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 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 9 | 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") |
| 10 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 11 | licenses(["notice"]) |
| 12 | |
| 13 | exports_files(["LICENSE"]) |
| 14 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 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 = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 53 | "src/add-nc.c", |
| 54 | "src/argmax-pooling-nhwc.c", |
| 55 | "src/average-pooling-nhwc.c", |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 56 | "src/binary-elementwise-nd.c", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 57 | "src/channel-pad-nc.c", |
| 58 | "src/channel-shuffle-nc.c", |
| 59 | "src/clamp-nc.c", |
| 60 | "src/convolution-nchw.c", |
| 61 | "src/convolution-nhwc.c", |
| 62 | "src/deconvolution-nhwc.c", |
| 63 | "src/fully-connected-nc.c", |
| 64 | "src/global-average-pooling-ncw.c", |
| 65 | "src/global-average-pooling-nwc.c", |
| 66 | "src/hardswish-nc.c", |
| 67 | "src/leaky-relu-nc.c", |
| 68 | "src/max-pooling-nhwc.c", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 69 | "src/prelu-nc.c", |
| 70 | "src/resize-bilinear-nhwc.c", |
| 71 | "src/sigmoid-nc.c", |
| 72 | "src/softargmax-nc.c", |
| 73 | "src/unpooling-nhwc.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 74 | ] |
| 75 | |
| 76 | SCALAR_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 77 | "src/f32-argmaxpool/4x-scalar-c1.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 78 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 79 | "src/f32-argmaxpool/9x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 80 | "src/f32-avgpool/mp9p8q-scalar.c", |
| 81 | "src/f32-avgpool/up9-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 82 | "src/f32-bilinear/gen/scalar-c1.c", |
| 83 | "src/f32-bilinear/gen/scalar-c2.c", |
| 84 | "src/f32-bilinear/gen/scalar-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 85 | "src/f32-clamp/scalar.c", |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 86 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
Marat Dukhan | 6b7dfae | 2019-12-04 16:00:52 -0800 | [diff] [blame] | 87 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 88 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c", |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 89 | "src/f32-dwconv-spchw/3x3p1-scalar.c", |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 90 | "src/f32-dwconv-spchw/3x3s2p1-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 91 | "src/f32-dwconv-spchw/5x5p2-scalar.c", |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 92 | "src/f32-dwconv-spchw/5x5s2p2-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 93 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 94 | "src/f32-dwconv/gen/up1x25-scalar.c", |
| 95 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 96 | "src/f32-dwconv/gen/up1x4-scalar.c", |
| 97 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 98 | "src/f32-dwconv/gen/up1x9-scalar.c", |
| 99 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 100 | "src/f32-dwconv/gen/up2x25-scalar.c", |
| 101 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 102 | "src/f32-dwconv/gen/up2x4-scalar.c", |
| 103 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 104 | "src/f32-dwconv/gen/up2x9-scalar.c", |
Erich Elsen | 34dc2c0 | 2019-10-16 05:11:41 -0700 | [diff] [blame] | 105 | "src/f32-gavgpool-spchw/scalar-x1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 106 | "src/f32-gavgpool/mp7p7q-scalar.c", |
| 107 | "src/f32-gavgpool/up7-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 108 | "src/f32-gemm/gen-inc/1x4-scalar.c", |
| 109 | "src/f32-gemm/gen-inc/2x4-scalar.c", |
| 110 | "src/f32-gemm/gen-inc/4x4-scalar.c", |
| 111 | "src/f32-gemm/gen/1x4-scalar.c", |
| 112 | "src/f32-gemm/gen/2x4-scalar.c", |
| 113 | "src/f32-gemm/gen/4x2-scalar.c", |
| 114 | "src/f32-gemm/gen/4x4-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 115 | "src/f32-hswish/scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 116 | "src/f32-igemm/gen/1x4-scalar.c", |
| 117 | "src/f32-igemm/gen/2x4-scalar.c", |
| 118 | "src/f32-igemm/gen/4x2-scalar.c", |
| 119 | "src/f32-igemm/gen/4x4-scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 120 | "src/f32-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 121 | "src/f32-pavgpool/mp9p8q-scalar.c", |
| 122 | "src/f32-pavgpool/up9-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 123 | "src/f32-ppmm/gen/2x4-scalar.c", |
| 124 | "src/f32-ppmm/gen/3x3-scalar.c", |
| 125 | "src/f32-ppmm/gen/4x2-scalar.c", |
| 126 | "src/f32-ppmm/gen/4x4-scalar.c", |
| 127 | "src/f32-prelu/gen/scalar-2x1.c", |
| 128 | "src/f32-prelu/gen/scalar-2x4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 129 | "src/f32-rmax/scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 130 | "src/f32-spmm/gen/1x1-scalar-pipelined.c", |
| 131 | "src/f32-spmm/gen/1x1-scalar.c", |
| 132 | "src/f32-spmm/gen/2x1-scalar-pipelined.c", |
| 133 | "src/f32-spmm/gen/2x1-scalar.c", |
| 134 | "src/f32-spmm/gen/4x1-scalar-pipelined.c", |
| 135 | "src/f32-spmm/gen/4x1-scalar.c", |
| 136 | "src/f32-spmm/gen/8x1-scalar-pipelined.c", |
| 137 | "src/f32-spmm/gen/8x1-scalar.c", |
| 138 | "src/f32-spmm/gen/8x2-scalar.c", |
| 139 | "src/f32-spmm/gen/8x4-scalar.c", |
| 140 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 141 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 142 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
| 143 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 144 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 145 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
| 146 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 147 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 148 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
| 149 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 150 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 151 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
| 152 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 153 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 154 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
| 155 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 156 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 157 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
| 158 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 159 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 160 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
| 161 | "src/f32-vmulcaddc/gen/c1-scalar-2x.c", |
| 162 | "src/f32-vmulcaddc/gen/c2-scalar-2x.c", |
| 163 | "src/f32-vmulcaddc/gen/c4-scalar-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 164 | "src/q8-avgpool/mp9p8q-scalar.c", |
| 165 | "src/q8-avgpool/up9-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 166 | "src/q8-dwconv/up1x9-scalar.c", |
| 167 | "src/q8-gavgpool/mp7p7q-scalar.c", |
| 168 | "src/q8-gavgpool/up7-scalar.c", |
| 169 | "src/q8-gemm/2x2-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 170 | "src/q8-igemm/2x2-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 171 | "src/q8-vadd/scalar.c", |
| 172 | "src/u8-clamp/scalar.c", |
| 173 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 174 | "src/u8-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 175 | "src/u8-rmax/scalar.c", |
| 176 | "src/x32-packx/x2-scalar.c", |
| 177 | "src/x32-packx/x3-scalar.c", |
| 178 | "src/x32-packx/x4-scalar.c", |
| 179 | "src/x32-pad/x2-scalar.c", |
| 180 | "src/x32-unpool/scalar.c", |
| 181 | "src/x32-zip/x2-scalar.c", |
| 182 | "src/x32-zip/x3-scalar.c", |
| 183 | "src/x32-zip/x4-scalar.c", |
| 184 | "src/x32-zip/xm-scalar.c", |
| 185 | "src/x8-lut/scalar.c", |
| 186 | "src/x8-zip/x2-scalar.c", |
| 187 | "src/x8-zip/x3-scalar.c", |
| 188 | "src/x8-zip/x4-scalar.c", |
| 189 | "src/x8-zip/xm-scalar.c", |
| 190 | ] |
| 191 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 192 | WASM_UKERNELS = [ |
| 193 | "src/f32-avgpool/mp9p8q-wasm.c", |
| 194 | "src/f32-avgpool/up9-wasm.c", |
| 195 | "src/f32-clamp/wasm.c", |
| 196 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 197 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 198 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 199 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 200 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 201 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 202 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 203 | "src/f32-dwconv/gen/up2x25-wasm.c", |
| 204 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 205 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 206 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 207 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| 208 | "src/f32-gavgpool/mp7p7q-wasm.c", |
| 209 | "src/f32-gavgpool/up7-wasm.c", |
| 210 | "src/f32-gemm/gen-inc/1x4-wasm.c", |
| 211 | "src/f32-gemm/gen-inc/2x4-wasm.c", |
| 212 | "src/f32-gemm/gen-inc/4x4-wasm.c", |
| 213 | "src/f32-gemm/gen/1x4-wasm.c", |
| 214 | "src/f32-gemm/gen/2x4-wasm.c", |
| 215 | "src/f32-gemm/gen/4x2-wasm.c", |
| 216 | "src/f32-gemm/gen/4x4-wasm.c", |
| 217 | "src/f32-hswish/wasm.c", |
| 218 | "src/f32-igemm/gen/1x4-wasm.c", |
| 219 | "src/f32-igemm/gen/2x4-wasm.c", |
| 220 | "src/f32-igemm/gen/4x2-wasm.c", |
| 221 | "src/f32-igemm/gen/4x4-wasm.c", |
| 222 | "src/f32-maxpool/9p8x-wasm-c1.c", |
| 223 | "src/f32-pavgpool/mp9p8q-wasm.c", |
| 224 | "src/f32-pavgpool/up9-wasm.c", |
| 225 | "src/f32-prelu/gen/wasm-2x1.c", |
| 226 | "src/f32-prelu/gen/wasm-2x4.c", |
| 227 | "src/f32-vbinary/gen/vadd-wasm-x1.c", |
| 228 | "src/f32-vbinary/gen/vadd-wasm-x2.c", |
| 229 | "src/f32-vbinary/gen/vadd-wasm-x4.c", |
| 230 | "src/f32-vbinary/gen/vaddc-wasm-x1.c", |
| 231 | "src/f32-vbinary/gen/vaddc-wasm-x2.c", |
| 232 | "src/f32-vbinary/gen/vaddc-wasm-x4.c", |
| 233 | "src/f32-vbinary/gen/vmul-wasm-x1.c", |
| 234 | "src/f32-vbinary/gen/vmul-wasm-x2.c", |
| 235 | "src/f32-vbinary/gen/vmul-wasm-x4.c", |
| 236 | "src/f32-vbinary/gen/vmulc-wasm-x1.c", |
| 237 | "src/f32-vbinary/gen/vmulc-wasm-x2.c", |
| 238 | "src/f32-vbinary/gen/vmulc-wasm-x4.c", |
| 239 | "src/f32-vbinary/gen/vrsubc-wasm-x1.c", |
| 240 | "src/f32-vbinary/gen/vrsubc-wasm-x2.c", |
| 241 | "src/f32-vbinary/gen/vrsubc-wasm-x4.c", |
| 242 | "src/f32-vbinary/gen/vsub-wasm-x1.c", |
| 243 | "src/f32-vbinary/gen/vsub-wasm-x2.c", |
| 244 | "src/f32-vbinary/gen/vsub-wasm-x4.c", |
| 245 | "src/f32-vbinary/gen/vsubc-wasm-x1.c", |
| 246 | "src/f32-vbinary/gen/vsubc-wasm-x2.c", |
| 247 | "src/f32-vbinary/gen/vsubc-wasm-x4.c", |
| 248 | "src/f32-vmulcaddc/gen/c1-wasm-2x.c", |
| 249 | "src/f32-vmulcaddc/gen/c2-wasm-2x.c", |
| 250 | "src/f32-vmulcaddc/gen/c4-wasm-2x.c", |
| 251 | ] |
| 252 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 253 | PSIMD_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 254 | "src/f32-argmaxpool/4x-psimd-c4.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 255 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 256 | "src/f32-argmaxpool/9x-psimd-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 257 | "src/f32-avgpool/mp9p8q-psimd.c", |
| 258 | "src/f32-avgpool/up9-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 259 | "src/f32-bilinear/gen/psimd-c4.c", |
| 260 | "src/f32-bilinear/gen/psimd-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 261 | "src/f32-clamp/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 262 | "src/f32-dwconv/gen/up4x25-psimd-acc2.c", |
| 263 | "src/f32-dwconv/gen/up4x25-psimd.c", |
| 264 | "src/f32-dwconv/gen/up4x4-psimd-acc2.c", |
| 265 | "src/f32-dwconv/gen/up4x4-psimd.c", |
| 266 | "src/f32-dwconv/gen/up4x9-psimd-acc2.c", |
| 267 | "src/f32-dwconv/gen/up4x9-psimd.c", |
| 268 | "src/f32-dwconv/gen/up8x25-psimd-acc2.c", |
| 269 | "src/f32-dwconv/gen/up8x25-psimd.c", |
| 270 | "src/f32-dwconv/gen/up8x4-psimd-acc2.c", |
| 271 | "src/f32-dwconv/gen/up8x4-psimd.c", |
| 272 | "src/f32-dwconv/gen/up8x9-psimd-acc2.c", |
| 273 | "src/f32-dwconv/gen/up8x9-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 274 | "src/f32-gavgpool/mp7p7q-psimd.c", |
| 275 | "src/f32-gavgpool/up7-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 276 | "src/f32-gemm/gen/1x8-psimd-loadsplat.c", |
| 277 | "src/f32-gemm/gen/1x8-psimd-splat.c", |
| 278 | "src/f32-gemm/gen/1x8s4-psimd.c", |
| 279 | "src/f32-gemm/gen/4x8-psimd-loadsplat.c", |
| 280 | "src/f32-gemm/gen/4x8-psimd-splat.c", |
| 281 | "src/f32-gemm/gen/4x8s4-psimd.c", |
| 282 | "src/f32-gemm/gen/6x8-psimd-loadsplat.c", |
| 283 | "src/f32-gemm/gen/6x8-psimd-splat.c", |
| 284 | "src/f32-gemm/gen/6x8s4-psimd.c", |
| 285 | "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c", |
| 286 | "src/f32-gemm/gen-inc/1x8-psimd-splat.c", |
| 287 | "src/f32-gemm/gen-inc/1x8s4-psimd.c", |
| 288 | "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c", |
| 289 | "src/f32-gemm/gen-inc/4x8-psimd-splat.c", |
| 290 | "src/f32-gemm/gen-inc/4x8s4-psimd.c", |
| 291 | "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c", |
| 292 | "src/f32-gemm/gen-inc/6x8-psimd-splat.c", |
| 293 | "src/f32-gemm/gen-inc/6x8s4-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 294 | "src/f32-hswish/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 295 | "src/f32-igemm/gen/1x8-psimd-loadsplat.c", |
| 296 | "src/f32-igemm/gen/1x8-psimd-splat.c", |
| 297 | "src/f32-igemm/gen/1x8s4-psimd.c", |
| 298 | "src/f32-igemm/gen/4x2c4-psimd.c", |
| 299 | "src/f32-igemm/gen/4x8-psimd-loadsplat.c", |
| 300 | "src/f32-igemm/gen/4x8-psimd-splat.c", |
| 301 | "src/f32-igemm/gen/4x8s4-psimd.c", |
| 302 | "src/f32-igemm/gen/6x8-psimd-loadsplat.c", |
| 303 | "src/f32-igemm/gen/6x8-psimd-splat.c", |
| 304 | "src/f32-igemm/gen/6x8s4-psimd.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 305 | "src/f32-maxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 306 | "src/f32-pavgpool/mp9p8q-psimd.c", |
| 307 | "src/f32-pavgpool/up9-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 308 | "src/f32-ppmm/gen/4x8-psimd.c", |
| 309 | "src/f32-prelu/gen/psimd-2x4.c", |
| 310 | "src/f32-prelu/gen/psimd-2x8.c", |
| 311 | "src/f32-vbinary/gen/vadd-psimd-x4.c", |
| 312 | "src/f32-vbinary/gen/vadd-psimd-x8.c", |
| 313 | "src/f32-vbinary/gen/vaddc-psimd-x4.c", |
| 314 | "src/f32-vbinary/gen/vaddc-psimd-x8.c", |
| 315 | "src/f32-vbinary/gen/vmul-psimd-x4.c", |
| 316 | "src/f32-vbinary/gen/vmul-psimd-x8.c", |
| 317 | "src/f32-vbinary/gen/vmulc-psimd-x4.c", |
| 318 | "src/f32-vbinary/gen/vmulc-psimd-x8.c", |
| 319 | "src/f32-vbinary/gen/vrsubc-psimd-x4.c", |
| 320 | "src/f32-vbinary/gen/vrsubc-psimd-x8.c", |
| 321 | "src/f32-vbinary/gen/vsub-psimd-x4.c", |
| 322 | "src/f32-vbinary/gen/vsub-psimd-x8.c", |
| 323 | "src/f32-vbinary/gen/vsubc-psimd-x4.c", |
| 324 | "src/f32-vbinary/gen/vsubc-psimd-x8.c", |
| 325 | "src/f32-vmulcaddc/gen/c4-psimd-2x.c", |
| 326 | "src/f32-vmulcaddc/gen/c8-psimd-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 327 | "src/x32-packx/x4-psimd.c", |
| 328 | "src/x32-pad/x2-psimd.c", |
| 329 | "src/x32-unpool/psimd.c", |
| 330 | "src/x32-zip/x2-psimd.c", |
| 331 | "src/x32-zip/x3-psimd.c", |
| 332 | "src/x32-zip/x4-psimd.c", |
| 333 | "src/x32-zip/xm-psimd.c", |
| 334 | ] |
| 335 | |
| 336 | # ISA-specific micro-kernels |
| 337 | NEON_UKERNELS = [ |
| 338 | "src/f32-avgpool/mp9p8q-neon.c", |
| 339 | "src/f32-avgpool/up9-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 340 | "src/f32-bilinear/gen/neon-c4.c", |
| 341 | "src/f32-bilinear/gen/neon-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 342 | "src/f32-clamp/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 343 | "src/f32-dwconv/gen/up4x9-neon.c", |
| 344 | "src/f32-dwconv/gen/up4x9-neon-acc2.c", |
| 345 | "src/f32-dwconv/gen/up8x9-neon.c", |
| 346 | "src/f32-dwconv/gen/up8x9-neon-acc2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 347 | "src/f32-gavgpool-spchw/neon-x4.c", |
| 348 | "src/f32-gavgpool/mp7p7q-neon.c", |
| 349 | "src/f32-gavgpool/up7-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 350 | "src/f32-gemm/gen/1x8-neon-lane-ld64.c", |
| 351 | "src/f32-gemm/gen/4x2-neon-lane-ld64.c", |
| 352 | "src/f32-gemm/gen/4x8-neon-lane-ld128.c", |
| 353 | "src/f32-gemm/gen/4x8-neon-lane-ld64.c", |
| 354 | "src/f32-gemm/gen/5x8-neon-lane-ld64.c", |
| 355 | "src/f32-gemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 356 | "src/f32-gemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 357 | "src/f32-gemm/gen/1x8-neon-dup-ld64.c", |
| 358 | "src/f32-gemm/gen/4x8-neon-dup-ld128.c", |
| 359 | "src/f32-gemm/gen/4x8-neon-dup-ld64.c", |
| 360 | "src/f32-gemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 361 | "src/f32-gemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 362 | "src/f32-gemm/gen/1x8s4-neon.c", |
| 363 | "src/f32-gemm/gen/4x8s4-neon.c", |
| 364 | "src/f32-gemm/gen/6x8s4-neon.c", |
| 365 | "src/f32-gemm/gen/8x8s4-neon.c", |
| 366 | "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c", |
| 367 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c", |
| 368 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c", |
| 369 | "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c", |
| 370 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 371 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 372 | "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c", |
| 373 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c", |
| 374 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c", |
| 375 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 376 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 377 | "src/f32-gemm/gen-inc/1x8s4-neon.c", |
| 378 | "src/f32-gemm/gen-inc/4x8s4-neon.c", |
| 379 | "src/f32-gemm/gen-inc/6x8s4-neon.c", |
| 380 | "src/f32-gemm/gen-inc/8x8s4-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 381 | "src/f32-hswish/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 382 | "src/f32-igemm/gen/1x8-neon-lane-ld64.c", |
| 383 | "src/f32-igemm/gen/4x2-neon-lane-ld64.c", |
| 384 | "src/f32-igemm/gen/4x4-neon-lane-ld64.c", |
| 385 | "src/f32-igemm/gen/4x8-neon-lane-ld128.c", |
| 386 | "src/f32-igemm/gen/4x8-neon-lane-ld64.c", |
| 387 | "src/f32-igemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 388 | "src/f32-igemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 389 | "src/f32-igemm/gen/1x8-neon-dup-ld64.c", |
| 390 | "src/f32-igemm/gen/4x8-neon-dup-ld128.c", |
| 391 | "src/f32-igemm/gen/4x8-neon-dup-ld64.c", |
| 392 | "src/f32-igemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 393 | "src/f32-igemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 394 | "src/f32-igemm/gen/1x8s4-neon.c", |
| 395 | "src/f32-igemm/gen/4x8s4-neon.c", |
| 396 | "src/f32-igemm/gen/6x8s4-neon.c", |
| 397 | "src/f32-igemm/gen/8x8s4-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 398 | "src/f32-pavgpool/mp9p8q-neon.c", |
| 399 | "src/f32-pavgpool/up9-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 400 | "src/f32-ppmm/gen/4x8-neon.c", |
| 401 | "src/f32-ppmm/gen/8x8-neon.c", |
| 402 | "src/f32-prelu/gen/neon-2x4.c", |
| 403 | "src/f32-prelu/gen/neon-2x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 404 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 405 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
| 406 | "src/f32-vbinary/gen/vadd-neon-x4.c", |
| 407 | "src/f32-vbinary/gen/vadd-neon-x8.c", |
| 408 | "src/f32-vbinary/gen/vaddc-neon-x4.c", |
| 409 | "src/f32-vbinary/gen/vaddc-neon-x8.c", |
| 410 | "src/f32-vbinary/gen/vmul-neon-x4.c", |
| 411 | "src/f32-vbinary/gen/vmul-neon-x8.c", |
| 412 | "src/f32-vbinary/gen/vmulc-neon-x4.c", |
| 413 | "src/f32-vbinary/gen/vmulc-neon-x8.c", |
| 414 | "src/f32-vbinary/gen/vrsubc-neon-x4.c", |
| 415 | "src/f32-vbinary/gen/vrsubc-neon-x8.c", |
| 416 | "src/f32-vbinary/gen/vsub-neon-x4.c", |
| 417 | "src/f32-vbinary/gen/vsub-neon-x8.c", |
| 418 | "src/f32-vbinary/gen/vsubc-neon-x4.c", |
| 419 | "src/f32-vbinary/gen/vsubc-neon-x8.c", |
| 420 | "src/f32-vmulcaddc/gen/c4-neon-2x.c", |
| 421 | "src/f32-vmulcaddc/gen/c8-neon-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 422 | "src/q8-avgpool/mp9p8q-neon.c", |
| 423 | "src/q8-avgpool/up9-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 424 | "src/q8-dwconv/up8x9-neon.c", |
| 425 | "src/q8-gavgpool/mp7p7q-neon.c", |
| 426 | "src/q8-gavgpool/up7-neon.c", |
| 427 | "src/q8-gemm/4x8-neon.c", |
| 428 | "src/q8-gemm/8x8-neon.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 429 | "src/q8-igemm/4x8-neon.c", |
| 430 | "src/q8-igemm/8x8-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 431 | "src/q8-vadd/neon.c", |
| 432 | "src/u8-clamp/neon.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 433 | "src/u8-maxpool/9p8x-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 434 | "src/u8-rmax/neon.c", |
| 435 | "src/x32-packx/x4-neon-st4.c", |
| 436 | "src/x32-pad/x2-neon.c", |
| 437 | "src/x32-zip/x2-neon.c", |
| 438 | "src/x32-zip/x3-neon.c", |
| 439 | "src/x32-zip/x4-neon.c", |
| 440 | "src/x32-zip/xm-neon.c", |
| 441 | "src/x8-zip/x2-neon.c", |
| 442 | "src/x8-zip/x3-neon.c", |
| 443 | "src/x8-zip/x4-neon.c", |
| 444 | "src/x8-zip/xm-neon.c", |
| 445 | ] |
| 446 | |
| 447 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 448 | "src/f32-bilinear/gen/neonfma-c4.c", |
| 449 | "src/f32-bilinear/gen/neonfma-c8.c", |
| 450 | "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c", |
| 451 | "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c", |
| 452 | "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c", |
| 453 | "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 454 | "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 455 | "src/f32-igemm/gen/1x8s4-neonfma.c", |
| 456 | "src/f32-igemm/gen/4x8s4-neonfma.c", |
| 457 | "src/f32-igemm/gen/6x8s4-neonfma.c", |
| 458 | "src/f32-igemm/gen/8x8s4-neonfma.c", |
| 459 | "src/f32-dwconv/gen/up4x9-neonfma.c", |
| 460 | "src/f32-dwconv/gen/up4x9-neonfma-acc2.c", |
| 461 | "src/f32-dwconv/gen/up8x9-neonfma.c", |
| 462 | "src/f32-dwconv/gen/up8x9-neonfma-acc2.c", |
| 463 | "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c", |
| 464 | "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c", |
| 465 | "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c", |
| 466 | "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 467 | "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 468 | "src/f32-gemm/gen/1x8s4-neonfma.c", |
| 469 | "src/f32-gemm/gen/4x8s4-neonfma.c", |
| 470 | "src/f32-gemm/gen/6x8s4-neonfma.c", |
| 471 | "src/f32-gemm/gen/8x8s4-neonfma.c", |
| 472 | "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c", |
| 473 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c", |
| 474 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c", |
| 475 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 476 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 477 | "src/f32-gemm/gen-inc/1x8s4-neonfma.c", |
| 478 | "src/f32-gemm/gen-inc/4x8s4-neonfma.c", |
| 479 | "src/f32-gemm/gen-inc/6x8s4-neonfma.c", |
| 480 | "src/f32-gemm/gen-inc/8x8s4-neonfma.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 481 | "src/f32-hswish/neonfma.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 482 | "src/f32-ppmm/gen/4x8-neonfma.c", |
| 483 | "src/f32-ppmm/gen/8x8-neonfma.c", |
| 484 | "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c", |
| 485 | "src/f32-vmulcaddc/gen/c4-neonfma-2x.c", |
| 486 | "src/f32-vmulcaddc/gen/c8-neonfma-2x.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 487 | "src/math/exp-neonfma-lut64-p2.c", |
| 488 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 489 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 490 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 491 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 91f8d86 | 2019-11-27 12:25:42 -0800 | [diff] [blame] | 492 | "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c", |
| 493 | "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c", |
| 494 | "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c", |
Marat Dukhan | 22aae13 | 2019-11-22 17:10:29 -0800 | [diff] [blame] | 495 | "src/math/sigmoid-neonfma-p5-nr1recps1fma.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 496 | "src/math/sigmoid-neonfma-p5-nr2fma.c", |
Marat Dukhan | 22aae13 | 2019-11-22 17:10:29 -0800 | [diff] [blame] | 497 | "src/math/sigmoid-neonfma-p5-nr2recps.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 498 | ] |
| 499 | |
| 500 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 501 | "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c", |
| 502 | "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c", |
| 503 | "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c", |
| 504 | "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c", |
| 505 | "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c", |
| 506 | "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 507 | "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 508 | "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c", |
| 509 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c", |
| 510 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c", |
| 511 | "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c", |
| 512 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 513 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 514 | "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c", |
| 515 | "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c", |
| 516 | "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c", |
| 517 | "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c", |
| 518 | "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c", |
| 519 | "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 520 | "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 521 | "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c", |
| 522 | "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c", |
| 523 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c", |
| 524 | "src/f32-dwconv-spchw/3x3p1-neonfma.c", |
| 525 | "src/f32-dwconv-spchw/5x5p2-neonfma.c", |
| 526 | "src/f32-dwconv-spchw/3x3s2p1-neonfma.c", |
| 527 | "src/f32-dwconv-spchw/5x5s2p2-neonfma.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 528 | "src/f32-spmm/gen/12x1-neonfma.c", |
| 529 | "src/f32-spmm/gen/12x2-neonfma.c", |
| 530 | "src/f32-spmm/gen/12x4-neonfma.c", |
| 531 | "src/f32-spmm/gen/16x1-neonfma-pipelined.c", |
| 532 | "src/f32-spmm/gen/16x1-neonfma-unroll2.c", |
| 533 | "src/f32-spmm/gen/16x1-neonfma.c", |
| 534 | "src/f32-spmm/gen/16x2-neonfma.c", |
| 535 | "src/f32-spmm/gen/16x4-neonfma.c", |
| 536 | "src/f32-spmm/gen/4x1-neonfma-pipelined.c", |
| 537 | "src/f32-spmm/gen/4x1-neonfma-unroll2.c", |
| 538 | "src/f32-spmm/gen/4x1-neonfma.c", |
| 539 | "src/f32-spmm/gen/4x2-neonfma.c", |
| 540 | "src/f32-spmm/gen/4x4-neonfma.c", |
| 541 | "src/f32-spmm/gen/8x1-neonfma-pipelined.c", |
| 542 | "src/f32-spmm/gen/8x1-neonfma-unroll2.c", |
| 543 | "src/f32-spmm/gen/8x1-neonfma.c", |
| 544 | "src/f32-spmm/gen/8x2-neonfma.c", |
| 545 | "src/f32-spmm/gen/8x4-neonfma.c", |
Marat Dukhan | 22aae13 | 2019-11-22 17:10:29 -0800 | [diff] [blame] | 546 | "src/math/sigmoid-neonfma-p5-div.c", |
Marat Dukhan | 91f8d86 | 2019-11-27 12:25:42 -0800 | [diff] [blame] | 547 | "src/math/sigmoid-neonfma-lut2048-p1-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 548 | ] |
| 549 | |
| 550 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 551 | "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c", |
| 552 | "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c", |
| 553 | "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 554 | ] |
| 555 | |
| 556 | SSE_UKERNELS = [ |
| 557 | "src/f32-avgpool/mp9p8q-sse.c", |
| 558 | "src/f32-avgpool/up9-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 559 | "src/f32-bilinear/gen/sse-c4.c", |
| 560 | "src/f32-bilinear/gen/sse-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 561 | "src/f32-clamp/sse.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 562 | "src/f32-dwconv-spchw/3x3p1-sse.c", |
| 563 | "src/f32-dwconv-spchw/3x3s2p1-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 564 | "src/f32-dwconv/gen/up4x25-sse-acc2.c", |
| 565 | "src/f32-dwconv/gen/up4x25-sse.c", |
| 566 | "src/f32-dwconv/gen/up4x4-sse-acc2.c", |
| 567 | "src/f32-dwconv/gen/up4x4-sse.c", |
| 568 | "src/f32-dwconv/gen/up4x9-sse-acc2.c", |
| 569 | "src/f32-dwconv/gen/up4x9-sse.c", |
| 570 | "src/f32-dwconv/gen/up8x25-sse-acc2.c", |
| 571 | "src/f32-dwconv/gen/up8x25-sse.c", |
| 572 | "src/f32-dwconv/gen/up8x4-sse-acc2.c", |
| 573 | "src/f32-dwconv/gen/up8x4-sse.c", |
| 574 | "src/f32-dwconv/gen/up8x9-sse-acc2.c", |
| 575 | "src/f32-dwconv/gen/up8x9-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 576 | "src/f32-gavgpool-spchw/sse-x4.c", |
| 577 | "src/f32-gavgpool/mp7p7q-sse.c", |
| 578 | "src/f32-gavgpool/up7-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 579 | "src/f32-gemm/gen/1x8-sse-dup.c", |
| 580 | "src/f32-gemm/gen/1x8-sse-load1.c", |
| 581 | "src/f32-gemm/gen/1x8s4-sse.c", |
| 582 | "src/f32-gemm/gen/4x8-sse-dup.c", |
| 583 | "src/f32-gemm/gen/4x8-sse-load1.c", |
| 584 | "src/f32-gemm/gen/4x8s4-sse.c", |
| 585 | "src/f32-gemm/gen-inc/1x8-sse-dup.c", |
| 586 | "src/f32-gemm/gen-inc/1x8-sse-load1.c", |
| 587 | "src/f32-gemm/gen-inc/1x8s4-sse.c", |
| 588 | "src/f32-gemm/gen-inc/4x8-sse-dup.c", |
| 589 | "src/f32-gemm/gen-inc/4x8-sse-load1.c", |
| 590 | "src/f32-gemm/gen-inc/4x8s4-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 591 | "src/f32-hswish/sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 592 | "src/f32-igemm/gen/1x8-sse-dup.c", |
| 593 | "src/f32-igemm/gen/1x8-sse-load1.c", |
| 594 | "src/f32-igemm/gen/1x8s4-sse.c", |
| 595 | "src/f32-igemm/gen/4x2c4-sse.c", |
| 596 | "src/f32-igemm/gen/4x8-sse-dup.c", |
| 597 | "src/f32-igemm/gen/4x8-sse-load1.c", |
| 598 | "src/f32-igemm/gen/4x8s4-sse.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 599 | "src/f32-maxpool/9p8x-sse-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 600 | "src/f32-pavgpool/mp9p8q-sse.c", |
| 601 | "src/f32-pavgpool/up9-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 602 | "src/f32-ppmm/gen/4x8-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 603 | "src/f32-rmax/sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 604 | "src/f32-spmm/gen/4x1-sse.c", |
| 605 | "src/f32-spmm/gen/8x1-sse.c", |
| 606 | "src/f32-vbinary/gen/vadd-sse-x4.c", |
| 607 | "src/f32-vbinary/gen/vadd-sse-x8.c", |
| 608 | "src/f32-vbinary/gen/vaddc-sse-x4.c", |
| 609 | "src/f32-vbinary/gen/vaddc-sse-x8.c", |
| 610 | "src/f32-vbinary/gen/vmul-sse-x4.c", |
| 611 | "src/f32-vbinary/gen/vmul-sse-x8.c", |
| 612 | "src/f32-vbinary/gen/vmulc-sse-x4.c", |
| 613 | "src/f32-vbinary/gen/vmulc-sse-x8.c", |
| 614 | "src/f32-vbinary/gen/vrsubc-sse-x4.c", |
| 615 | "src/f32-vbinary/gen/vrsubc-sse-x8.c", |
| 616 | "src/f32-vbinary/gen/vsub-sse-x4.c", |
| 617 | "src/f32-vbinary/gen/vsub-sse-x8.c", |
| 618 | "src/f32-vbinary/gen/vsubc-sse-x4.c", |
| 619 | "src/f32-vbinary/gen/vsubc-sse-x8.c", |
| 620 | "src/f32-vmulcaddc/gen/c4-sse-2x.c", |
| 621 | "src/f32-vmulcaddc/gen/c8-sse-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 622 | "src/x32-packx/x4-sse.c", |
| 623 | ] |
| 624 | |
| 625 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 626 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 627 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 628 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 629 | "src/f32-prelu/gen/sse2-2x4.c", |
| 630 | "src/f32-prelu/gen/sse2-2x8.c", |
| 631 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
| 632 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 633 | "src/q8-avgpool/mp9p8q-sse2.c", |
| 634 | "src/q8-avgpool/up9-sse2.c", |
| 635 | "src/q8-igemm/4x4c2-sse2.c", |
| 636 | "src/q8-dwconv/up8x9-sse2.c", |
| 637 | "src/q8-gavgpool/mp7p7q-sse2.c", |
| 638 | "src/q8-gavgpool/up7-sse2.c", |
| 639 | "src/q8-gemm/2x4c8-sse2.c", |
| 640 | "src/q8-gemm/4x4c2-sse2.c", |
| 641 | "src/q8-vadd/sse2.c", |
| 642 | "src/u8-clamp/sse2.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 643 | "src/u8-maxpool/9p8x-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 644 | "src/u8-rmax/sse2.c", |
| 645 | "src/x32-pad/x2-sse2.c", |
| 646 | "src/x32-zip/x2-sse2.c", |
| 647 | "src/x32-zip/x3-sse2.c", |
| 648 | "src/x32-zip/x4-sse2.c", |
| 649 | "src/x32-zip/xm-sse2.c", |
| 650 | "src/x8-zip/x2-sse2.c", |
| 651 | "src/x8-zip/x3-sse2.c", |
| 652 | "src/x8-zip/x4-sse2.c", |
| 653 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 654 | "src/math/exp-sse2-p5.c", |
| 655 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 656 | "src/math/sigmoid-sse2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 657 | ] |
| 658 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 659 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 660 | "src/f32-prelu/gen/sse41-2x4.c", |
| 661 | "src/f32-prelu/gen/sse41-2x8.c", |
| 662 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
| 663 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 664 | ] |
| 665 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 666 | AVX_UKERNELS = [ |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 667 | "src/f32-clamp/avx.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 668 | "src/f32-dwconv/gen/up16x4-avx-acc2.c", |
| 669 | "src/f32-dwconv/gen/up16x4-avx.c", |
| 670 | "src/f32-dwconv/gen/up8x4-avx-acc2.c", |
| 671 | "src/f32-dwconv/gen/up8x4-avx.c", |
| 672 | "src/f32-dwconv/gen/up16x9-avx-acc2.c", |
| 673 | "src/f32-dwconv/gen/up16x9-avx.c", |
| 674 | "src/f32-dwconv/gen/up8x9-avx-acc2.c", |
| 675 | "src/f32-dwconv/gen/up8x9-avx.c", |
| 676 | "src/f32-dwconv/gen/up16x25-avx-acc2.c", |
| 677 | "src/f32-dwconv/gen/up16x25-avx.c", |
| 678 | "src/f32-dwconv/gen/up8x25-avx-acc2.c", |
| 679 | "src/f32-dwconv/gen/up8x25-avx.c", |
| 680 | "src/f32-gemm/gen/1x8-avx-broadcast.c", |
| 681 | "src/f32-gemm/gen/4x8-avx-broadcast.c", |
| 682 | "src/f32-gemm/gen/5x8-avx-broadcast.c", |
| 683 | "src/f32-gemm/gen/6x8-avx-broadcast.c", |
| 684 | "src/f32-gemm/gen/7x8-avx-broadcast.c", |
| 685 | "src/f32-gemm/gen-inc/1x8-avx-broadcast.c", |
| 686 | "src/f32-gemm/gen-inc/4x8-avx-broadcast.c", |
| 687 | "src/f32-gemm/gen-inc/5x8-avx-broadcast.c", |
| 688 | "src/f32-gemm/gen-inc/6x8-avx-broadcast.c", |
| 689 | "src/f32-gemm/gen-inc/7x8-avx-broadcast.c", |
| 690 | "src/f32-igemm/gen/1x8-avx-broadcast.c", |
| 691 | "src/f32-igemm/gen/4x8-avx-broadcast.c", |
| 692 | "src/f32-igemm/gen/5x8-avx-broadcast.c", |
| 693 | "src/f32-igemm/gen/6x8-avx-broadcast.c", |
| 694 | "src/f32-igemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 695 | "src/f32-rmax/avx.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 696 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 697 | ] |
| 698 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 699 | FMA3_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 700 | "src/f32-dwconv/gen/up16x4-fma3-acc2.c", |
| 701 | "src/f32-dwconv/gen/up16x4-fma3.c", |
| 702 | "src/f32-dwconv/gen/up8x4-fma3-acc2.c", |
| 703 | "src/f32-dwconv/gen/up8x4-fma3.c", |
| 704 | "src/f32-dwconv/gen/up16x9-fma3-acc2.c", |
| 705 | "src/f32-dwconv/gen/up16x9-fma3.c", |
| 706 | "src/f32-dwconv/gen/up8x9-fma3-acc2.c", |
| 707 | "src/f32-dwconv/gen/up8x9-fma3.c", |
| 708 | "src/f32-dwconv/gen/up16x25-fma3-acc2.c", |
| 709 | "src/f32-dwconv/gen/up16x25-fma3.c", |
| 710 | "src/f32-dwconv/gen/up8x25-fma3-acc2.c", |
| 711 | "src/f32-dwconv/gen/up8x25-fma3.c", |
| 712 | "src/f32-gemm/gen/1x8-fma3-broadcast.c", |
| 713 | "src/f32-gemm/gen/4x8-fma3-broadcast.c", |
| 714 | "src/f32-gemm/gen/5x8-fma3-broadcast.c", |
| 715 | "src/f32-gemm/gen/6x8-fma3-broadcast.c", |
| 716 | "src/f32-gemm/gen/7x8-fma3-broadcast.c", |
| 717 | "src/f32-gemm/gen/8x8-fma3-broadcast.c", |
| 718 | "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c", |
| 719 | "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c", |
| 720 | "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c", |
| 721 | "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c", |
| 722 | "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c", |
| 723 | "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c", |
| 724 | "src/f32-igemm/gen/1x8-fma3-broadcast.c", |
| 725 | "src/f32-igemm/gen/4x8-fma3-broadcast.c", |
| 726 | "src/f32-igemm/gen/5x8-fma3-broadcast.c", |
| 727 | "src/f32-igemm/gen/6x8-fma3-broadcast.c", |
| 728 | "src/f32-igemm/gen/7x8-fma3-broadcast.c", |
| 729 | "src/f32-igemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 730 | ] |
| 731 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 732 | AVX2_UKERNELS = [ |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 733 | "src/f32-raddexpminusmax/avx2-p5-unroll64.c", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 734 | "src/f32-raddextexp/avx2-p5-unroll64.c", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 735 | "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c", |
| 736 | "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 737 | "src/f32-vscaleextexp/avx2-p5-unroll64.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 738 | "src/math/exp-avx2-p5.c", |
| 739 | "src/math/exp-avx2-perm-p3.c", |
| 740 | "src/math/exp-avx2-perm-p4.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 741 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 742 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 743 | ] |
| 744 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 745 | AVX512F_UKERNELS = [ |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 746 | "src/f32-clamp/avx512f.c", |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 747 | "src/f32-dwconv/gen/up32x4-avx512f-acc2.c", |
| 748 | "src/f32-dwconv/gen/up32x4-avx512f.c", |
| 749 | "src/f32-dwconv/gen/up16x4-avx512f-acc2.c", |
| 750 | "src/f32-dwconv/gen/up16x4-avx512f.c", |
| 751 | "src/f32-dwconv/gen/up32x9-avx512f-acc2.c", |
| 752 | "src/f32-dwconv/gen/up32x9-avx512f.c", |
| 753 | "src/f32-dwconv/gen/up16x9-avx512f-acc2.c", |
| 754 | "src/f32-dwconv/gen/up16x9-avx512f.c", |
| 755 | "src/f32-dwconv/gen/up32x25-avx512f-acc2.c", |
| 756 | "src/f32-dwconv/gen/up32x25-avx512f.c", |
| 757 | "src/f32-dwconv/gen/up16x25-avx512f-acc2.c", |
| 758 | "src/f32-dwconv/gen/up16x25-avx512f.c", |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 759 | "src/f32-gemm/gen/1x16-avx512f-broadcast.c", |
| 760 | "src/f32-gemm/gen/4x16-avx512f-broadcast.c", |
| 761 | "src/f32-gemm/gen/5x16-avx512f-broadcast.c", |
| 762 | "src/f32-gemm/gen/6x16-avx512f-broadcast.c", |
| 763 | "src/f32-gemm/gen/7x16-avx512f-broadcast.c", |
| 764 | "src/f32-gemm/gen/8x16-avx512f-broadcast.c", |
| 765 | "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c", |
| 766 | "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c", |
| 767 | "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c", |
| 768 | "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c", |
| 769 | "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c", |
| 770 | "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c", |
| 771 | "src/f32-igemm/gen/1x16-avx512f-broadcast.c", |
| 772 | "src/f32-igemm/gen/4x16-avx512f-broadcast.c", |
| 773 | "src/f32-igemm/gen/5x16-avx512f-broadcast.c", |
| 774 | "src/f32-igemm/gen/6x16-avx512f-broadcast.c", |
| 775 | "src/f32-igemm/gen/7x16-avx512f-broadcast.c", |
| 776 | "src/f32-igemm/gen/8x16-avx512f-broadcast.c", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 777 | "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 778 | "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 779 | "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 780 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 781 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 782 | "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 783 | "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 784 | "src/math/exp-avx512f-p5-scalef.c", |
| 785 | "src/math/exp-avx512f-p5.c", |
| 786 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | feb4923 | 2019-10-28 11:03:31 -0700 | [diff] [blame] | 787 | "src/math/exp-avx512f-perm2-p2.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 788 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 789 | ] |
| 790 | |
| 791 | AARCH32_ASM_UKERNELS = [ |
| 792 | "src/q8-dwconv/up8x9-aarch32-neon.S", |
Frank Barchard | 3e237f2 | 2019-12-04 23:08:51 -0800 | [diff] [blame] | 793 | "src/f32-gemm/4x8-aarch32-neon-cortex-a75.S", |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 794 | "src/f32-gemm/4x8-aarch32-neon-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 795 | ] |
| 796 | |
| 797 | AARCH64_ASM_UKERNELS = [ |
| 798 | "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S", |
| 799 | "src/f32-dwconv/up4x9-aarch64-neonfma.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 800 | "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S", |
| 801 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S", |
| 802 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 803 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
| 804 | "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S", |
| 805 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S", |
| 806 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 807 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 808 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S", |
| 809 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S", |
| 810 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
| 811 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S", |
| 812 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
| 813 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S", |
| 814 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
| 815 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S", |
| 816 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S", |
| 817 | "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S", |
| 818 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S", |
| 819 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S", |
| 820 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S", |
| 821 | "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S", |
| 822 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S", |
| 823 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S", |
| 824 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S", |
| 825 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S", |
| 826 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S", |
| 827 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S", |
| 828 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S", |
| 829 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S", |
| 830 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S", |
| 831 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S", |
| 832 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S", |
| 833 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 834 | "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 835 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 836 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S", |
| 837 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S", |
| 838 | "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 839 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 840 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S", |
| 841 | "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 842 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 843 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S", |
| 844 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S", |
| 845 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S", |
| 846 | ] |
| 847 | |
| 848 | INTERNAL_MICROKERNEL_HDRS = [ |
| 849 | "src/xnnpack/argmaxpool.h", |
| 850 | "src/xnnpack/avgpool.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 851 | "src/xnnpack/bilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 852 | "src/xnnpack/clamp.h", |
| 853 | "src/xnnpack/common.h", |
| 854 | "src/xnnpack/conv.h", |
| 855 | "src/xnnpack/dwconv.h", |
| 856 | "src/xnnpack/gavgpool.h", |
| 857 | "src/xnnpack/gemm.h", |
| 858 | "src/xnnpack/hswish.h", |
| 859 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame^] | 860 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 861 | "src/xnnpack/lut.h", |
| 862 | "src/xnnpack/math.h", |
| 863 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 864 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 865 | "src/xnnpack/packx.h", |
| 866 | "src/xnnpack/pad.h", |
| 867 | "src/xnnpack/params.h", |
| 868 | "src/xnnpack/pavgpool.h", |
| 869 | "src/xnnpack/ppmm.h", |
| 870 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 871 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 872 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 873 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 874 | "src/xnnpack/rmax.h", |
| 875 | "src/xnnpack/scalar-utils.h", |
| 876 | "src/xnnpack/spmm.h", |
| 877 | "src/xnnpack/unpool.h", |
| 878 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 879 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 880 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 881 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 882 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 883 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 884 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 885 | "src/xnnpack/zip.h", |
| 886 | ] |
| 887 | |
| 888 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 889 | "include/xnnpack.h", |
| 890 | "src/xnnpack/allocator.h", |
| 891 | "src/xnnpack/compute.h", |
| 892 | "src/xnnpack/im2col.h", |
| 893 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 894 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 895 | "src/xnnpack/operator.h", |
| 896 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 897 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 898 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 899 | "src/xnnpack/requantization.h", |
| 900 | ] |
| 901 | |
| 902 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 903 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 904 | ] |
| 905 | |
| 906 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 907 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 908 | "include/xnnpack.h", |
| 909 | ] |
| 910 | |
| 911 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 912 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 913 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 914 | "src/xnnpack/requantization.h", |
| 915 | "include/xnnpack.h", |
| 916 | ] |
| 917 | |
| 918 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 919 | "src/xnnpack/params.h", |
| 920 | "src/xnnpack/common.h", |
| 921 | ] |
| 922 | |
| 923 | WEIGHTS_PACK_HDRS = [ |
| 924 | "src/xnnpack/pack.h", |
| 925 | "src/xnnpack/operator.h", |
| 926 | "src/xnnpack/compute.h", |
| 927 | ] |
| 928 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 929 | LOGGING_COPTS = select({ |
| 930 | # No logging in optimized mode |
| 931 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 932 | # Full logging in debug mode |
| 933 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 934 | # Error-only logging in default (fastbuild) mode |
| 935 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 936 | }) |
| 937 | |
| 938 | LOGGING_HDRS = [ |
| 939 | "src/xnnpack/log.h", |
| 940 | ] |
| 941 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 942 | xnnpack_cc_library( |
| 943 | name = "scalar_ukernels", |
| 944 | srcs = SCALAR_UKERNELS, |
| 945 | hdrs = INTERNAL_HDRS, |
| 946 | aarch32_copts = ["-marm"], |
| 947 | copts = xnnpack_std_copts(), |
| 948 | deps = [ |
| 949 | "@FP16", |
| 950 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 951 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 952 | ], |
| 953 | ) |
| 954 | |
| 955 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 956 | name = "wasm_ukernels", |
| 957 | hdrs = INTERNAL_HDRS, |
| 958 | copts = xnnpack_std_copts(), |
| 959 | wasm_srcs = WASM_UKERNELS, |
| 960 | deps = [ |
| 961 | "@FP16", |
| 962 | "@FXdiv", |
| 963 | "@pthreadpool", |
| 964 | ], |
| 965 | ) |
| 966 | |
| 967 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 968 | name = "psimd_ukernels", |
| 969 | srcs = PSIMD_UKERNELS, |
| 970 | hdrs = INTERNAL_HDRS, |
| 971 | aarch32_copts = [ |
| 972 | "-marm", |
| 973 | "-mfpu=neon", |
| 974 | ], |
| 975 | copts = xnnpack_std_copts(), |
| 976 | optimized_copts = [ |
| 977 | "-O3", |
| 978 | "-ffast-math", |
| 979 | ], |
| 980 | deps = [ |
| 981 | "@FP16", |
| 982 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 983 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 984 | ], |
| 985 | ) |
| 986 | |
| 987 | xnnpack_cc_library( |
| 988 | name = "neon_ukernels", |
| 989 | hdrs = INTERNAL_HDRS, |
| 990 | aarch32_copts = [ |
| 991 | "-marm", |
| 992 | "-mfpu=neon", |
| 993 | ], |
| 994 | aarch32_srcs = NEON_UKERNELS, |
| 995 | aarch64_srcs = NEON_UKERNELS, |
| 996 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 997 | deps = [ |
| 998 | "@FP16", |
| 999 | "@pthreadpool", |
| 1000 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1001 | ) |
| 1002 | |
| 1003 | xnnpack_cc_library( |
| 1004 | name = "neonfma_ukernels", |
| 1005 | hdrs = INTERNAL_HDRS, |
| 1006 | aarch32_copts = [ |
| 1007 | "-marm", |
| 1008 | "-mfpu=neon-vfpv4", |
| 1009 | ], |
| 1010 | aarch32_srcs = NEONFMA_UKERNELS, |
| 1011 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 1012 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1013 | deps = [ |
| 1014 | "@FP16", |
| 1015 | "@pthreadpool", |
| 1016 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1017 | ) |
| 1018 | |
| 1019 | xnnpack_cc_library( |
| 1020 | name = "neonfp16arith_ukernels", |
| 1021 | hdrs = INTERNAL_HDRS, |
| 1022 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 1023 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 1024 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1025 | deps = [ |
| 1026 | "@FP16", |
| 1027 | "@pthreadpool", |
| 1028 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1029 | ) |
| 1030 | |
| 1031 | xnnpack_cc_library( |
| 1032 | name = "sse2_ukernels", |
| 1033 | hdrs = INTERNAL_HDRS, |
| 1034 | copts = xnnpack_std_copts(), |
| 1035 | x86_copts = ["-msse2"], |
| 1036 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1037 | deps = [ |
| 1038 | "@FP16", |
| 1039 | "@pthreadpool", |
| 1040 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1041 | ) |
| 1042 | |
| 1043 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1044 | name = "sse41_ukernels", |
| 1045 | hdrs = INTERNAL_HDRS, |
| 1046 | copts = xnnpack_std_copts(), |
| 1047 | x86_copts = ["-msse4.1"], |
| 1048 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1049 | deps = [ |
| 1050 | "@FP16", |
| 1051 | "@pthreadpool", |
| 1052 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1053 | ) |
| 1054 | |
| 1055 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1056 | name = "avx_ukernels", |
| 1057 | hdrs = INTERNAL_HDRS, |
| 1058 | copts = xnnpack_std_copts(), |
| 1059 | x86_copts = ["-mavx"], |
| 1060 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1061 | deps = [ |
| 1062 | "@FP16", |
| 1063 | "@pthreadpool", |
| 1064 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1065 | ) |
| 1066 | |
| 1067 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1068 | name = "fma3_ukernels", |
| 1069 | hdrs = INTERNAL_HDRS, |
| 1070 | copts = xnnpack_std_copts(), |
| 1071 | x86_copts = [ |
| 1072 | "-mfma", |
| 1073 | ], |
| 1074 | x86_srcs = FMA3_UKERNELS, |
| 1075 | deps = [ |
| 1076 | "@FP16", |
| 1077 | "@pthreadpool", |
| 1078 | ], |
| 1079 | ) |
| 1080 | |
| 1081 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1082 | name = "avx2_ukernels", |
| 1083 | hdrs = INTERNAL_HDRS, |
| 1084 | copts = xnnpack_std_copts(), |
| 1085 | x86_copts = [ |
| 1086 | "-mfma", |
| 1087 | "-mavx2", |
| 1088 | ], |
| 1089 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1090 | deps = [ |
| 1091 | "@FP16", |
| 1092 | "@pthreadpool", |
| 1093 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1094 | ) |
| 1095 | |
| 1096 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1097 | name = "avx512f_ukernels", |
| 1098 | hdrs = INTERNAL_HDRS, |
| 1099 | copts = xnnpack_std_copts(), |
| 1100 | x86_copts = ["-mavx512f"], |
| 1101 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1102 | deps = [ |
| 1103 | "@FP16", |
| 1104 | "@pthreadpool", |
| 1105 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1106 | ) |
| 1107 | |
| 1108 | xnnpack_cc_library( |
| 1109 | name = "asm_ukernels", |
| 1110 | hdrs = ["src/xnnpack/assembly.h"], |
| 1111 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
| 1112 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 1113 | ) |
| 1114 | |
| 1115 | xnnpack_aggregate_library( |
| 1116 | name = "ukernels", |
| 1117 | aarch32_deps = [ |
| 1118 | ":psimd_ukernels", |
| 1119 | ":neon_ukernels", |
| 1120 | ":neonfma_ukernels", |
| 1121 | ":asm_ukernels", |
| 1122 | ], |
| 1123 | aarch64_deps = [ |
| 1124 | ":psimd_ukernels", |
| 1125 | ":neon_ukernels", |
| 1126 | ":neonfma_ukernels", |
| 1127 | ":neonfp16arith_ukernels", |
| 1128 | ":asm_ukernels", |
| 1129 | ], |
| 1130 | generic_deps = [":scalar_ukernels"], |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1131 | wasm_deps = [ |
| 1132 | ":wasm_ukernels", |
| 1133 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1134 | wasmsimd_deps = [ |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1135 | ":wasm_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1136 | ":psimd_ukernels", |
| 1137 | ], |
| 1138 | x86_deps = [ |
| 1139 | ":psimd_ukernels", |
| 1140 | ":sse2_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1141 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1142 | ":avx_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1143 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1144 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1145 | ":avx512f_ukernels", |
| 1146 | ], |
| 1147 | ) |
| 1148 | |
| 1149 | xnnpack_cc_library( |
| 1150 | name = "im2col", |
| 1151 | srcs = ["src/im2col.c"], |
| 1152 | hdrs = [ |
| 1153 | "src/xnnpack/common.h", |
| 1154 | "src/xnnpack/im2col.h", |
| 1155 | ], |
| 1156 | copts = xnnpack_std_copts(), |
| 1157 | ) |
| 1158 | |
| 1159 | xnnpack_cc_library( |
| 1160 | name = "indirection", |
| 1161 | srcs = ["src/indirection.c"], |
| 1162 | hdrs = INTERNAL_HDRS, |
| 1163 | copts = xnnpack_std_copts(), |
| 1164 | deps = [ |
| 1165 | "@FP16", |
| 1166 | "@FXdiv", |
| 1167 | "@pthreadpool", |
| 1168 | ], |
| 1169 | ) |
| 1170 | |
| 1171 | xnnpack_cc_library( |
| 1172 | name = "operator_run", |
| 1173 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1174 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 1175 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1176 | # Wrappers for multi-pass microkernels use VLAs for temporary buffers. |
| 1177 | "-Wno-vla", |
| 1178 | ], |
| 1179 | deps = [ |
| 1180 | "@FP16", |
| 1181 | "@FXdiv", |
| 1182 | "@clog", |
| 1183 | "@pthreadpool", |
| 1184 | ], |
| 1185 | ) |
| 1186 | |
| 1187 | cc_library( |
| 1188 | name = "enable_assembly", |
| 1189 | defines = select({ |
| 1190 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 1191 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 1192 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1193 | }), |
| 1194 | ) |
| 1195 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1196 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1197 | name = "operators", |
| 1198 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1199 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1200 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1201 | ], |
| 1202 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1203 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1204 | "-Isrc", |
| 1205 | "-Iinclude", |
| 1206 | ] + select({ |
| 1207 | ":debug_build": [], |
| 1208 | "//conditions:default": xnnpack_min_size_copts(), |
| 1209 | }), |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1210 | wasm_srcs = ["src/wasm-stubs.c"], |
| 1211 | wasmsimd_srcs = ["src/wasm-stubs.c"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1212 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1213 | ":indirection", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1214 | "@FP16", |
| 1215 | "@FXdiv", |
| 1216 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1217 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1218 | ], |
| 1219 | ) |
| 1220 | |
| 1221 | cc_library( |
| 1222 | name = "XNNPACK", |
| 1223 | srcs = [ |
| 1224 | "src/init.c", |
| 1225 | ], |
| 1226 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1227 | "-Isrc", |
| 1228 | "-Iinclude", |
| 1229 | ] + select({ |
| 1230 | ":debug_build": [], |
| 1231 | "//conditions:default": xnnpack_min_size_copts(), |
| 1232 | }), |
| 1233 | includes = ["include"], |
| 1234 | linkstatic = True, |
| 1235 | textual_hdrs = ["include/xnnpack.h"], |
| 1236 | visibility = xnnpack_visibility(), |
| 1237 | deps = [ |
| 1238 | ":enable_assembly", |
| 1239 | ":ukernels", |
| 1240 | ":operator_run", |
| 1241 | ":operators", |
| 1242 | "@clog", |
| 1243 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 1244 | ] + select({ |
| 1245 | ":emscripten": [], |
| 1246 | "//conditions:default": ["@cpuinfo"], |
| 1247 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1248 | ) |
| 1249 | |
| 1250 | cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1251 | name = "xnnpack_operators_nhwc_f32", |
| 1252 | srcs = [ |
| 1253 | "src/init.c", |
| 1254 | ], |
| 1255 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1256 | "-Isrc", |
| 1257 | "-Iinclude", |
| 1258 | ] + select({ |
| 1259 | ":debug_build": [], |
| 1260 | "//conditions:default": xnnpack_min_size_copts(), |
| 1261 | }), |
| 1262 | defines = [ |
| 1263 | "XNN_NO_Q8_OPERATORS", |
| 1264 | "XNN_NO_U8_OPERATORS", |
| 1265 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1266 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1267 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1268 | includes = ["include"], |
| 1269 | linkstatic = True, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1270 | textual_hdrs = ["include/xnnpack.h"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1271 | visibility = xnnpack_visibility(), |
| 1272 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1273 | ":enable_assembly", |
| 1274 | ":ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1275 | ":operator_run", |
| 1276 | ":operators", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1277 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1278 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1279 | ] + select({ |
| 1280 | ":emscripten": [], |
| 1281 | "//conditions:default": ["@cpuinfo"], |
| 1282 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1283 | ) |
| 1284 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1285 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1286 | name = "bench_utils", |
| 1287 | srcs = ["bench/utils.cc"], |
| 1288 | hdrs = ["bench/utils.h"], |
| 1289 | copts = ["-Wno-unused-result"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 1290 | deps = [ |
| 1291 | "@com_google_benchmark//:benchmark", |
| 1292 | "@cpuinfo", |
| 1293 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1294 | ) |
| 1295 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1296 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1297 | |
| 1298 | xnnpack_benchmark( |
| 1299 | name = "q8_gemm_bench", |
| 1300 | srcs = [ |
| 1301 | "bench/gemm.h", |
| 1302 | "bench/q8-gemm.cc", |
| 1303 | "src/xnnpack/AlignedAllocator.h", |
| 1304 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1305 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 1306 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
| 1307 | ) |
| 1308 | |
| 1309 | xnnpack_benchmark( |
| 1310 | name = "f16_gemm_bench", |
| 1311 | srcs = [ |
| 1312 | "bench/f16-gemm.cc", |
| 1313 | "bench/gemm.h", |
| 1314 | "src/xnnpack/AlignedAllocator.h", |
| 1315 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1316 | copts = ["-Wno-unused-function"], |
| 1317 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1318 | ) |
| 1319 | |
| 1320 | xnnpack_benchmark( |
| 1321 | name = "f32_igemm_bench", |
| 1322 | srcs = [ |
| 1323 | "bench/f32-igemm.cc", |
| 1324 | "bench/conv.h", |
| 1325 | "src/xnnpack/AlignedAllocator.h", |
| 1326 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1327 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1328 | ) |
| 1329 | |
| 1330 | xnnpack_benchmark( |
| 1331 | name = "f32_conv_hwc_bench", |
| 1332 | srcs = [ |
| 1333 | "bench/f32-conv-hwc.cc", |
| 1334 | "bench/dconv.h", |
| 1335 | "src/xnnpack/AlignedAllocator.h", |
| 1336 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1337 | copts = ["-Wno-unused-function"], |
| 1338 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1339 | ) |
| 1340 | |
| 1341 | xnnpack_benchmark( |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 1342 | name = "f32_conv_hwc2spchw_bench", |
| 1343 | srcs = [ |
| 1344 | "bench/f32-conv-hwc2spchw.cc", |
| 1345 | "bench/dconv.h", |
| 1346 | "src/xnnpack/AlignedAllocator.h", |
| 1347 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1348 | copts = ["-Wno-unused-function"], |
| 1349 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1350 | ) |
| 1351 | |
| 1352 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1353 | name = "f32_dwconv_bench", |
| 1354 | srcs = [ |
| 1355 | "bench/f32-dwconv.cc", |
| 1356 | "bench/dwconv.h", |
| 1357 | "src/xnnpack/AlignedAllocator.h", |
| 1358 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1359 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 1360 | ) |
| 1361 | |
| 1362 | xnnpack_benchmark( |
| 1363 | name = "f32_dwconv_spchw_bench", |
| 1364 | srcs = [ |
| 1365 | "bench/f32-dwconv-spchw.cc", |
| 1366 | "bench/dwconv.h", |
| 1367 | "src/xnnpack/AlignedAllocator.h", |
| 1368 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1369 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 1370 | ) |
| 1371 | |
| 1372 | xnnpack_benchmark( |
| 1373 | name = "f32_gemm_bench", |
| 1374 | srcs = [ |
| 1375 | "bench/f32-gemm.cc", |
| 1376 | "bench/gemm.h", |
| 1377 | "src/xnnpack/AlignedAllocator.h", |
| 1378 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1379 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(), |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1380 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1381 | ) |
| 1382 | |
| 1383 | xnnpack_benchmark( |
| 1384 | name = "f32_rmax_bench", |
| 1385 | srcs = [ |
| 1386 | "bench/f32-rmax.cc", |
| 1387 | "src/xnnpack/AlignedAllocator.h", |
| 1388 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1389 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1390 | ) |
| 1391 | |
| 1392 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 1393 | name = "f32_sigmoid_bench", |
| 1394 | srcs = [ |
| 1395 | "bench/f32-sigmoid.cc", |
| 1396 | "src/xnnpack/AlignedAllocator.h", |
| 1397 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1398 | copts = ["-Wno-unused-function"], |
| 1399 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1400 | ) |
| 1401 | |
| 1402 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1403 | name = "f32_spmm_bench", |
| 1404 | srcs = [ |
| 1405 | "bench/f32-spmm.cc", |
| 1406 | "bench/gemm.h", |
| 1407 | "src/xnnpack/AlignedAllocator.h", |
| 1408 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1409 | copts = ["-Wno-unused-function"], |
| 1410 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1411 | ) |
| 1412 | |
| 1413 | xnnpack_benchmark( |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 1414 | name = "f32_softargmax_bench", |
| 1415 | srcs = [ |
| 1416 | "bench/f32-softargmax.cc", |
| 1417 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1418 | copts = ["-Wno-unused-function"], |
| 1419 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1420 | ) |
| 1421 | |
| 1422 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1423 | name = "f32_im2col_gemm_bench", |
| 1424 | srcs = [ |
| 1425 | "bench/f32-im2col-gemm.cc", |
| 1426 | "bench/conv.h", |
| 1427 | "src/xnnpack/AlignedAllocator.h", |
| 1428 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1429 | deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"], |
| 1430 | ) |
| 1431 | |
| 1432 | ########################### Benchmarks for operators ########################### |
| 1433 | |
| 1434 | xnnpack_benchmark( |
| 1435 | name = "add_bench", |
| 1436 | srcs = ["bench/add.cc"], |
| 1437 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1438 | ) |
| 1439 | |
| 1440 | xnnpack_benchmark( |
| 1441 | name = "average_pooling_bench", |
| 1442 | srcs = ["bench/average-pooling.cc"], |
| 1443 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1444 | ) |
| 1445 | |
| 1446 | xnnpack_benchmark( |
| 1447 | name = "channel_shuffle_bench", |
| 1448 | srcs = ["bench/channel-shuffle.cc"], |
| 1449 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1450 | ) |
| 1451 | |
| 1452 | xnnpack_benchmark( |
| 1453 | name = "convolution_bench", |
| 1454 | srcs = ["bench/convolution.cc"], |
| 1455 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
| 1456 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
| 1457 | ) |
| 1458 | |
| 1459 | xnnpack_benchmark( |
| 1460 | name = "deconvolution_bench", |
| 1461 | srcs = ["bench/deconvolution.cc"], |
| 1462 | copts = xnnpack_optional_tflite_copts(), |
| 1463 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 1464 | ) |
| 1465 | |
| 1466 | xnnpack_benchmark( |
| 1467 | name = "global_average_pooling_bench", |
| 1468 | srcs = ["bench/global-average-pooling.cc"], |
| 1469 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1470 | ) |
| 1471 | |
| 1472 | xnnpack_benchmark( |
| 1473 | name = "max_pooling_bench", |
| 1474 | srcs = ["bench/max-pooling.cc"], |
| 1475 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1476 | ) |
| 1477 | |
| 1478 | xnnpack_benchmark( |
| 1479 | name = "sigmoid_bench", |
| 1480 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 1481 | copts = xnnpack_optional_tflite_copts(), |
| 1482 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1483 | ) |
| 1484 | |
| 1485 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 1486 | name = "prelu_bench", |
| 1487 | srcs = ["bench/prelu.cc"], |
| 1488 | copts = xnnpack_optional_tflite_copts(), |
| 1489 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 1490 | ) |
| 1491 | |
| 1492 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1493 | name = "softargmax_bench", |
| 1494 | srcs = ["bench/softargmax.cc"], |
| 1495 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1496 | ) |
| 1497 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1498 | ############################# End-to-end benchmarks ############################ |
| 1499 | |
| 1500 | cc_library( |
| 1501 | name = "mobilenet_v1", |
| 1502 | srcs = ["models/mobilenet-v1.cc"], |
| 1503 | hdrs = ["models/models.h"], |
| 1504 | linkstatic = True, |
| 1505 | deps = [ |
| 1506 | ":XNNPACK", |
| 1507 | "@pthreadpool", |
| 1508 | ], |
| 1509 | ) |
| 1510 | |
| 1511 | cc_library( |
| 1512 | name = "mobilenet_v2", |
| 1513 | srcs = ["models/mobilenet-v2.cc"], |
| 1514 | hdrs = ["models/models.h"], |
| 1515 | linkstatic = True, |
| 1516 | deps = [ |
| 1517 | ":XNNPACK", |
| 1518 | "@pthreadpool", |
| 1519 | ], |
| 1520 | ) |
| 1521 | |
| 1522 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 1523 | name = "f32_dwconv_e2e_bench", |
| 1524 | srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS, |
| 1525 | copts = ["-Wno-unused-function"], |
| 1526 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 1527 | ":XNNPACK", |
| 1528 | ":mobilenet_v1", |
| 1529 | ":mobilenet_v2", |
| 1530 | ], |
| 1531 | ) |
| 1532 | |
| 1533 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 1534 | name = "f32_gemm_e2e_bench", |
| 1535 | srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS, |
| 1536 | copts = ["-Wno-unused-function"], |
| 1537 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 1538 | ":XNNPACK", |
| 1539 | ":mobilenet_v1", |
| 1540 | ":mobilenet_v2", |
| 1541 | ], |
| 1542 | ) |
| 1543 | |
| 1544 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1545 | name = "end2end_bench", |
| 1546 | srcs = ["bench/end2end.cc"], |
| 1547 | deps = [ |
| 1548 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 1549 | ":bench_utils", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1550 | ":mobilenet_v1", |
| 1551 | ":mobilenet_v2", |
| 1552 | "@pthreadpool", |
| 1553 | ], |
| 1554 | ) |
| 1555 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1556 | #################### Accuracy evaluation for math functions #################### |
| 1557 | |
| 1558 | xnnpack_benchmark( |
| 1559 | name = "f32_exp_eval", |
| 1560 | srcs = [ |
| 1561 | "eval/f32-exp.cc", |
| 1562 | "src/xnnpack/AlignedAllocator.h", |
| 1563 | ] + ACCURACY_EVAL_HDRS, |
| 1564 | deps = ACCURACY_EVAL_DEPS, |
| 1565 | ) |
| 1566 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 1567 | xnnpack_benchmark( |
| 1568 | name = "f32_expminus_eval", |
| 1569 | srcs = [ |
| 1570 | "eval/f32-expminus.cc", |
| 1571 | "src/xnnpack/AlignedAllocator.h", |
| 1572 | ] + ACCURACY_EVAL_HDRS, |
| 1573 | deps = ACCURACY_EVAL_DEPS, |
| 1574 | ) |
| 1575 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1576 | xnnpack_benchmark( |
| 1577 | name = "f32_extexp_eval", |
| 1578 | srcs = [ |
| 1579 | "eval/f32-extexp.cc", |
| 1580 | "src/xnnpack/AlignedAllocator.h", |
| 1581 | ] + ACCURACY_EVAL_HDRS, |
| 1582 | deps = ACCURACY_EVAL_DEPS, |
| 1583 | ) |
| 1584 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1585 | xnnpack_benchmark( |
| 1586 | name = "f32_sigmoid_eval", |
| 1587 | srcs = [ |
| 1588 | "eval/f32-sigmoid.cc", |
| 1589 | "src/xnnpack/AlignedAllocator.h", |
| 1590 | ] + ACCURACY_EVAL_HDRS, |
| 1591 | deps = ACCURACY_EVAL_DEPS, |
| 1592 | ) |
| 1593 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1594 | ######################### Unit tests for micro-kernels ######################### |
| 1595 | |
| 1596 | xnnpack_unit_test( |
| 1597 | name = "f16_gemm_test", |
| 1598 | srcs = [ |
| 1599 | "test/f16-gemm.cc", |
| 1600 | "test/gemm-microkernel-tester.h", |
| 1601 | "src/xnnpack/AlignedAllocator.h", |
| 1602 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1603 | deps = MICROKERNEL_TEST_DEPS, |
| 1604 | ) |
| 1605 | |
| 1606 | xnnpack_unit_test( |
| 1607 | name = "f32_argmaxpool_test", |
| 1608 | srcs = [ |
| 1609 | "test/f32-argmaxpool.cc", |
| 1610 | "test/argmaxpool-microkernel-tester.h", |
| 1611 | "src/xnnpack/AlignedAllocator.h", |
| 1612 | ] + MICROKERNEL_TEST_HDRS, |
| 1613 | deps = MICROKERNEL_TEST_DEPS, |
| 1614 | ) |
| 1615 | |
| 1616 | xnnpack_unit_test( |
| 1617 | name = "f32_avgpool_test", |
| 1618 | srcs = [ |
| 1619 | "test/f32-avgpool.cc", |
| 1620 | "test/avgpool-microkernel-tester.h", |
| 1621 | "src/xnnpack/AlignedAllocator.h", |
| 1622 | ] + MICROKERNEL_TEST_HDRS, |
| 1623 | deps = MICROKERNEL_TEST_DEPS, |
| 1624 | ) |
| 1625 | |
| 1626 | xnnpack_unit_test( |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 1627 | name = "f32_bilinear_test", |
| 1628 | srcs = [ |
| 1629 | "test/f32-bilinear.cc", |
| 1630 | "test/bilinear-microkernel-tester.h", |
| 1631 | "src/xnnpack/AlignedAllocator.h", |
| 1632 | ] + MICROKERNEL_TEST_HDRS, |
| 1633 | deps = MICROKERNEL_TEST_DEPS, |
| 1634 | ) |
| 1635 | |
| 1636 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1637 | name = "f32_clamp_test", |
| 1638 | srcs = [ |
| 1639 | "test/f32-clamp.cc", |
| 1640 | "test/clamp-microkernel-tester.h", |
| 1641 | ] + MICROKERNEL_TEST_HDRS, |
| 1642 | deps = MICROKERNEL_TEST_DEPS, |
| 1643 | ) |
| 1644 | |
| 1645 | xnnpack_unit_test( |
| 1646 | name = "f32_igemm_test", |
| 1647 | srcs = [ |
| 1648 | "test/f32-igemm.cc", |
| 1649 | "test/gemm-microkernel-tester.h", |
| 1650 | "src/xnnpack/AlignedAllocator.h", |
| 1651 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1652 | deps = MICROKERNEL_TEST_DEPS, |
| 1653 | ) |
| 1654 | |
| 1655 | xnnpack_unit_test( |
| 1656 | name = "f32_conv_hwc_test", |
| 1657 | srcs = [ |
| 1658 | "test/f32-conv-hwc.cc", |
| 1659 | "test/conv-hwc-microkernel-tester.h", |
| 1660 | "src/xnnpack/AlignedAllocator.h", |
| 1661 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1662 | deps = MICROKERNEL_TEST_DEPS, |
| 1663 | ) |
| 1664 | |
| 1665 | xnnpack_unit_test( |
| 1666 | name = "f32_conv_hwc2spchw_test", |
| 1667 | srcs = [ |
| 1668 | "test/f32-conv-hwc2spchw.cc", |
| 1669 | "test/conv-hwc2spchw-microkernel-tester.h", |
| 1670 | "src/xnnpack/AlignedAllocator.h", |
| 1671 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1672 | deps = MICROKERNEL_TEST_DEPS, |
| 1673 | ) |
| 1674 | |
| 1675 | xnnpack_unit_test( |
| 1676 | name = "f32_dwconv_test", |
| 1677 | srcs = [ |
| 1678 | "test/f32-dwconv.cc", |
| 1679 | "test/dwconv-microkernel-tester.h", |
| 1680 | "src/xnnpack/AlignedAllocator.h", |
| 1681 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1682 | deps = MICROKERNEL_TEST_DEPS, |
| 1683 | ) |
| 1684 | |
| 1685 | xnnpack_unit_test( |
| 1686 | name = "f32_dwconv_spchw_test", |
| 1687 | srcs = [ |
| 1688 | "test/f32-dwconv-spchw.cc", |
| 1689 | "test/dwconv-spchw-microkernel-tester.h", |
| 1690 | "src/xnnpack/AlignedAllocator.h", |
| 1691 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1692 | deps = MICROKERNEL_TEST_DEPS, |
| 1693 | ) |
| 1694 | |
| 1695 | xnnpack_unit_test( |
| 1696 | name = "f32_gavgpool_test", |
| 1697 | srcs = [ |
| 1698 | "test/f32-gavgpool.cc", |
| 1699 | "test/gavgpool-microkernel-tester.h", |
| 1700 | "src/xnnpack/AlignedAllocator.h", |
| 1701 | ] + MICROKERNEL_TEST_HDRS, |
| 1702 | deps = MICROKERNEL_TEST_DEPS, |
| 1703 | ) |
| 1704 | |
| 1705 | xnnpack_unit_test( |
| 1706 | name = "f32_gavgpool_spchw_test", |
| 1707 | srcs = [ |
| 1708 | "test/f32-gavgpool-spchw.cc", |
| 1709 | "test/gavgpool-spchw-microkernel-tester.h", |
| 1710 | "src/xnnpack/AlignedAllocator.h", |
| 1711 | ] + MICROKERNEL_TEST_HDRS, |
| 1712 | deps = MICROKERNEL_TEST_DEPS, |
| 1713 | ) |
| 1714 | |
| 1715 | xnnpack_unit_test( |
| 1716 | name = "f32_gemm_test", |
| 1717 | srcs = [ |
| 1718 | "test/f32-gemm.cc", |
| 1719 | "test/gemm-microkernel-tester.h", |
| 1720 | "src/xnnpack/AlignedAllocator.h", |
| 1721 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1722 | deps = MICROKERNEL_TEST_DEPS, |
| 1723 | ) |
| 1724 | |
| 1725 | xnnpack_unit_test( |
| 1726 | name = "f32_gemminc_test", |
| 1727 | srcs = [ |
| 1728 | "test/f32-gemminc.cc", |
| 1729 | "test/gemm-microkernel-tester.h", |
| 1730 | "src/xnnpack/AlignedAllocator.h", |
| 1731 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1732 | deps = MICROKERNEL_TEST_DEPS, |
| 1733 | ) |
| 1734 | |
| 1735 | xnnpack_unit_test( |
| 1736 | name = "f32_hswish_test", |
| 1737 | srcs = [ |
| 1738 | "test/f32-hswish.cc", |
| 1739 | "test/hswish-microkernel-tester.h", |
| 1740 | ] + MICROKERNEL_TEST_HDRS, |
| 1741 | deps = MICROKERNEL_TEST_DEPS, |
| 1742 | ) |
| 1743 | |
| 1744 | xnnpack_unit_test( |
| 1745 | name = "f32_maxpool_test", |
| 1746 | srcs = [ |
| 1747 | "test/f32-maxpool.cc", |
| 1748 | "test/maxpool-microkernel-tester.h", |
| 1749 | ] + MICROKERNEL_TEST_HDRS, |
| 1750 | deps = MICROKERNEL_TEST_DEPS, |
| 1751 | ) |
| 1752 | |
| 1753 | xnnpack_unit_test( |
| 1754 | name = "f32_pavgpool_test", |
| 1755 | srcs = [ |
| 1756 | "test/f32-pavgpool.cc", |
| 1757 | "test/avgpool-microkernel-tester.h", |
| 1758 | "src/xnnpack/AlignedAllocator.h", |
| 1759 | ] + MICROKERNEL_TEST_HDRS, |
| 1760 | deps = MICROKERNEL_TEST_DEPS, |
| 1761 | ) |
| 1762 | |
| 1763 | xnnpack_unit_test( |
| 1764 | name = "f32_ppmm_test", |
| 1765 | srcs = [ |
| 1766 | "test/f32-ppmm.cc", |
| 1767 | "test/gemm-microkernel-tester.h", |
| 1768 | "src/xnnpack/AlignedAllocator.h", |
| 1769 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1770 | deps = MICROKERNEL_TEST_DEPS, |
| 1771 | ) |
| 1772 | |
| 1773 | xnnpack_unit_test( |
| 1774 | name = "f32_prelu_test", |
| 1775 | srcs = [ |
| 1776 | "test/f32-prelu.cc", |
| 1777 | "test/prelu-microkernel-tester.h", |
| 1778 | "src/xnnpack/AlignedAllocator.h", |
| 1779 | ] + MICROKERNEL_TEST_HDRS, |
| 1780 | deps = MICROKERNEL_TEST_DEPS, |
| 1781 | ) |
| 1782 | |
| 1783 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1784 | name = "f32_raddexpminusmax_test", |
| 1785 | srcs = [ |
| 1786 | "test/f32-raddexpminusmax.cc", |
| 1787 | "test/raddexpminusmax-microkernel-tester.h", |
| 1788 | ] + MICROKERNEL_TEST_HDRS, |
| 1789 | deps = MICROKERNEL_TEST_DEPS, |
| 1790 | ) |
| 1791 | |
| 1792 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1793 | name = "f32_raddextexp_test", |
| 1794 | srcs = [ |
| 1795 | "test/f32-raddextexp.cc", |
| 1796 | "test/raddextexp-microkernel-tester.h", |
| 1797 | ] + MICROKERNEL_TEST_HDRS, |
| 1798 | deps = MICROKERNEL_TEST_DEPS, |
| 1799 | ) |
| 1800 | |
| 1801 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1802 | name = "f32_raddstoreexpminusmax_test", |
| 1803 | srcs = [ |
| 1804 | "test/f32-raddstoreexpminusmax.cc", |
| 1805 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 1806 | ] + MICROKERNEL_TEST_HDRS, |
| 1807 | deps = MICROKERNEL_TEST_DEPS, |
| 1808 | ) |
| 1809 | |
| 1810 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1811 | name = "f32_rmax_test", |
| 1812 | srcs = [ |
| 1813 | "test/f32-rmax.cc", |
| 1814 | "test/rmax-microkernel-tester.h", |
| 1815 | ] + MICROKERNEL_TEST_HDRS, |
| 1816 | deps = MICROKERNEL_TEST_DEPS, |
| 1817 | ) |
| 1818 | |
| 1819 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1820 | name = "f32_sigmoid_test", |
| 1821 | srcs = [ |
| 1822 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1823 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1824 | ] + MICROKERNEL_TEST_HDRS, |
| 1825 | deps = MICROKERNEL_TEST_DEPS, |
| 1826 | ) |
| 1827 | |
| 1828 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1829 | name = "f32_spmm_test", |
| 1830 | srcs = [ |
| 1831 | "test/f32-spmm.cc", |
| 1832 | "test/spmm-microkernel-tester.h", |
| 1833 | "src/xnnpack/AlignedAllocator.h", |
| 1834 | ] + MICROKERNEL_TEST_HDRS, |
| 1835 | deps = MICROKERNEL_TEST_DEPS, |
| 1836 | ) |
| 1837 | |
| 1838 | xnnpack_unit_test( |
| 1839 | name = "f32_vadd_test", |
| 1840 | srcs = [ |
| 1841 | "test/f32-vadd.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1842 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 1843 | ] + MICROKERNEL_TEST_HDRS, |
| 1844 | deps = MICROKERNEL_TEST_DEPS, |
| 1845 | ) |
| 1846 | |
| 1847 | xnnpack_unit_test( |
| 1848 | name = "f32_vaddc_test", |
| 1849 | srcs = [ |
| 1850 | "test/f32-vaddc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1851 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1852 | ] + MICROKERNEL_TEST_HDRS, |
| 1853 | deps = MICROKERNEL_TEST_DEPS, |
| 1854 | ) |
| 1855 | |
| 1856 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1857 | name = "f32_vmul_test", |
| 1858 | srcs = [ |
| 1859 | "test/f32-vmul.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1860 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 1861 | ] + MICROKERNEL_TEST_HDRS, |
| 1862 | deps = MICROKERNEL_TEST_DEPS, |
| 1863 | ) |
| 1864 | |
| 1865 | xnnpack_unit_test( |
| 1866 | name = "f32_vmulc_test", |
| 1867 | srcs = [ |
| 1868 | "test/f32-vmulc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1869 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1870 | ] + MICROKERNEL_TEST_HDRS, |
| 1871 | deps = MICROKERNEL_TEST_DEPS, |
| 1872 | ) |
| 1873 | |
| 1874 | xnnpack_unit_test( |
| 1875 | name = "f32_vmulcaddc_test", |
| 1876 | srcs = [ |
| 1877 | "test/f32-vmulcaddc.cc", |
| 1878 | "test/vmulcaddc-microkernel-tester.h", |
| 1879 | "src/xnnpack/AlignedAllocator.h", |
| 1880 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1881 | deps = MICROKERNEL_TEST_DEPS, |
| 1882 | ) |
| 1883 | |
| 1884 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1885 | name = "f32_vscale_test", |
| 1886 | srcs = [ |
| 1887 | "test/f32-vscale.cc", |
| 1888 | "test/vscale-microkernel-tester.h", |
| 1889 | ] + MICROKERNEL_TEST_HDRS, |
| 1890 | deps = MICROKERNEL_TEST_DEPS, |
| 1891 | ) |
| 1892 | |
| 1893 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1894 | name = "f32_vscaleexpminusmax_test", |
| 1895 | srcs = [ |
| 1896 | "test/f32-vscaleexpminusmax.cc", |
| 1897 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 1898 | ] + MICROKERNEL_TEST_HDRS, |
| 1899 | deps = MICROKERNEL_TEST_DEPS, |
| 1900 | ) |
| 1901 | |
| 1902 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1903 | name = "f32_vscaleextexp_test", |
| 1904 | srcs = [ |
| 1905 | "test/f32-vscaleextexp.cc", |
| 1906 | "test/vscaleextexp-microkernel-tester.h", |
| 1907 | ] + MICROKERNEL_TEST_HDRS, |
| 1908 | deps = MICROKERNEL_TEST_DEPS, |
| 1909 | ) |
| 1910 | |
| 1911 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1912 | name = "f32_vsub_test", |
| 1913 | srcs = [ |
| 1914 | "test/f32-vsub.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1915 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 1916 | ] + MICROKERNEL_TEST_HDRS, |
| 1917 | deps = MICROKERNEL_TEST_DEPS, |
| 1918 | ) |
| 1919 | |
| 1920 | xnnpack_unit_test( |
| 1921 | name = "f32_vsubc_test", |
| 1922 | srcs = [ |
| 1923 | "test/f32-vsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1924 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 1925 | ] + MICROKERNEL_TEST_HDRS, |
| 1926 | deps = MICROKERNEL_TEST_DEPS, |
| 1927 | ) |
| 1928 | |
| 1929 | xnnpack_unit_test( |
| 1930 | name = "f32_vrsubc_test", |
| 1931 | srcs = [ |
| 1932 | "test/f32-vrsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1933 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1934 | ] + MICROKERNEL_TEST_HDRS, |
| 1935 | deps = MICROKERNEL_TEST_DEPS, |
| 1936 | ) |
| 1937 | |
| 1938 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1939 | name = "q8_avgpool_test", |
| 1940 | srcs = [ |
| 1941 | "test/q8-avgpool.cc", |
| 1942 | "test/avgpool-microkernel-tester.h", |
| 1943 | "src/xnnpack/AlignedAllocator.h", |
| 1944 | ] + MICROKERNEL_TEST_HDRS, |
| 1945 | deps = MICROKERNEL_TEST_DEPS, |
| 1946 | ) |
| 1947 | |
| 1948 | xnnpack_unit_test( |
| 1949 | name = "q8_igemm_test", |
| 1950 | srcs = [ |
| 1951 | "test/q8-igemm.cc", |
| 1952 | "test/gemm-microkernel-tester.h", |
| 1953 | "src/xnnpack/AlignedAllocator.h", |
| 1954 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1955 | deps = MICROKERNEL_TEST_DEPS, |
| 1956 | ) |
| 1957 | |
| 1958 | xnnpack_unit_test( |
| 1959 | name = "q8_dwconv_test", |
| 1960 | srcs = [ |
| 1961 | "test/q8-dwconv.cc", |
| 1962 | "test/dwconv-microkernel-tester.h", |
| 1963 | "src/xnnpack/AlignedAllocator.h", |
| 1964 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1965 | deps = MICROKERNEL_TEST_DEPS, |
| 1966 | ) |
| 1967 | |
| 1968 | xnnpack_unit_test( |
| 1969 | name = "q8_gavgpool_test", |
| 1970 | srcs = [ |
| 1971 | "test/q8-gavgpool.cc", |
| 1972 | "test/gavgpool-microkernel-tester.h", |
| 1973 | "src/xnnpack/AlignedAllocator.h", |
| 1974 | ] + MICROKERNEL_TEST_HDRS, |
| 1975 | deps = MICROKERNEL_TEST_DEPS, |
| 1976 | ) |
| 1977 | |
| 1978 | xnnpack_unit_test( |
| 1979 | name = "q8_gemm_test", |
| 1980 | srcs = [ |
| 1981 | "test/q8-gemm.cc", |
| 1982 | "test/gemm-microkernel-tester.h", |
| 1983 | "src/xnnpack/AlignedAllocator.h", |
| 1984 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1985 | deps = MICROKERNEL_TEST_DEPS, |
| 1986 | ) |
| 1987 | |
| 1988 | xnnpack_unit_test( |
| 1989 | name = "q8_vadd_test", |
| 1990 | srcs = [ |
| 1991 | "test/q8-vadd.cc", |
| 1992 | "test/vadd-microkernel-tester.h", |
| 1993 | ] + MICROKERNEL_TEST_HDRS, |
| 1994 | deps = MICROKERNEL_TEST_DEPS, |
| 1995 | ) |
| 1996 | |
| 1997 | xnnpack_unit_test( |
| 1998 | name = "u8_clamp_test", |
| 1999 | srcs = [ |
| 2000 | "test/u8-clamp.cc", |
| 2001 | "test/clamp-microkernel-tester.h", |
| 2002 | ] + MICROKERNEL_TEST_HDRS, |
| 2003 | deps = MICROKERNEL_TEST_DEPS, |
| 2004 | ) |
| 2005 | |
| 2006 | xnnpack_unit_test( |
| 2007 | name = "u8_lut32norm_test", |
| 2008 | srcs = [ |
| 2009 | "test/u8-lut32norm.cc", |
| 2010 | "test/lut-norm-microkernel-tester.h", |
| 2011 | ] + MICROKERNEL_TEST_HDRS, |
| 2012 | deps = MICROKERNEL_TEST_DEPS, |
| 2013 | ) |
| 2014 | |
| 2015 | xnnpack_unit_test( |
| 2016 | name = "u8_maxpool_test", |
| 2017 | srcs = [ |
| 2018 | "test/u8-maxpool.cc", |
| 2019 | "test/maxpool-microkernel-tester.h", |
| 2020 | ] + MICROKERNEL_TEST_HDRS, |
| 2021 | deps = MICROKERNEL_TEST_DEPS, |
| 2022 | ) |
| 2023 | |
| 2024 | xnnpack_unit_test( |
| 2025 | name = "u8_rmax_test", |
| 2026 | srcs = [ |
| 2027 | "test/u8-rmax.cc", |
| 2028 | "test/rmax-microkernel-tester.h", |
| 2029 | ] + MICROKERNEL_TEST_HDRS, |
| 2030 | deps = MICROKERNEL_TEST_DEPS, |
| 2031 | ) |
| 2032 | |
| 2033 | xnnpack_unit_test( |
| 2034 | name = "x32_packx_test", |
| 2035 | srcs = [ |
| 2036 | "test/x32-packx.cc", |
| 2037 | "test/pack-microkernel-tester.h", |
| 2038 | "src/xnnpack/AlignedAllocator.h", |
| 2039 | ] + MICROKERNEL_TEST_HDRS, |
| 2040 | deps = MICROKERNEL_TEST_DEPS, |
| 2041 | ) |
| 2042 | |
| 2043 | xnnpack_unit_test( |
| 2044 | name = "x32_pad_test", |
| 2045 | srcs = [ |
| 2046 | "test/x32-pad.cc", |
| 2047 | "test/pad-microkernel-tester.h", |
| 2048 | ] + MICROKERNEL_TEST_HDRS, |
| 2049 | deps = MICROKERNEL_TEST_DEPS, |
| 2050 | ) |
| 2051 | |
| 2052 | xnnpack_unit_test( |
| 2053 | name = "x32_unpool_test", |
| 2054 | srcs = [ |
| 2055 | "test/x32-unpool.cc", |
| 2056 | "test/unpool-microkernel-tester.h", |
| 2057 | ] + MICROKERNEL_TEST_HDRS, |
| 2058 | deps = MICROKERNEL_TEST_DEPS, |
| 2059 | ) |
| 2060 | |
| 2061 | xnnpack_unit_test( |
| 2062 | name = "x32_zip_test", |
| 2063 | srcs = [ |
| 2064 | "test/x32-zip.cc", |
| 2065 | "test/zip-microkernel-tester.h", |
| 2066 | ] + MICROKERNEL_TEST_HDRS, |
| 2067 | deps = MICROKERNEL_TEST_DEPS, |
| 2068 | ) |
| 2069 | |
| 2070 | xnnpack_unit_test( |
| 2071 | name = "x8_lut_test", |
| 2072 | srcs = [ |
| 2073 | "test/x8-lut.cc", |
| 2074 | "test/lut-microkernel-tester.h", |
| 2075 | ] + MICROKERNEL_TEST_HDRS, |
| 2076 | deps = MICROKERNEL_TEST_DEPS, |
| 2077 | ) |
| 2078 | |
| 2079 | xnnpack_unit_test( |
| 2080 | name = "x8_zip_test", |
| 2081 | srcs = [ |
| 2082 | "test/x8-zip.cc", |
| 2083 | "test/zip-microkernel-tester.h", |
| 2084 | ] + MICROKERNEL_TEST_HDRS, |
| 2085 | deps = MICROKERNEL_TEST_DEPS, |
| 2086 | ) |
| 2087 | |
| 2088 | ########################### Size test for the library ########################## |
| 2089 | |
| 2090 | xnnpack_binary( |
| 2091 | name = "size_test", |
| 2092 | srcs = ["test/size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2093 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2094 | ) |
| 2095 | |
| 2096 | ########################### Unit tests for operators ########################### |
| 2097 | |
| 2098 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2099 | name = "add_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2100 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2101 | "test/add-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2102 | "test/add-operator-tester.h", |
| 2103 | ], |
| 2104 | deps = OPERATOR_TEST_DEPS, |
| 2105 | ) |
| 2106 | |
| 2107 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2108 | name = "add_nd_test", |
| 2109 | srcs = [ |
| 2110 | "test/add-nd.cc", |
| 2111 | "test/binary-elementwise-operator-tester.h", |
| 2112 | ], |
| 2113 | deps = OPERATOR_TEST_DEPS, |
| 2114 | ) |
| 2115 | |
| 2116 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2117 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2118 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2119 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2120 | "test/argmax-pooling-operator-tester.h", |
| 2121 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2122 | deps = OPERATOR_TEST_DEPS, |
| 2123 | ) |
| 2124 | |
| 2125 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2126 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2127 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2128 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2129 | "test/average-pooling-operator-tester.h", |
| 2130 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2131 | deps = OPERATOR_TEST_DEPS, |
| 2132 | ) |
| 2133 | |
| 2134 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2135 | name = "channel_pad_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2136 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2137 | "test/channel-pad-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2138 | "test/channel-pad-operator-tester.h", |
| 2139 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2140 | deps = OPERATOR_TEST_DEPS, |
| 2141 | ) |
| 2142 | |
| 2143 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2144 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2145 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2146 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2147 | "test/channel-shuffle-operator-tester.h", |
| 2148 | ], |
| 2149 | deps = OPERATOR_TEST_DEPS, |
| 2150 | ) |
| 2151 | |
| 2152 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2153 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2154 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2155 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2156 | "test/clamp-operator-tester.h", |
| 2157 | ], |
| 2158 | deps = OPERATOR_TEST_DEPS, |
| 2159 | ) |
| 2160 | |
| 2161 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2162 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2163 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2164 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2165 | "test/convolution-operator-tester.h", |
| 2166 | ], |
| 2167 | deps = OPERATOR_TEST_DEPS, |
| 2168 | ) |
| 2169 | |
| 2170 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2171 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2172 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2173 | "test/convolution-nchw.cc", |
| 2174 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2175 | ], |
| 2176 | deps = OPERATOR_TEST_DEPS, |
| 2177 | ) |
| 2178 | |
| 2179 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2180 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2181 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2182 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2183 | "test/deconvolution-operator-tester.h", |
| 2184 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2185 | deps = OPERATOR_TEST_DEPS, |
| 2186 | ) |
| 2187 | |
| 2188 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2189 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2190 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2191 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2192 | "test/fully-connected-operator-tester.h", |
| 2193 | ], |
| 2194 | deps = OPERATOR_TEST_DEPS, |
| 2195 | ) |
| 2196 | |
| 2197 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2198 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2199 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2200 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2201 | "test/global-average-pooling-operator-tester.h", |
| 2202 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2203 | deps = OPERATOR_TEST_DEPS, |
| 2204 | ) |
| 2205 | |
| 2206 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2207 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2208 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2209 | "test/global-average-pooling-ncw.cc", |
| 2210 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2211 | ], |
| 2212 | deps = OPERATOR_TEST_DEPS, |
| 2213 | ) |
| 2214 | |
| 2215 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2216 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2217 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2218 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2219 | "test/hardswish-operator-tester.h", |
| 2220 | ], |
| 2221 | deps = OPERATOR_TEST_DEPS, |
| 2222 | ) |
| 2223 | |
| 2224 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2225 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2226 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2227 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2228 | "test/leaky-relu-operator-tester.h", |
| 2229 | ], |
| 2230 | deps = OPERATOR_TEST_DEPS, |
| 2231 | ) |
| 2232 | |
| 2233 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2234 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2235 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2236 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2237 | "test/max-pooling-operator-tester.h", |
| 2238 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2239 | deps = OPERATOR_TEST_DEPS, |
| 2240 | ) |
| 2241 | |
| 2242 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2243 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 2244 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2245 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2246 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 2247 | ], |
| 2248 | deps = OPERATOR_TEST_DEPS, |
| 2249 | ) |
| 2250 | |
| 2251 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2252 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2253 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2254 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2255 | "test/prelu-operator-tester.h", |
| 2256 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2257 | deps = OPERATOR_TEST_DEPS, |
| 2258 | ) |
| 2259 | |
| 2260 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2261 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 2262 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2263 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 2264 | "test/resize-bilinear-operator-tester.h", |
| 2265 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2266 | deps = OPERATOR_TEST_DEPS, |
| 2267 | ) |
| 2268 | |
| 2269 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2270 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2271 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2272 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2273 | "test/sigmoid-operator-tester.h", |
| 2274 | ], |
| 2275 | deps = OPERATOR_TEST_DEPS, |
| 2276 | ) |
| 2277 | |
| 2278 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2279 | name = "softargmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2280 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2281 | "test/softargmax-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2282 | "test/softargmax-operator-tester.h", |
| 2283 | ], |
| 2284 | deps = OPERATOR_TEST_DEPS, |
| 2285 | ) |
| 2286 | |
| 2287 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 2288 | name = "subtract_nd_test", |
| 2289 | srcs = [ |
| 2290 | "test/binary-elementwise-operator-tester.h", |
| 2291 | "test/subtract-nd.cc", |
| 2292 | ], |
| 2293 | deps = OPERATOR_TEST_DEPS, |
| 2294 | ) |
| 2295 | |
| 2296 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2297 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2298 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2299 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2300 | "test/unpooling-operator-tester.h", |
| 2301 | ], |
| 2302 | deps = OPERATOR_TEST_DEPS, |
| 2303 | ) |
| 2304 | |
| 2305 | ############################# Build configurations ############################# |
| 2306 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2307 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2308 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2309 | name = "xnn_enable_assembly_explicit_true", |
| 2310 | define_values = {"xnn_enable_assembly": "true"}, |
| 2311 | ) |
| 2312 | |
| 2313 | # Disables usage of assembly kernels. |
| 2314 | config_setting( |
| 2315 | name = "xnn_enable_assembly_explicit_false", |
| 2316 | define_values = {"xnn_enable_assembly": "false"}, |
| 2317 | ) |
| 2318 | |
| 2319 | # Builds with -c dbg |
| 2320 | config_setting( |
| 2321 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2322 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2323 | "compilation_mode": "dbg", |
| 2324 | }, |
| 2325 | ) |
| 2326 | |
| 2327 | # Builds with -c opt |
| 2328 | config_setting( |
| 2329 | name = "optimized_build", |
| 2330 | values = { |
| 2331 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2332 | }, |
| 2333 | ) |
| 2334 | |
| 2335 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2336 | name = "linux_k8", |
| 2337 | values = {"cpu": "k8"}, |
| 2338 | ) |
| 2339 | |
| 2340 | config_setting( |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 2341 | name = "linux_aarch64", |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2342 | values = {"cpu": "aarch64"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 2343 | ) |
| 2344 | |
| 2345 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2346 | name = "android", |
| 2347 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 2348 | ) |
| 2349 | |
| 2350 | config_setting( |
| 2351 | name = "android_armv7", |
| 2352 | values = { |
| 2353 | "crosstool_top": "//external:android/crosstool", |
| 2354 | "cpu": "armeabi-v7a", |
| 2355 | }, |
| 2356 | ) |
| 2357 | |
| 2358 | config_setting( |
| 2359 | name = "android_arm64", |
| 2360 | values = { |
| 2361 | "crosstool_top": "//external:android/crosstool", |
| 2362 | "cpu": "arm64-v8a", |
| 2363 | }, |
| 2364 | ) |
| 2365 | |
| 2366 | config_setting( |
| 2367 | name = "android_x86", |
| 2368 | values = { |
| 2369 | "crosstool_top": "//external:android/crosstool", |
| 2370 | "cpu": "x86", |
| 2371 | }, |
| 2372 | ) |
| 2373 | |
| 2374 | config_setting( |
| 2375 | name = "android_x86_64", |
| 2376 | values = { |
| 2377 | "crosstool_top": "//external:android/crosstool", |
| 2378 | "cpu": "x86_64", |
| 2379 | }, |
| 2380 | ) |
| 2381 | |
| 2382 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 2383 | name = "macos_x86_64", |
| 2384 | values = { |
| 2385 | "apple_platform_type": "macos", |
| 2386 | "cpu": "darwin", |
| 2387 | }, |
| 2388 | ) |
| 2389 | |
| 2390 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2391 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2392 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2393 | ) |
| 2394 | |
| 2395 | config_setting( |
| 2396 | name = "emscripten_wasm", |
| 2397 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2398 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2399 | "cpu": "wasm", |
| 2400 | }, |
| 2401 | ) |
| 2402 | |
| 2403 | config_setting( |
| 2404 | name = "emscripten_wasmsimd", |
| 2405 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2406 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2407 | "cpu": "wasm", |
Marat Dukhan | 8c19e3c | 2019-10-30 12:14:58 -0700 | [diff] [blame] | 2408 | "features": "wasm_simd", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2409 | }, |
| 2410 | ) |
| 2411 | |
| 2412 | config_setting( |
| 2413 | name = "emscripten_asmjs", |
| 2414 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2415 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2416 | "cpu": "asmjs", |
| 2417 | }, |
| 2418 | ) |