2_UKERNELS = [ |
| 378 | "src/f32-argmaxpool/mp9p8q-sse2.c", |
| 379 | "src/f32-argmaxpool/up4-sse2.c", |
| 380 | "src/f32-argmaxpool/up9-sse2.c", |
| 381 | "src/q8-avgpool/mp9p8q-sse2.c", |
| 382 | "src/q8-avgpool/up9-sse2.c", |
| 383 | "src/q8-igemm/4x4c2-sse2.c", |
| 384 | "src/q8-dwconv/up8x9-sse2.c", |
| 385 | "src/q8-gavgpool/mp7p7q-sse2.c", |
| 386 | "src/q8-gavgpool/up7-sse2.c", |
| 387 | "src/q8-gemm/2x4c8-sse2.c", |
| 388 | "src/q8-gemm/4x4c2-sse2.c", |
| 389 | "src/q8-vadd/sse2.c", |
| 390 | "src/u8-clamp/sse2.c", |
| 391 | "src/u8-maxpool/9p8q-sse2.c", |
| 392 | "src/u8-rmax/sse2.c", |
| 393 | "src/x32-pad/x2-sse2.c", |
| 394 | "src/x32-zip/x2-sse2.c", |
| 395 | "src/x32-zip/x3-sse2.c", |
| 396 | "src/x32-zip/x4-sse2.c", |
| 397 | "src/x32-zip/xm-sse2.c", |
| 398 | "src/x8-zip/x2-sse2.c", |
| 399 | "src/x8-zip/x3-sse2.c", |
| 400 | "src/x8-zip/x4-sse2.c", |
| 401 | "src/x8-zip/xm-sse2.c", |
| 402 | ] |
| 403 | |
| 404 | AVX_UKERNELS = [ |
| 405 | "src/f32-rmax/avx.c", |
| 406 | ] |
| 407 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 408 | AVX2_UKERNELS = [ |
| 409 | "src/math/exp-avx2-p5.c", |
| 410 | "src/math/exp-avx2-perm-p3.c", |
| 411 | "src/math/exp-avx2-perm-p4.c", |
| 412 | ] |
| 413 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 414 | AVX512F_UKERNELS = [ |
| 415 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 416 | "src/math/exp-avx512f-p5-scalef.c", |
| 417 | "src/math/exp-avx512f-p5.c", |
| 418 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 419 | ] |
| 420 | |
| 421 | AARCH32_ASM_UKERNELS = [ |
| 422 | "src/q8-dwconv/up8x9-aarch32-neon.S", |
| 423 | ] |
| 424 | |
| 425 | AARCH64_ASM_UKERNELS = [ |
| 426 | "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S", |
| 427 | "src/f32-dwconv/up4x9-aarch64-neonfma.S", |
| 428 | "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 429 | "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 430 | "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S", |
| 431 | "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S", |
| 432 | "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S", |
| 433 | "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S", |
| 434 | "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S", |
| 435 | "src/f32-gemm/4x8-aarch64-neonfma-ld128.S", |
| 436 | "src/f32-gemm/4x8-aarch64-neonfma-ld64.S", |
| 437 | "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S", |
| 438 | "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S", |
| 439 | "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S", |
| 440 | "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S", |
| 441 | "src/f32-gemm/6x8-aarch64-neonfma-ld128.S", |
| 442 | "src/f32-gemm/6x8-aarch64-neonfma-ld64.S", |
| 443 | "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 444 | "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 445 | "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S", |
| 446 | "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S", |
| 447 | "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S", |
| 448 | "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S", |
| 449 | "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S", |
| 450 | "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S", |
| 451 | "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S", |
| 452 | "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S", |
| 453 | "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S", |
| 454 | "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S", |
| 455 | "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S", |
| 456 | "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S", |
| 457 | "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S", |
| 458 | "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 459 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 460 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S", |
| 461 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S", |
| 462 | "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S", |
| 463 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S", |
| 464 | "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S", |
| 465 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S", |
| 466 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S", |
| 467 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S", |
| 468 | ] |
| 469 | |
| 470 | INTERNAL_MICROKERNEL_HDRS = [ |
| 471 | "src/xnnpack/argmaxpool.h", |
| 472 | "src/xnnpack/avgpool.h", |
| 473 | "src/xnnpack/clamp.h", |
| 474 | "src/xnnpack/common.h", |
| 475 | "src/xnnpack/conv.h", |
| 476 | "src/xnnpack/dwconv.h", |
| 477 | "src/xnnpack/gavgpool.h", |
| 478 | "src/xnnpack/gemm.h", |
| 479 | "src/xnnpack/hswish.h", |
| 480 | "src/xnnpack/igemm.h", |
| 481 | "src/xnnpack/lut.h", |
| 482 | "src/xnnpack/math.h", |
| 483 | "src/xnnpack/maxpool.h", |
| 484 | "src/xnnpack/packx.h", |
| 485 | "src/xnnpack/pad.h", |
| 486 | "src/xnnpack/params.h", |
| 487 | "src/xnnpack/pavgpool.h", |
| 488 | "src/xnnpack/ppmm.h", |
| 489 | "src/xnnpack/prelu.h", |
| 490 | "src/xnnpack/rmax.h", |
| 491 | "src/xnnpack/scalar-utils.h", |
| 492 | "src/xnnpack/spmm.h", |
| 493 | "src/xnnpack/unpool.h", |
| 494 | "src/xnnpack/vadd.h", |
| 495 | "src/xnnpack/vmul.h", |
| 496 | "src/xnnpack/vmulcaddc.h", |
| 497 | "src/xnnpack/vsub.h", |
| 498 | "src/xnnpack/zip.h", |
| 499 | ] |
| 500 | |
| 501 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 502 | "include/xnnpack.h", |
| 503 | "src/xnnpack/allocator.h", |
| 504 | "src/xnnpack/compute.h", |
| 505 | "src/xnnpack/im2col.h", |
| 506 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 507 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 508 | "src/xnnpack/operator.h", |
| 509 | "src/xnnpack/pack.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 510 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 511 | "src/xnnpack/requantization.h", |
| 512 | ] |
| 513 | |
| 514 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 515 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 516 | ] |
| 517 | |
| 518 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 519 | "src/xnnpack/requantization.h", |
| 520 | "include/xnnpack.h", |
| 521 | ] |
| 522 | |
| 523 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 524 | "src/xnnpack/isa-checks.h", |
| 525 | "src/xnnpack/requantization.h", |
| 526 | "include/xnnpack.h", |
| 527 | ] |
| 528 | |
| 529 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 530 | "src/xnnpack/params.h", |
| 531 | "src/xnnpack/common.h", |
| 532 | ] |
| 533 | |
| 534 | WEIGHTS_PACK_HDRS = [ |
| 535 | "src/xnnpack/pack.h", |
| 536 | "src/xnnpack/operator.h", |
| 537 | "src/xnnpack/compute.h", |
| 538 | ] |
| 539 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 540 | LOGGING_COPTS = select({ |
| 541 | # No logging in optimized mode |
| 542 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 543 | # Full logging in debug mode |
| 544 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 545 | # Error-only logging in default (fastbuild) mode |
| 546 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 547 | }) |
| 548 | |
| 549 | LOGGING_HDRS = [ |
| 550 | "src/xnnpack/log.h", |
| 551 | ] |
| 552 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 553 | xnnpack_cc_library( |
| 554 | name = "scalar_ukernels", |
| 555 | srcs = SCALAR_UKERNELS, |
| 556 | hdrs = INTERNAL_HDRS, |
| 557 | aarch32_copts = ["-marm"], |
| 558 | copts = xnnpack_std_copts(), |
| 559 | deps = [ |
| 560 | "@FP16", |
| 561 | "@FXdiv", |
| 562 | ], |
| 563 | ) |
| 564 | |
| 565 | xnnpack_cc_library( |
| 566 | name = "psimd_ukernels", |
| 567 | srcs = PSIMD_UKERNELS, |
| 568 | hdrs = INTERNAL_HDRS, |
| 569 | aarch32_copts = [ |
| 570 | "-marm", |
| 571 | "-mfpu=neon", |
| 572 | ], |
| 573 | copts = xnnpack_std_copts(), |
| 574 | optimized_copts = [ |
| 575 | "-O3", |
| 576 | "-ffast-math", |
| 577 | ], |
| 578 | deps = [ |
| 579 | "@FP16", |
| 580 | "@psimd", |
| 581 | ], |
| 582 | ) |
| 583 | |
| 584 | xnnpack_cc_library( |
| 585 | name = "neon_ukernels", |
| 586 | hdrs = INTERNAL_HDRS, |
| 587 | aarch32_copts = [ |
| 588 | "-marm", |
| 589 | "-mfpu=neon", |
| 590 | ], |
| 591 | aarch32_srcs = NEON_UKERNELS, |
| 592 | aarch64_srcs = NEON_UKERNELS, |
| 593 | copts = xnnpack_std_copts(), |
| 594 | deps = ["@FP16"], |
| 595 | ) |
| 596 | |
| 597 | xnnpack_cc_library( |
| 598 | name = "neonfma_ukernels", |
| 599 | hdrs = INTERNAL_HDRS, |
| 600 | aarch32_copts = [ |
| 601 | "-marm", |
| 602 | "-mfpu=neon-vfpv4", |
| 603 | ], |
| 604 | aarch32_srcs = NEONFMA_UKERNELS, |
| 605 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 606 | copts = xnnpack_std_copts(), |
| 607 | deps = ["@FP16"], |
| 608 | ) |
| 609 | |
| 610 | xnnpack_cc_library( |
| 611 | name = "neonfp16arith_ukernels", |
| 612 | hdrs = INTERNAL_HDRS, |
| 613 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 614 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 615 | copts = xnnpack_std_copts(), |
| 616 | deps = ["@FP16"], |
| 617 | ) |
| 618 | |
| 619 | xnnpack_cc_library( |
| 620 | name = "sse2_ukernels", |
| 621 | hdrs = INTERNAL_HDRS, |
| 622 | copts = xnnpack_std_copts(), |
| 623 | x86_copts = ["-msse2"], |
| 624 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| 625 | deps = ["@FP16"], |
| 626 | ) |
| 627 | |
| 628 | xnnpack_cc_library( |
| 629 | name = "avx_ukernels", |
| 630 | hdrs = INTERNAL_HDRS, |
| 631 | copts = xnnpack_std_copts(), |
| 632 | x86_copts = ["-mavx"], |
| 633 | x86_srcs = AVX_UKERNELS, |
| 634 | deps = ["@FP16"], |
| 635 | ) |
| 636 | |
| 637 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 638 | name = "avx2_ukernels", |
| 639 | hdrs = INTERNAL_HDRS, |
| 640 | copts = xnnpack_std_copts(), |
| 641 | x86_copts = [ |
| 642 | "-mfma", |
| 643 | "-mavx2", |
| 644 | ], |
| 645 | x86_srcs = AVX2_UKERNELS, |
| 646 | deps = ["@FP16"], |
| 647 | ) |
| 648 | |
| 649 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 650 | name = "avx512f_ukernels", |
| 651 | hdrs = INTERNAL_HDRS, |
| 652 | copts = xnnpack_std_copts(), |
| 653 | x86_copts = ["-mavx512f"], |
| 654 | x86_srcs = AVX512F_UKERNELS, |
| 655 | deps = ["@FP16"], |
| 656 | ) |
| 657 | |
| 658 | xnnpack_cc_library( |
| 659 | name = "asm_ukernels", |
| 660 | hdrs = ["src/xnnpack/assembly.h"], |
| 661 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
| 662 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 663 | ) |
| 664 | |
| 665 | xnnpack_aggregate_library( |
| 666 | name = "ukernels", |
| 667 | aarch32_deps = [ |
| 668 | ":psimd_ukernels", |
| 669 | ":neon_ukernels", |
| 670 | ":neonfma_ukernels", |
| 671 | ":asm_ukernels", |
| 672 | ], |
| 673 | aarch64_deps = [ |
| 674 | ":psimd_ukernels", |
| 675 | ":neon_ukernels", |
| 676 | ":neonfma_ukernels", |
| 677 | ":neonfp16arith_ukernels", |
| 678 | ":asm_ukernels", |
| 679 | ], |
| 680 | generic_deps = [":scalar_ukernels"], |
| 681 | wasmsimd_deps = [ |
| 682 | ":psimd_ukernels", |
| 683 | ], |
| 684 | x86_deps = [ |
| 685 | ":psimd_ukernels", |
| 686 | ":sse2_ukernels", |
| 687 | ":avx_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 688 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 689 | ":avx512f_ukernels", |
| 690 | ], |
| 691 | ) |
| 692 | |
| 693 | xnnpack_cc_library( |
| 694 | name = "im2col", |
| 695 | srcs = ["src/im2col.c"], |
| 696 | hdrs = [ |
| 697 | "src/xnnpack/common.h", |
| 698 | "src/xnnpack/im2col.h", |
| 699 | ], |
| 700 | copts = xnnpack_std_copts(), |
| 701 | ) |
| 702 | |
| 703 | xnnpack_cc_library( |
| 704 | name = "indirection", |
| 705 | srcs = ["src/indirection.c"], |
| 706 | hdrs = INTERNAL_HDRS, |
| 707 | copts = xnnpack_std_copts(), |
| 708 | deps = [ |
| 709 | "@FP16", |
| 710 | "@FXdiv", |
| 711 | "@pthreadpool", |
| 712 | ], |
| 713 | ) |
| 714 | |
| 715 | xnnpack_cc_library( |
| 716 | name = "operator_run", |
| 717 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 718 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 719 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 720 | # Wrappers for multi-pass microkernels use VLAs for temporary buffers. |
| 721 | "-Wno-vla", |
| 722 | ], |
| 723 | deps = [ |
| 724 | "@FP16", |
| 725 | "@FXdiv", |
| 726 | "@clog", |
| 727 | "@pthreadpool", |
| 728 | ], |
| 729 | ) |
| 730 | |
| 731 | cc_library( |
| 732 | name = "enable_assembly", |
| 733 | defines = select({ |
| 734 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 735 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 736 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 737 | }), |
| 738 | ) |
| 739 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 740 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 741 | name = "operators", |
| 742 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 743 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 744 | ], |
| 745 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 746 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 747 | "-Isrc", |
| 748 | "-Iinclude", |
| 749 | ] + select({ |
| 750 | ":debug_build": [], |
| 751 | "//conditions:default": xnnpack_min_size_copts(), |
| 752 | }), |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 753 | wasm_srcs = ["src/wasm-stubs.c"], |
| 754 | wasmsimd_srcs = ["src/wasm-stubs.c"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 755 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 756 | ":indirection", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 757 | "@FP16", |
| 758 | "@FXdiv", |
| 759 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 760 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 761 | ], |
| 762 | ) |
| 763 | |
| 764 | cc_library( |
| 765 | name = "XNNPACK", |
| 766 | srcs = [ |
| 767 | "src/init.c", |
| 768 | ], |
| 769 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 770 | "-Isrc", |
| 771 | "-Iinclude", |
| 772 | ] + select({ |
| 773 | ":debug_build": [], |
| 774 | "//conditions:default": xnnpack_min_size_copts(), |
| 775 | }), |
| 776 | includes = ["include"], |
| 777 | linkstatic = True, |
| 778 | textual_hdrs = ["include/xnnpack.h"], |
| 779 | visibility = xnnpack_visibility(), |
| 780 | deps = [ |
| 781 | ":enable_assembly", |
| 782 | ":ukernels", |
| 783 | ":operator_run", |
| 784 | ":operators", |
| 785 | "@clog", |
| 786 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 787 | ] + select({ |
| 788 | ":emscripten": [], |
| 789 | "//conditions:default": ["@cpuinfo"], |
| 790 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 791 | ) |
| 792 | |
| 793 | cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 794 | name = "xnnpack_operators_nhwc_f32", |
| 795 | srcs = [ |
| 796 | "src/init.c", |
| 797 | ], |
| 798 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 799 | "-Isrc", |
| 800 | "-Iinclude", |
| 801 | ] + select({ |
| 802 | ":debug_build": [], |
| 803 | "//conditions:default": xnnpack_min_size_copts(), |
| 804 | }), |
| 805 | defines = [ |
| 806 | "XNN_NO_Q8_OPERATORS", |
| 807 | "XNN_NO_U8_OPERATORS", |
| 808 | "XNN_NO_X8_OPERATORS", |
| 809 | "XNN_NO_SPNCHW_OPERATORS", |
| 810 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 811 | includes = ["include"], |
| 812 | linkstatic = True, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 813 | textual_hdrs = ["include/xnnpack.h"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 814 | visibility = xnnpack_visibility(), |
| 815 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 816 | ":enable_assembly", |
| 817 | ":ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 818 | ":operator_run", |
| 819 | ":operators", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 820 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 821 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 822 | ] + select({ |
| 823 | ":emscripten": [], |
| 824 | "//conditions:default": ["@cpuinfo"], |
| 825 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 826 | ) |
| 827 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 828 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 829 | name = "bench_utils", |
| 830 | srcs = ["bench/utils.cc"], |
| 831 | hdrs = ["bench/utils.h"], |
| 832 | copts = ["-Wno-unused-result"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 833 | deps = ["@cpuinfo"], |
| 834 | ) |
| 835 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 836 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 837 | |
| 838 | xnnpack_benchmark( |
| 839 | name = "q8_gemm_bench", |
| 840 | srcs = [ |
| 841 | "bench/gemm.h", |
| 842 | "bench/q8-gemm.cc", |
| 843 | "src/xnnpack/AlignedAllocator.h", |
| 844 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 845 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 846 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
| 847 | ) |
| 848 | |
| 849 | xnnpack_benchmark( |
| 850 | name = "f16_gemm_bench", |
| 851 | srcs = [ |
| 852 | "bench/f16-gemm.cc", |
| 853 | "bench/gemm.h", |
| 854 | "src/xnnpack/AlignedAllocator.h", |
| 855 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 856 | copts = ["-Wno-unused-function"], |
| 857 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 858 | ) |
| 859 | |
| 860 | xnnpack_benchmark( |
| 861 | name = "f32_igemm_bench", |
| 862 | srcs = [ |
| 863 | "bench/f32-igemm.cc", |
| 864 | "bench/conv.h", |
| 865 | "src/xnnpack/AlignedAllocator.h", |
| 866 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 867 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 868 | ) |
| 869 | |
| 870 | xnnpack_benchmark( |
| 871 | name = "f32_conv_hwc_bench", |
| 872 | srcs = [ |
| 873 | "bench/f32-conv-hwc.cc", |
| 874 | "bench/dconv.h", |
| 875 | "src/xnnpack/AlignedAllocator.h", |
| 876 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 877 | copts = ["-Wno-unused-function"], |
| 878 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 879 | ) |
| 880 | |
| 881 | xnnpack_benchmark( |
| 882 | name = "f32_dwconv_bench", |
| 883 | srcs = [ |
| 884 | "bench/f32-dwconv.cc", |
| 885 | "bench/dwconv.h", |
| 886 | "src/xnnpack/AlignedAllocator.h", |
| 887 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 888 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 889 | ) |
| 890 | |
| 891 | xnnpack_benchmark( |
| 892 | name = "f32_dwconv_spchw_bench", |
| 893 | srcs = [ |
| 894 | "bench/f32-dwconv-spchw.cc", |
| 895 | "bench/dwconv.h", |
| 896 | "src/xnnpack/AlignedAllocator.h", |
| 897 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 898 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 899 | ) |
| 900 | |
| 901 | xnnpack_benchmark( |
| 902 | name = "f32_gemm_bench", |
| 903 | srcs = [ |
| 904 | "bench/f32-gemm.cc", |
| 905 | "bench/gemm.h", |
| 906 | "src/xnnpack/AlignedAllocator.h", |
| 907 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 908 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(), |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 909 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 910 | ) |
| 911 | |
| 912 | xnnpack_benchmark( |
| 913 | name = "f32_rmax_bench", |
| 914 | srcs = [ |
| 915 | "bench/f32-rmax.cc", |
| 916 | "src/xnnpack/AlignedAllocator.h", |
| 917 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 918 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 919 | ) |
| 920 | |
| 921 | xnnpack_benchmark( |
| 922 | name = "f32_spmm_bench", |
| 923 | srcs = [ |
| 924 | "bench/f32-spmm.cc", |
| 925 | "bench/gemm.h", |
| 926 | "src/xnnpack/AlignedAllocator.h", |
| 927 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 928 | copts = ["-Wno-unused-function"], |
| 929 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 930 | ) |
| 931 | |
| 932 | xnnpack_benchmark( |
| 933 | name = "f32_im2col_gemm_bench", |
| 934 | srcs = [ |
| 935 | "bench/f32-im2col-gemm.cc", |
| 936 | "bench/conv.h", |
| 937 | "src/xnnpack/AlignedAllocator.h", |
| 938 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 939 | deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"], |
| 940 | ) |
| 941 | |
| 942 | ########################### Benchmarks for operators ########################### |
| 943 | |
| 944 | xnnpack_benchmark( |
| 945 | name = "add_bench", |
| 946 | srcs = ["bench/add.cc"], |
| 947 | deps = OPERATOR_BENCHMARK_DEPS, |
| 948 | ) |
| 949 | |
| 950 | xnnpack_benchmark( |
| 951 | name = "average_pooling_bench", |
| 952 | srcs = ["bench/average-pooling.cc"], |
| 953 | deps = OPERATOR_BENCHMARK_DEPS, |
| 954 | ) |
| 955 | |
| 956 | xnnpack_benchmark( |
| 957 | name = "channel_shuffle_bench", |
| 958 | srcs = ["bench/channel-shuffle.cc"], |
| 959 | deps = OPERATOR_BENCHMARK_DEPS, |
| 960 | ) |
| 961 | |
| 962 | xnnpack_benchmark( |
| 963 | name = "convolution_bench", |
| 964 | srcs = ["bench/convolution.cc"], |
| 965 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
| 966 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
| 967 | ) |
| 968 | |
| 969 | xnnpack_benchmark( |
| 970 | name = "deconvolution_bench", |
| 971 | srcs = ["bench/deconvolution.cc"], |
| 972 | copts = xnnpack_optional_tflite_copts(), |
| 973 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 974 | ) |
| 975 | |
| 976 | xnnpack_benchmark( |
| 977 | name = "global_average_pooling_bench", |
| 978 | srcs = ["bench/global-average-pooling.cc"], |
| 979 | deps = OPERATOR_BENCHMARK_DEPS, |
| 980 | ) |
| 981 | |
| 982 | xnnpack_benchmark( |
| 983 | name = "max_pooling_bench", |
| 984 | srcs = ["bench/max-pooling.cc"], |
| 985 | deps = OPERATOR_BENCHMARK_DEPS, |
| 986 | ) |
| 987 | |
| 988 | xnnpack_benchmark( |
| 989 | name = "sigmoid_bench", |
| 990 | srcs = ["bench/sigmoid.cc"], |
| 991 | deps = OPERATOR_BENCHMARK_DEPS, |
| 992 | ) |
| 993 | |
| 994 | xnnpack_benchmark( |
| 995 | name = "softargmax_bench", |
| 996 | srcs = ["bench/softargmax.cc"], |
| 997 | deps = OPERATOR_BENCHMARK_DEPS, |
| 998 | ) |
| 999 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1000 | ############################# End-to-end benchmarks ############################ |
| 1001 | |
| 1002 | cc_library( |
| 1003 | name = "mobilenet_v1", |
| 1004 | srcs = ["models/mobilenet-v1.cc"], |
| 1005 | hdrs = ["models/models.h"], |
| 1006 | linkstatic = True, |
| 1007 | deps = [ |
| 1008 | ":XNNPACK", |
| 1009 | "@pthreadpool", |
| 1010 | ], |
| 1011 | ) |
| 1012 | |
| 1013 | cc_library( |
| 1014 | name = "mobilenet_v2", |
| 1015 | srcs = ["models/mobilenet-v2.cc"], |
| 1016 | hdrs = ["models/models.h"], |
| 1017 | linkstatic = True, |
| 1018 | deps = [ |
| 1019 | ":XNNPACK", |
| 1020 | "@pthreadpool", |
| 1021 | ], |
| 1022 | ) |
| 1023 | |
| 1024 | xnnpack_benchmark( |
| 1025 | name = "end2end_bench", |
| 1026 | srcs = ["bench/end2end.cc"], |
| 1027 | deps = [ |
| 1028 | ":XNNPACK", |
| 1029 | ":mobilenet_v1", |
| 1030 | ":mobilenet_v2", |
| 1031 | "@pthreadpool", |
| 1032 | ], |
| 1033 | ) |
| 1034 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1035 | #################### Accuracy evaluation for math functions #################### |
| 1036 | |
| 1037 | xnnpack_benchmark( |
| 1038 | name = "f32_exp_eval", |
| 1039 | srcs = [ |
| 1040 | "eval/f32-exp.cc", |
| 1041 | "src/xnnpack/AlignedAllocator.h", |
| 1042 | ] + ACCURACY_EVAL_HDRS, |
| 1043 | deps = ACCURACY_EVAL_DEPS, |
| 1044 | ) |
| 1045 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1046 | ######################### Unit tests for micro-kernels ######################### |
| 1047 | |
| 1048 | xnnpack_unit_test( |
| 1049 | name = "f16_gemm_test", |
| 1050 | srcs = [ |
| 1051 | "test/f16-gemm.cc", |
| 1052 | "test/gemm-microkernel-tester.h", |
| 1053 | "src/xnnpack/AlignedAllocator.h", |
| 1054 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1055 | deps = MICROKERNEL_TEST_DEPS, |
| 1056 | ) |
| 1057 | |
| 1058 | xnnpack_unit_test( |
| 1059 | name = "f32_argmaxpool_test", |
| 1060 | srcs = [ |
| 1061 | "test/f32-argmaxpool.cc", |
| 1062 | "test/argmaxpool-microkernel-tester.h", |
| 1063 | "src/xnnpack/AlignedAllocator.h", |
| 1064 | ] + MICROKERNEL_TEST_HDRS, |
| 1065 | deps = MICROKERNEL_TEST_DEPS, |
| 1066 | ) |
| 1067 | |
| 1068 | xnnpack_unit_test( |
| 1069 | name = "f32_avgpool_test", |
| 1070 | srcs = [ |
| 1071 | "test/f32-avgpool.cc", |
| 1072 | "test/avgpool-microkernel-tester.h", |
| 1073 | "src/xnnpack/AlignedAllocator.h", |
| 1074 | ] + MICROKERNEL_TEST_HDRS, |
| 1075 | deps = MICROKERNEL_TEST_DEPS, |
| 1076 | ) |
| 1077 | |
| 1078 | xnnpack_unit_test( |
| 1079 | name = "f32_clamp_test", |
| 1080 | srcs = [ |
| 1081 | "test/f32-clamp.cc", |
| 1082 | "test/clamp-microkernel-tester.h", |
| 1083 | ] + MICROKERNEL_TEST_HDRS, |
| 1084 | deps = MICROKERNEL_TEST_DEPS, |
| 1085 | ) |
| 1086 | |
| 1087 | xnnpack_unit_test( |
| 1088 | name = "f32_igemm_test", |
| 1089 | srcs = [ |
| 1090 | "test/f32-igemm.cc", |
| 1091 | "test/gemm-microkernel-tester.h", |
| 1092 | "src/xnnpack/AlignedAllocator.h", |
| 1093 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1094 | deps = MICROKERNEL_TEST_DEPS, |
| 1095 | ) |
| 1096 | |
| 1097 | xnnpack_unit_test( |
| 1098 | name = "f32_conv_hwc_test", |
| 1099 | srcs = [ |
| 1100 | "test/f32-conv-hwc.cc", |
| 1101 | "test/conv-hwc-microkernel-tester.h", |
| 1102 | "src/xnnpack/AlignedAllocator.h", |
| 1103 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1104 | deps = MICROKERNEL_TEST_DEPS, |
| 1105 | ) |
| 1106 | |
| 1107 | xnnpack_unit_test( |
| 1108 | name = "f32_conv_hwc2spchw_test", |
| 1109 | srcs = [ |
| 1110 | "test/f32-conv-hwc2spchw.cc", |
| 1111 | "test/conv-hwc2spchw-microkernel-tester.h", |
| 1112 | "src/xnnpack/AlignedAllocator.h", |
| 1113 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1114 | deps = MICROKERNEL_TEST_DEPS, |
| 1115 | ) |
| 1116 | |
| 1117 | xnnpack_unit_test( |
| 1118 | name = "f32_dwconv_test", |
| 1119 | srcs = [ |
| 1120 | "test/f32-dwconv.cc", |
| 1121 | "test/dwconv-microkernel-tester.h", |
| 1122 | "src/xnnpack/AlignedAllocator.h", |
| 1123 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1124 | deps = MICROKERNEL_TEST_DEPS, |
| 1125 | ) |
| 1126 | |
| 1127 | xnnpack_unit_test( |
| 1128 | name = "f32_dwconv_spchw_test", |
| 1129 | srcs = [ |
| 1130 | "test/f32-dwconv-spchw.cc", |
| 1131 | "test/dwconv-spchw-microkernel-tester.h", |
| 1132 | "src/xnnpack/AlignedAllocator.h", |
| 1133 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1134 | deps = MICROKERNEL_TEST_DEPS, |
| 1135 | ) |
| 1136 | |
| 1137 | xnnpack_unit_test( |
| 1138 | name = "f32_gavgpool_test", |
| 1139 | srcs = [ |
| 1140 | "test/f32-gavgpool.cc", |
| 1141 | "test/gavgpool-microkernel-tester.h", |
| 1142 | "src/xnnpack/AlignedAllocator.h", |
| 1143 | ] + MICROKERNEL_TEST_HDRS, |
| 1144 | deps = MICROKERNEL_TEST_DEPS, |
| 1145 | ) |
| 1146 | |
| 1147 | xnnpack_unit_test( |
| 1148 | name = "f32_gavgpool_spchw_test", |
| 1149 | srcs = [ |
| 1150 | "test/f32-gavgpool-spchw.cc", |
| 1151 | "test/gavgpool-spchw-microkernel-tester.h", |
| 1152 | "src/xnnpack/AlignedAllocator.h", |
| 1153 | ] + MICROKERNEL_TEST_HDRS, |
| 1154 | deps = MICROKERNEL_TEST_DEPS, |
| 1155 | ) |
| 1156 | |
| 1157 | xnnpack_unit_test( |
| 1158 | name = "f32_gemm_test", |
| 1159 | srcs = [ |
| 1160 | "test/f32-gemm.cc", |
| 1161 | "test/gemm-microkernel-tester.h", |
| 1162 | "src/xnnpack/AlignedAllocator.h", |
| 1163 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1164 | deps = MICROKERNEL_TEST_DEPS, |
| 1165 | ) |
| 1166 | |
| 1167 | xnnpack_unit_test( |
| 1168 | name = "f32_gemminc_test", |
| 1169 | srcs = [ |
| 1170 | "test/f32-gemminc.cc", |
| 1171 | "test/gemm-microkernel-tester.h", |
| 1172 | "src/xnnpack/AlignedAllocator.h", |
| 1173 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1174 | deps = MICROKERNEL_TEST_DEPS, |
| 1175 | ) |
| 1176 | |
| 1177 | xnnpack_unit_test( |
| 1178 | name = "f32_hswish_test", |
| 1179 | srcs = [ |
| 1180 | "test/f32-hswish.cc", |
| 1181 | "test/hswish-microkernel-tester.h", |
| 1182 | ] + MICROKERNEL_TEST_HDRS, |
| 1183 | deps = MICROKERNEL_TEST_DEPS, |
| 1184 | ) |
| 1185 | |
| 1186 | xnnpack_unit_test( |
| 1187 | name = "f32_maxpool_test", |
| 1188 | srcs = [ |
| 1189 | "test/f32-maxpool.cc", |
| 1190 | "test/maxpool-microkernel-tester.h", |
| 1191 | ] + MICROKERNEL_TEST_HDRS, |
| 1192 | deps = MICROKERNEL_TEST_DEPS, |
| 1193 | ) |
| 1194 | |
| 1195 | xnnpack_unit_test( |
| 1196 | name = "f32_pavgpool_test", |
| 1197 | srcs = [ |
| 1198 | "test/f32-pavgpool.cc", |
| 1199 | "test/avgpool-microkernel-tester.h", |
| 1200 | "src/xnnpack/AlignedAllocator.h", |
| 1201 | ] + MICROKERNEL_TEST_HDRS, |
| 1202 | deps = MICROKERNEL_TEST_DEPS, |
| 1203 | ) |
| 1204 | |
| 1205 | xnnpack_unit_test( |
| 1206 | name = "f32_ppmm_test", |
| 1207 | srcs = [ |
| 1208 | "test/f32-ppmm.cc", |
| 1209 | "test/gemm-microkernel-tester.h", |
| 1210 | "src/xnnpack/AlignedAllocator.h", |
| 1211 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1212 | deps = MICROKERNEL_TEST_DEPS, |
| 1213 | ) |
| 1214 | |
| 1215 | xnnpack_unit_test( |
| 1216 | name = "f32_prelu_test", |
| 1217 | srcs = [ |
| 1218 | "test/f32-prelu.cc", |
| 1219 | "test/prelu-microkernel-tester.h", |
| 1220 | "src/xnnpack/AlignedAllocator.h", |
| 1221 | ] + MICROKERNEL_TEST_HDRS, |
| 1222 | deps = MICROKERNEL_TEST_DEPS, |
| 1223 | ) |
| 1224 | |
| 1225 | xnnpack_unit_test( |
| 1226 | name = "f32_rmax_test", |
| 1227 | srcs = [ |
| 1228 | "test/f32-rmax.cc", |
| 1229 | "test/rmax-microkernel-tester.h", |
| 1230 | ] + MICROKERNEL_TEST_HDRS, |
| 1231 | deps = MICROKERNEL_TEST_DEPS, |
| 1232 | ) |
| 1233 | |
| 1234 | xnnpack_unit_test( |
| 1235 | name = "f32_spmm_test", |
| 1236 | srcs = [ |
| 1237 | "test/f32-spmm.cc", |
| 1238 | "test/spmm-microkernel-tester.h", |
| 1239 | "src/xnnpack/AlignedAllocator.h", |
| 1240 | ] + MICROKERNEL_TEST_HDRS, |
| 1241 | deps = MICROKERNEL_TEST_DEPS, |
| 1242 | ) |
| 1243 | |
| 1244 | xnnpack_unit_test( |
| 1245 | name = "f32_vadd_test", |
| 1246 | srcs = [ |
| 1247 | "test/f32-vadd.cc", |
| 1248 | "test/vadd-microkernel-tester.h", |
| 1249 | ] + MICROKERNEL_TEST_HDRS, |
| 1250 | deps = MICROKERNEL_TEST_DEPS, |
| 1251 | ) |
| 1252 | |
| 1253 | xnnpack_unit_test( |
| 1254 | name = "f32_vsub_test", |
| 1255 | srcs = [ |
| 1256 | "test/f32-vsub.cc", |
| 1257 | "test/vsub-microkernel-tester.h", |
| 1258 | ] + MICROKERNEL_TEST_HDRS, |
| 1259 | deps = MICROKERNEL_TEST_DEPS, |
| 1260 | ) |
| 1261 | |
| 1262 | xnnpack_unit_test( |
| 1263 | name = "f32_vmul_test", |
| 1264 | srcs = [ |
| 1265 | "test/f32-vmul.cc", |
| 1266 | "test/vmul-microkernel-tester.h", |
| 1267 | ] + MICROKERNEL_TEST_HDRS, |
| 1268 | deps = MICROKERNEL_TEST_DEPS, |
| 1269 | ) |
| 1270 | |
| 1271 | xnnpack_unit_test( |
| 1272 | name = "f32_vmulcaddc_test", |
| 1273 | srcs = [ |
| 1274 | "test/f32-vmulcaddc.cc", |
| 1275 | "test/vmulcaddc-microkernel-tester.h", |
| 1276 | "src/xnnpack/AlignedAllocator.h", |
| 1277 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1278 | deps = MICROKERNEL_TEST_DEPS, |
| 1279 | ) |
| 1280 | |
| 1281 | xnnpack_unit_test( |
| 1282 | name = "q8_avgpool_test", |
| 1283 | srcs = [ |
| 1284 | "test/q8-avgpool.cc", |
| 1285 | "test/avgpool-microkernel-tester.h", |
| 1286 | "src/xnnpack/AlignedAllocator.h", |
| 1287 | ] + MICROKERNEL_TEST_HDRS, |
| 1288 | deps = MICROKERNEL_TEST_DEPS, |
| 1289 | ) |
| 1290 | |
| 1291 | xnnpack_unit_test( |
| 1292 | name = "q8_igemm_test", |
| 1293 | srcs = [ |
| 1294 | "test/q8-igemm.cc", |
| 1295 | "test/gemm-microkernel-tester.h", |
| 1296 | "src/xnnpack/AlignedAllocator.h", |
| 1297 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1298 | deps = MICROKERNEL_TEST_DEPS, |
| 1299 | ) |
| 1300 | |
| 1301 | xnnpack_unit_test( |
| 1302 | name = "q8_dwconv_test", |
| 1303 | srcs = [ |
| 1304 | "test/q8-dwconv.cc", |
| 1305 | "test/dwconv-microkernel-tester.h", |
| 1306 | "src/xnnpack/AlignedAllocator.h", |
| 1307 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1308 | deps = MICROKERNEL_TEST_DEPS, |
| 1309 | ) |
| 1310 | |
| 1311 | xnnpack_unit_test( |
| 1312 | name = "q8_gavgpool_test", |
| 1313 | srcs = [ |
| 1314 | "test/q8-gavgpool.cc", |
| 1315 | "test/gavgpool-microkernel-tester.h", |
| 1316 | "src/xnnpack/AlignedAllocator.h", |
| 1317 | ] + MICROKERNEL_TEST_HDRS, |
| 1318 | deps = MICROKERNEL_TEST_DEPS, |
| 1319 | ) |
| 1320 | |
| 1321 | xnnpack_unit_test( |
| 1322 | name = "q8_gemm_test", |
| 1323 | srcs = [ |
| 1324 | "test/q8-gemm.cc", |
| 1325 | "test/gemm-microkernel-tester.h", |
| 1326 | "src/xnnpack/AlignedAllocator.h", |
| 1327 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1328 | deps = MICROKERNEL_TEST_DEPS, |
| 1329 | ) |
| 1330 | |
| 1331 | xnnpack_unit_test( |
| 1332 | name = "q8_vadd_test", |
| 1333 | srcs = [ |
| 1334 | "test/q8-vadd.cc", |
| 1335 | "test/vadd-microkernel-tester.h", |
| 1336 | ] + MICROKERNEL_TEST_HDRS, |
| 1337 | deps = MICROKERNEL_TEST_DEPS, |
| 1338 | ) |
| 1339 | |
| 1340 | xnnpack_unit_test( |
| 1341 | name = "u8_clamp_test", |
| 1342 | srcs = [ |
| 1343 | "test/u8-clamp.cc", |
| 1344 | "test/clamp-microkernel-tester.h", |
| 1345 | ] + MICROKERNEL_TEST_HDRS, |
| 1346 | deps = MICROKERNEL_TEST_DEPS, |
| 1347 | ) |
| 1348 | |
| 1349 | xnnpack_unit_test( |
| 1350 | name = "u8_lut32norm_test", |
| 1351 | srcs = [ |
| 1352 | "test/u8-lut32norm.cc", |
| 1353 | "test/lut-norm-microkernel-tester.h", |
| 1354 | ] + MICROKERNEL_TEST_HDRS, |
| 1355 | deps = MICROKERNEL_TEST_DEPS, |
| 1356 | ) |
| 1357 | |
| 1358 | xnnpack_unit_test( |
| 1359 | name = "u8_maxpool_test", |
| 1360 | srcs = [ |
| 1361 | "test/u8-maxpool.cc", |
| 1362 | "test/maxpool-microkernel-tester.h", |
| 1363 | ] + MICROKERNEL_TEST_HDRS, |
| 1364 | deps = MICROKERNEL_TEST_DEPS, |
| 1365 | ) |
| 1366 | |
| 1367 | xnnpack_unit_test( |
| 1368 | name = "u8_rmax_test", |
| 1369 | srcs = [ |
| 1370 | "test/u8-rmax.cc", |
| 1371 | "test/rmax-microkernel-tester.h", |
| 1372 | ] + MICROKERNEL_TEST_HDRS, |
| 1373 | deps = MICROKERNEL_TEST_DEPS, |
| 1374 | ) |
| 1375 | |
| 1376 | xnnpack_unit_test( |
| 1377 | name = "x32_packx_test", |
| 1378 | srcs = [ |
| 1379 | "test/x32-packx.cc", |
| 1380 | "test/pack-microkernel-tester.h", |
| 1381 | "src/xnnpack/AlignedAllocator.h", |
| 1382 | ] + MICROKERNEL_TEST_HDRS, |
| 1383 | deps = MICROKERNEL_TEST_DEPS, |
| 1384 | ) |
| 1385 | |
| 1386 | xnnpack_unit_test( |
| 1387 | name = "x32_pad_test", |
| 1388 | srcs = [ |
| 1389 | "test/x32-pad.cc", |
| 1390 | "test/pad-microkernel-tester.h", |
| 1391 | ] + MICROKERNEL_TEST_HDRS, |
| 1392 | deps = MICROKERNEL_TEST_DEPS, |
| 1393 | ) |
| 1394 | |
| 1395 | xnnpack_unit_test( |
| 1396 | name = "x32_unpool_test", |
| 1397 | srcs = [ |
| 1398 | "test/x32-unpool.cc", |
| 1399 | "test/unpool-microkernel-tester.h", |
| 1400 | ] + MICROKERNEL_TEST_HDRS, |
| 1401 | deps = MICROKERNEL_TEST_DEPS, |
| 1402 | ) |
| 1403 | |
| 1404 | xnnpack_unit_test( |
| 1405 | name = "x32_zip_test", |
| 1406 | srcs = [ |
| 1407 | "test/x32-zip.cc", |
| 1408 | "test/zip-microkernel-tester.h", |
| 1409 | ] + MICROKERNEL_TEST_HDRS, |
| 1410 | deps = MICROKERNEL_TEST_DEPS, |
| 1411 | ) |
| 1412 | |
| 1413 | xnnpack_unit_test( |
| 1414 | name = "x8_lut_test", |
| 1415 | srcs = [ |
| 1416 | "test/x8-lut.cc", |
| 1417 | "test/lut-microkernel-tester.h", |
| 1418 | ] + MICROKERNEL_TEST_HDRS, |
| 1419 | deps = MICROKERNEL_TEST_DEPS, |
| 1420 | ) |
| 1421 | |
| 1422 | xnnpack_unit_test( |
| 1423 | name = "x8_zip_test", |
| 1424 | srcs = [ |
| 1425 | "test/x8-zip.cc", |
| 1426 | "test/zip-microkernel-tester.h", |
| 1427 | ] + MICROKERNEL_TEST_HDRS, |
| 1428 | deps = MICROKERNEL_TEST_DEPS, |
| 1429 | ) |
| 1430 | |
| 1431 | ########################### Size test for the library ########################## |
| 1432 | |
| 1433 | xnnpack_binary( |
| 1434 | name = "size_test", |
| 1435 | srcs = ["test/size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1436 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1437 | ) |
| 1438 | |
| 1439 | ########################### Unit tests for operators ########################### |
| 1440 | |
| 1441 | xnnpack_unit_test( |
| 1442 | name = "add_test", |
| 1443 | srcs = [ |
| 1444 | "test/add.cc", |
| 1445 | "test/add-operator-tester.h", |
| 1446 | ], |
| 1447 | deps = OPERATOR_TEST_DEPS, |
| 1448 | ) |
| 1449 | |
| 1450 | xnnpack_unit_test( |
| 1451 | name = "argmax_pooling_test", |
| 1452 | srcs = [ |
| 1453 | "test/argmax-pooling.cc", |
| 1454 | "test/argmax-pooling-operator-tester.h", |
| 1455 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1456 | deps = OPERATOR_TEST_DEPS, |
| 1457 | ) |
| 1458 | |
| 1459 | xnnpack_unit_test( |
| 1460 | name = "average_pooling_test", |
| 1461 | srcs = [ |
| 1462 | "test/average-pooling.cc", |
| 1463 | "test/average-pooling-operator-tester.h", |
| 1464 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1465 | deps = OPERATOR_TEST_DEPS, |
| 1466 | ) |
| 1467 | |
| 1468 | xnnpack_unit_test( |
| 1469 | name = "channel_pad_test", |
| 1470 | srcs = [ |
| 1471 | "test/channel-pad.cc", |
| 1472 | "test/channel-pad-operator-tester.h", |
| 1473 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1474 | deps = OPERATOR_TEST_DEPS, |
| 1475 | ) |
| 1476 | |
| 1477 | xnnpack_unit_test( |
| 1478 | name = "channel_shuffle_test", |
| 1479 | srcs = [ |
| 1480 | "test/channel-shuffle.cc", |
| 1481 | "test/channel-shuffle-operator-tester.h", |
| 1482 | ], |
| 1483 | deps = OPERATOR_TEST_DEPS, |
| 1484 | ) |
| 1485 | |
| 1486 | xnnpack_unit_test( |
| 1487 | name = "clamp_test", |
| 1488 | srcs = [ |
| 1489 | "test/clamp.cc", |
| 1490 | "test/clamp-operator-tester.h", |
| 1491 | ], |
| 1492 | deps = OPERATOR_TEST_DEPS, |
| 1493 | ) |
| 1494 | |
| 1495 | xnnpack_unit_test( |
| 1496 | name = "convolution_test", |
| 1497 | srcs = [ |
| 1498 | "test/convolution.cc", |
| 1499 | "test/convolution-operator-tester.h", |
| 1500 | ], |
| 1501 | deps = OPERATOR_TEST_DEPS, |
| 1502 | ) |
| 1503 | |
| 1504 | xnnpack_unit_test( |
| 1505 | name = "convolution_spnchw_test", |
| 1506 | srcs = [ |
| 1507 | "test/convolution-spnchw.cc", |
| 1508 | "test/convolution-spnchw-operator-tester.h", |
| 1509 | ], |
| 1510 | deps = OPERATOR_TEST_DEPS, |
| 1511 | ) |
| 1512 | |
| 1513 | xnnpack_unit_test( |
| 1514 | name = "deconvolution_test", |
| 1515 | srcs = [ |
| 1516 | "test/deconvolution.cc", |
| 1517 | "test/deconvolution-operator-tester.h", |
| 1518 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1519 | deps = OPERATOR_TEST_DEPS, |
| 1520 | ) |
| 1521 | |
| 1522 | xnnpack_unit_test( |
| 1523 | name = "fully_connected_test", |
| 1524 | srcs = [ |
| 1525 | "test/fully-connected.cc", |
| 1526 | "test/fully-connected-operator-tester.h", |
| 1527 | ], |
| 1528 | deps = OPERATOR_TEST_DEPS, |
| 1529 | ) |
| 1530 | |
| 1531 | xnnpack_unit_test( |
| 1532 | name = "global_average_pooling_test", |
| 1533 | srcs = [ |
| 1534 | "test/global-average-pooling.cc", |
| 1535 | "test/global-average-pooling-operator-tester.h", |
| 1536 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1537 | deps = OPERATOR_TEST_DEPS, |
| 1538 | ) |
| 1539 | |
| 1540 | xnnpack_unit_test( |
| 1541 | name = "global_average_pooling_spnchw_test", |
| 1542 | srcs = [ |
| 1543 | "test/global-average-pooling-spnchw.cc", |
| 1544 | "test/global-average-pooling-spnchw-operator-tester.h", |
| 1545 | ], |
| 1546 | deps = OPERATOR_TEST_DEPS, |
| 1547 | ) |
| 1548 | |
| 1549 | xnnpack_unit_test( |
| 1550 | name = "hardswish_test", |
| 1551 | srcs = [ |
| 1552 | "test/hardswish.cc", |
| 1553 | "test/hardswish-operator-tester.h", |
| 1554 | ], |
| 1555 | deps = OPERATOR_TEST_DEPS, |
| 1556 | ) |
| 1557 | |
| 1558 | xnnpack_unit_test( |
| 1559 | name = "leaky_relu_test", |
| 1560 | srcs = [ |
| 1561 | "test/leaky-relu.cc", |
| 1562 | "test/leaky-relu-operator-tester.h", |
| 1563 | ], |
| 1564 | deps = OPERATOR_TEST_DEPS, |
| 1565 | ) |
| 1566 | |
| 1567 | xnnpack_unit_test( |
| 1568 | name = "max_pooling_test", |
| 1569 | srcs = [ |
| 1570 | "test/max-pooling.cc", |
| 1571 | "test/max-pooling-operator-tester.h", |
| 1572 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1573 | deps = OPERATOR_TEST_DEPS, |
| 1574 | ) |
| 1575 | |
| 1576 | xnnpack_unit_test( |
| 1577 | name = "prelu_test", |
| 1578 | srcs = [ |
| 1579 | "test/prelu.cc", |
| 1580 | "test/prelu-operator-tester.h", |
| 1581 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 1582 | deps = OPERATOR_TEST_DEPS, |
| 1583 | ) |
| 1584 | |
| 1585 | xnnpack_unit_test( |
| 1586 | name = "sigmoid_test", |
| 1587 | srcs = [ |
| 1588 | "test/sigmoid.cc", |
| 1589 | "test/sigmoid-operator-tester.h", |
| 1590 | ], |
| 1591 | deps = OPERATOR_TEST_DEPS, |
| 1592 | ) |
| 1593 | |
| 1594 | xnnpack_unit_test( |
| 1595 | name = "softargmax_test", |
| 1596 | srcs = [ |
| 1597 | "test/softargmax.cc", |
| 1598 | "test/softargmax-operator-tester.h", |
| 1599 | ], |
| 1600 | deps = OPERATOR_TEST_DEPS, |
| 1601 | ) |
| 1602 | |
| 1603 | xnnpack_unit_test( |
| 1604 | name = "unpooling_test", |
| 1605 | srcs = [ |
| 1606 | "test/unpooling.cc", |
| 1607 | "test/unpooling-operator-tester.h", |
| 1608 | ], |
| 1609 | deps = OPERATOR_TEST_DEPS, |
| 1610 | ) |
| 1611 | |
| 1612 | ############################# Build configurations ############################# |
| 1613 | |
| 1614 | config_setting( |
| 1615 | name = "linux_k8", |
| 1616 | values = { |
| 1617 | "cpu": "k8", |
| 1618 | }, |
| 1619 | ) |
| 1620 | |
| 1621 | config_setting( |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 1622 | name = "linux_aarch64", |
| 1623 | values = { |
| 1624 | "cpu": "aarch64", |
| 1625 | }, |
| 1626 | ) |
| 1627 | |
| 1628 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1629 | name = "android", |
| 1630 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 1631 | ) |
| 1632 | |
| 1633 | config_setting( |
| 1634 | name = "android_armv7", |
| 1635 | values = { |
| 1636 | "crosstool_top": "//external:android/crosstool", |
| 1637 | "cpu": "armeabi-v7a", |
| 1638 | }, |
| 1639 | ) |
| 1640 | |
| 1641 | config_setting( |
| 1642 | name = "android_arm64", |
| 1643 | values = { |
| 1644 | "crosstool_top": "//external:android/crosstool", |
| 1645 | "cpu": "arm64-v8a", |
| 1646 | }, |
| 1647 | ) |
| 1648 | |
| 1649 | config_setting( |
| 1650 | name = "android_x86", |
| 1651 | values = { |
| 1652 | "crosstool_top": "//external:android/crosstool", |
| 1653 | "cpu": "x86", |
| 1654 | }, |
| 1655 | ) |
| 1656 | |
| 1657 | config_setting( |
| 1658 | name = "android_x86_64", |
| 1659 | values = { |
| 1660 | "crosstool_top": "//external:android/crosstool", |
| 1661 | "cpu": "x86_64", |
| 1662 | }, |
| 1663 | ) |
| 1664 | |
| 1665 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 1666 | name = "macos_x86_64", |
| 1667 | values = { |
| 1668 | "apple_platform_type": "macos", |
| 1669 | "cpu": "darwin", |
| 1670 | }, |
| 1671 | ) |
| 1672 | |
| 1673 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1674 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 1675 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1676 | ) |
| 1677 | |
| 1678 | config_setting( |
| 1679 | name = "emscripten_wasm", |
| 1680 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 1681 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1682 | "cpu": "wasm", |
| 1683 | }, |
| 1684 | ) |
| 1685 | |
| 1686 | config_setting( |
| 1687 | name = "emscripten_wasmsimd", |
| 1688 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 1689 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1690 | "cpu": "wasm", |
| 1691 | "features": "wasmsimd", |
| 1692 | }, |
| 1693 | ) |
| 1694 | |
| 1695 | config_setting( |
| 1696 | name = "emscripten_asmjs", |
| 1697 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 1698 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1699 | "cpu": "asmjs", |
| 1700 | }, |
| 1701 | ) |
| 1702 | |
Marat Dukhan | 174706e | 2019-10-09 16:33:53 -0700 | [diff] [blame] | 1703 | # Builds with -c dbg |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1704 | config_setting( |
| 1705 | name = "debug_build", |
| 1706 | values = { |
| 1707 | "compilation_mode": "dbg", |
| 1708 | }, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1709 | ) |
| 1710 | |
Marat Dukhan | 174706e | 2019-10-09 16:33:53 -0700 | [diff] [blame] | 1711 | # Builds with -c opt |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1712 | config_setting( |
| 1713 | name = "optimized_build", |
| 1714 | values = { |
| 1715 | "compilation_mode": "opt", |
| 1716 | }, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1717 | ) |
| 1718 | |
| 1719 | # Enables usage of assembly kernels. |
| 1720 | config_setting( |
| 1721 | name = "xnn_enable_assembly_explicit_true", |
| 1722 | define_values = {"xnn_enable_assembly": "true"}, |
| 1723 | ) |
| 1724 | |
| 1725 | # Disables usage of assembly kernels. |
| 1726 | config_setting( |
| 1727 | name = "xnn_enable_assembly_explicit_false", |
| 1728 | define_values = {"xnn_enable_assembly": "false"}, |
| 1729 | ) |