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", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 146 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 147 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 148 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
| 149 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 150 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 151 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 152 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 153 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 154 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| 155 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 156 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 157 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| 158 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 159 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 160 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| 161 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 162 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 163 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 164 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 165 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 166 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
| 167 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 168 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 169 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 170 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 171 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 172 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 173 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 174 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 175 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
| 176 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 177 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 178 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
| 179 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 180 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 181 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
| 182 | "src/f32-vmulcaddc/gen/c1-scalar-2x.c", |
| 183 | "src/f32-vmulcaddc/gen/c2-scalar-2x.c", |
| 184 | "src/f32-vmulcaddc/gen/c4-scalar-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 185 | "src/q8-avgpool/mp9p8q-scalar.c", |
| 186 | "src/q8-avgpool/up9-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 187 | "src/q8-dwconv/up1x9-scalar.c", |
| 188 | "src/q8-gavgpool/mp7p7q-scalar.c", |
| 189 | "src/q8-gavgpool/up7-scalar.c", |
| 190 | "src/q8-gemm/2x2-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 191 | "src/q8-igemm/2x2-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 192 | "src/q8-vadd/scalar.c", |
| 193 | "src/u8-clamp/scalar.c", |
| 194 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 195 | "src/u8-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 196 | "src/u8-rmax/scalar.c", |
| 197 | "src/x32-packx/x2-scalar.c", |
| 198 | "src/x32-packx/x3-scalar.c", |
| 199 | "src/x32-packx/x4-scalar.c", |
| 200 | "src/x32-pad/x2-scalar.c", |
| 201 | "src/x32-unpool/scalar.c", |
| 202 | "src/x32-zip/x2-scalar.c", |
| 203 | "src/x32-zip/x3-scalar.c", |
| 204 | "src/x32-zip/x4-scalar.c", |
| 205 | "src/x32-zip/xm-scalar.c", |
| 206 | "src/x8-lut/scalar.c", |
| 207 | "src/x8-zip/x2-scalar.c", |
| 208 | "src/x8-zip/x3-scalar.c", |
| 209 | "src/x8-zip/x4-scalar.c", |
| 210 | "src/x8-zip/xm-scalar.c", |
| 211 | ] |
| 212 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 213 | WASM_UKERNELS = [ |
| 214 | "src/f32-avgpool/mp9p8q-wasm.c", |
| 215 | "src/f32-avgpool/up9-wasm.c", |
| 216 | "src/f32-clamp/wasm.c", |
| 217 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 218 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 219 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 220 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 221 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 222 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 223 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 224 | "src/f32-dwconv/gen/up2x25-wasm.c", |
| 225 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 226 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 227 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 228 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| 229 | "src/f32-gavgpool/mp7p7q-wasm.c", |
| 230 | "src/f32-gavgpool/up7-wasm.c", |
| 231 | "src/f32-gemm/gen-inc/1x4-wasm.c", |
| 232 | "src/f32-gemm/gen-inc/2x4-wasm.c", |
| 233 | "src/f32-gemm/gen-inc/4x4-wasm.c", |
| 234 | "src/f32-gemm/gen/1x4-wasm.c", |
| 235 | "src/f32-gemm/gen/2x4-wasm.c", |
| 236 | "src/f32-gemm/gen/4x2-wasm.c", |
| 237 | "src/f32-gemm/gen/4x4-wasm.c", |
| 238 | "src/f32-hswish/wasm.c", |
| 239 | "src/f32-igemm/gen/1x4-wasm.c", |
| 240 | "src/f32-igemm/gen/2x4-wasm.c", |
| 241 | "src/f32-igemm/gen/4x2-wasm.c", |
| 242 | "src/f32-igemm/gen/4x4-wasm.c", |
| 243 | "src/f32-maxpool/9p8x-wasm-c1.c", |
| 244 | "src/f32-pavgpool/mp9p8q-wasm.c", |
| 245 | "src/f32-pavgpool/up9-wasm.c", |
| 246 | "src/f32-prelu/gen/wasm-2x1.c", |
| 247 | "src/f32-prelu/gen/wasm-2x4.c", |
| 248 | "src/f32-vbinary/gen/vadd-wasm-x1.c", |
| 249 | "src/f32-vbinary/gen/vadd-wasm-x2.c", |
| 250 | "src/f32-vbinary/gen/vadd-wasm-x4.c", |
| 251 | "src/f32-vbinary/gen/vaddc-wasm-x1.c", |
| 252 | "src/f32-vbinary/gen/vaddc-wasm-x2.c", |
| 253 | "src/f32-vbinary/gen/vaddc-wasm-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 254 | "src/f32-vbinary/gen/vdiv-wasm-x1.c", |
| 255 | "src/f32-vbinary/gen/vdiv-wasm-x2.c", |
| 256 | "src/f32-vbinary/gen/vdiv-wasm-x4.c", |
| 257 | "src/f32-vbinary/gen/vdivc-wasm-x1.c", |
| 258 | "src/f32-vbinary/gen/vdivc-wasm-x2.c", |
| 259 | "src/f32-vbinary/gen/vdivc-wasm-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 260 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 261 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 262 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
| 263 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 264 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 265 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
| 266 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 267 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 268 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
| 269 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 270 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 271 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 272 | "src/f32-vbinary/gen/vmul-wasm-x1.c", |
| 273 | "src/f32-vbinary/gen/vmul-wasm-x2.c", |
| 274 | "src/f32-vbinary/gen/vmul-wasm-x4.c", |
| 275 | "src/f32-vbinary/gen/vmulc-wasm-x1.c", |
| 276 | "src/f32-vbinary/gen/vmulc-wasm-x2.c", |
| 277 | "src/f32-vbinary/gen/vmulc-wasm-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 278 | "src/f32-vbinary/gen/vrdivc-wasm-x1.c", |
| 279 | "src/f32-vbinary/gen/vrdivc-wasm-x2.c", |
| 280 | "src/f32-vbinary/gen/vrdivc-wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 281 | "src/f32-vbinary/gen/vrsubc-wasm-x1.c", |
| 282 | "src/f32-vbinary/gen/vrsubc-wasm-x2.c", |
| 283 | "src/f32-vbinary/gen/vrsubc-wasm-x4.c", |
| 284 | "src/f32-vbinary/gen/vsub-wasm-x1.c", |
| 285 | "src/f32-vbinary/gen/vsub-wasm-x2.c", |
| 286 | "src/f32-vbinary/gen/vsub-wasm-x4.c", |
| 287 | "src/f32-vbinary/gen/vsubc-wasm-x1.c", |
| 288 | "src/f32-vbinary/gen/vsubc-wasm-x2.c", |
| 289 | "src/f32-vbinary/gen/vsubc-wasm-x4.c", |
| 290 | "src/f32-vmulcaddc/gen/c1-wasm-2x.c", |
| 291 | "src/f32-vmulcaddc/gen/c2-wasm-2x.c", |
| 292 | "src/f32-vmulcaddc/gen/c4-wasm-2x.c", |
| 293 | ] |
| 294 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 295 | PSIMD_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 296 | "src/f32-argmaxpool/4x-psimd-c4.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 297 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 298 | "src/f32-argmaxpool/9x-psimd-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 299 | "src/f32-avgpool/mp9p8q-psimd.c", |
| 300 | "src/f32-avgpool/up9-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 301 | "src/f32-bilinear/gen/psimd-c4.c", |
| 302 | "src/f32-bilinear/gen/psimd-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 303 | "src/f32-clamp/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 304 | "src/f32-dwconv/gen/up4x25-psimd-acc2.c", |
| 305 | "src/f32-dwconv/gen/up4x25-psimd.c", |
| 306 | "src/f32-dwconv/gen/up4x4-psimd-acc2.c", |
| 307 | "src/f32-dwconv/gen/up4x4-psimd.c", |
| 308 | "src/f32-dwconv/gen/up4x9-psimd-acc2.c", |
| 309 | "src/f32-dwconv/gen/up4x9-psimd.c", |
| 310 | "src/f32-dwconv/gen/up8x25-psimd-acc2.c", |
| 311 | "src/f32-dwconv/gen/up8x25-psimd.c", |
| 312 | "src/f32-dwconv/gen/up8x4-psimd-acc2.c", |
| 313 | "src/f32-dwconv/gen/up8x4-psimd.c", |
| 314 | "src/f32-dwconv/gen/up8x9-psimd-acc2.c", |
| 315 | "src/f32-dwconv/gen/up8x9-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 316 | "src/f32-gavgpool/mp7p7q-psimd.c", |
| 317 | "src/f32-gavgpool/up7-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 318 | "src/f32-gemm/gen/1x8-psimd-loadsplat.c", |
| 319 | "src/f32-gemm/gen/1x8-psimd-splat.c", |
| 320 | "src/f32-gemm/gen/1x8s4-psimd.c", |
| 321 | "src/f32-gemm/gen/4x8-psimd-loadsplat.c", |
| 322 | "src/f32-gemm/gen/4x8-psimd-splat.c", |
| 323 | "src/f32-gemm/gen/4x8s4-psimd.c", |
| 324 | "src/f32-gemm/gen/6x8-psimd-loadsplat.c", |
| 325 | "src/f32-gemm/gen/6x8-psimd-splat.c", |
| 326 | "src/f32-gemm/gen/6x8s4-psimd.c", |
| 327 | "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c", |
| 328 | "src/f32-gemm/gen-inc/1x8-psimd-splat.c", |
| 329 | "src/f32-gemm/gen-inc/1x8s4-psimd.c", |
| 330 | "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c", |
| 331 | "src/f32-gemm/gen-inc/4x8-psimd-splat.c", |
| 332 | "src/f32-gemm/gen-inc/4x8s4-psimd.c", |
| 333 | "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c", |
| 334 | "src/f32-gemm/gen-inc/6x8-psimd-splat.c", |
| 335 | "src/f32-gemm/gen-inc/6x8s4-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 336 | "src/f32-hswish/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 337 | "src/f32-igemm/gen/1x8-psimd-loadsplat.c", |
| 338 | "src/f32-igemm/gen/1x8-psimd-splat.c", |
| 339 | "src/f32-igemm/gen/1x8s4-psimd.c", |
| 340 | "src/f32-igemm/gen/4x2c4-psimd.c", |
| 341 | "src/f32-igemm/gen/4x8-psimd-loadsplat.c", |
| 342 | "src/f32-igemm/gen/4x8-psimd-splat.c", |
| 343 | "src/f32-igemm/gen/4x8s4-psimd.c", |
| 344 | "src/f32-igemm/gen/6x8-psimd-loadsplat.c", |
| 345 | "src/f32-igemm/gen/6x8-psimd-splat.c", |
| 346 | "src/f32-igemm/gen/6x8s4-psimd.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 347 | "src/f32-maxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 348 | "src/f32-pavgpool/mp9p8q-psimd.c", |
| 349 | "src/f32-pavgpool/up9-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 350 | "src/f32-ppmm/gen/4x8-psimd.c", |
| 351 | "src/f32-prelu/gen/psimd-2x4.c", |
| 352 | "src/f32-prelu/gen/psimd-2x8.c", |
| 353 | "src/f32-vbinary/gen/vadd-psimd-x4.c", |
| 354 | "src/f32-vbinary/gen/vadd-psimd-x8.c", |
| 355 | "src/f32-vbinary/gen/vaddc-psimd-x4.c", |
| 356 | "src/f32-vbinary/gen/vaddc-psimd-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 357 | "src/f32-vbinary/gen/vdiv-psimd-x4.c", |
| 358 | "src/f32-vbinary/gen/vdiv-psimd-x8.c", |
| 359 | "src/f32-vbinary/gen/vdivc-psimd-x4.c", |
| 360 | "src/f32-vbinary/gen/vdivc-psimd-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 361 | "src/f32-vbinary/gen/vmax-psimd-x4.c", |
| 362 | "src/f32-vbinary/gen/vmax-psimd-x8.c", |
| 363 | "src/f32-vbinary/gen/vmaxc-psimd-x4.c", |
| 364 | "src/f32-vbinary/gen/vmaxc-psimd-x8.c", |
| 365 | "src/f32-vbinary/gen/vmin-psimd-x4.c", |
| 366 | "src/f32-vbinary/gen/vmin-psimd-x8.c", |
| 367 | "src/f32-vbinary/gen/vminc-psimd-x4.c", |
| 368 | "src/f32-vbinary/gen/vminc-psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 369 | "src/f32-vbinary/gen/vmul-psimd-x4.c", |
| 370 | "src/f32-vbinary/gen/vmul-psimd-x8.c", |
| 371 | "src/f32-vbinary/gen/vmulc-psimd-x4.c", |
| 372 | "src/f32-vbinary/gen/vmulc-psimd-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 373 | "src/f32-vbinary/gen/vrdivc-psimd-x4.c", |
| 374 | "src/f32-vbinary/gen/vrdivc-psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 375 | "src/f32-vbinary/gen/vrsubc-psimd-x4.c", |
| 376 | "src/f32-vbinary/gen/vrsubc-psimd-x8.c", |
| 377 | "src/f32-vbinary/gen/vsub-psimd-x4.c", |
| 378 | "src/f32-vbinary/gen/vsub-psimd-x8.c", |
| 379 | "src/f32-vbinary/gen/vsubc-psimd-x4.c", |
| 380 | "src/f32-vbinary/gen/vsubc-psimd-x8.c", |
| 381 | "src/f32-vmulcaddc/gen/c4-psimd-2x.c", |
| 382 | "src/f32-vmulcaddc/gen/c8-psimd-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 383 | "src/x32-packx/x4-psimd.c", |
| 384 | "src/x32-pad/x2-psimd.c", |
| 385 | "src/x32-unpool/psimd.c", |
| 386 | "src/x32-zip/x2-psimd.c", |
| 387 | "src/x32-zip/x3-psimd.c", |
| 388 | "src/x32-zip/x4-psimd.c", |
| 389 | "src/x32-zip/xm-psimd.c", |
| 390 | ] |
| 391 | |
| 392 | # ISA-specific micro-kernels |
| 393 | NEON_UKERNELS = [ |
| 394 | "src/f32-avgpool/mp9p8q-neon.c", |
| 395 | "src/f32-avgpool/up9-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 396 | "src/f32-bilinear/gen/neon-c4.c", |
| 397 | "src/f32-bilinear/gen/neon-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 398 | "src/f32-clamp/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 399 | "src/f32-dwconv/gen/up4x9-neon.c", |
| 400 | "src/f32-dwconv/gen/up4x9-neon-acc2.c", |
| 401 | "src/f32-dwconv/gen/up8x9-neon.c", |
| 402 | "src/f32-dwconv/gen/up8x9-neon-acc2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 403 | "src/f32-gavgpool-spchw/neon-x4.c", |
| 404 | "src/f32-gavgpool/mp7p7q-neon.c", |
| 405 | "src/f32-gavgpool/up7-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 406 | "src/f32-gemm/gen/1x8-neon-lane-ld64.c", |
| 407 | "src/f32-gemm/gen/4x2-neon-lane-ld64.c", |
| 408 | "src/f32-gemm/gen/4x8-neon-lane-ld128.c", |
| 409 | "src/f32-gemm/gen/4x8-neon-lane-ld64.c", |
| 410 | "src/f32-gemm/gen/5x8-neon-lane-ld64.c", |
| 411 | "src/f32-gemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 412 | "src/f32-gemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 413 | "src/f32-gemm/gen/1x8-neon-dup-ld64.c", |
| 414 | "src/f32-gemm/gen/4x8-neon-dup-ld128.c", |
| 415 | "src/f32-gemm/gen/4x8-neon-dup-ld64.c", |
| 416 | "src/f32-gemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 417 | "src/f32-gemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 418 | "src/f32-gemm/gen/1x8s4-neon.c", |
| 419 | "src/f32-gemm/gen/4x8s4-neon.c", |
| 420 | "src/f32-gemm/gen/6x8s4-neon.c", |
| 421 | "src/f32-gemm/gen/8x8s4-neon.c", |
| 422 | "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c", |
| 423 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c", |
| 424 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c", |
| 425 | "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c", |
| 426 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 427 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 428 | "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c", |
| 429 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c", |
| 430 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c", |
| 431 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 432 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 433 | "src/f32-gemm/gen-inc/1x8s4-neon.c", |
| 434 | "src/f32-gemm/gen-inc/4x8s4-neon.c", |
| 435 | "src/f32-gemm/gen-inc/6x8s4-neon.c", |
| 436 | "src/f32-gemm/gen-inc/8x8s4-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 437 | "src/f32-hswish/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 438 | "src/f32-igemm/gen/1x8-neon-lane-ld64.c", |
| 439 | "src/f32-igemm/gen/4x2-neon-lane-ld64.c", |
| 440 | "src/f32-igemm/gen/4x4-neon-lane-ld64.c", |
| 441 | "src/f32-igemm/gen/4x8-neon-lane-ld128.c", |
| 442 | "src/f32-igemm/gen/4x8-neon-lane-ld64.c", |
| 443 | "src/f32-igemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 444 | "src/f32-igemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 445 | "src/f32-igemm/gen/1x8-neon-dup-ld64.c", |
| 446 | "src/f32-igemm/gen/4x8-neon-dup-ld128.c", |
| 447 | "src/f32-igemm/gen/4x8-neon-dup-ld64.c", |
| 448 | "src/f32-igemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 449 | "src/f32-igemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 450 | "src/f32-igemm/gen/1x8s4-neon.c", |
| 451 | "src/f32-igemm/gen/4x8s4-neon.c", |
| 452 | "src/f32-igemm/gen/6x8s4-neon.c", |
| 453 | "src/f32-igemm/gen/8x8s4-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 454 | "src/f32-pavgpool/mp9p8q-neon.c", |
| 455 | "src/f32-pavgpool/up9-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 456 | "src/f32-ppmm/gen/4x8-neon.c", |
| 457 | "src/f32-ppmm/gen/8x8-neon.c", |
| 458 | "src/f32-prelu/gen/neon-2x4.c", |
| 459 | "src/f32-prelu/gen/neon-2x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 460 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 461 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
| 462 | "src/f32-vbinary/gen/vadd-neon-x4.c", |
| 463 | "src/f32-vbinary/gen/vadd-neon-x8.c", |
| 464 | "src/f32-vbinary/gen/vaddc-neon-x4.c", |
| 465 | "src/f32-vbinary/gen/vaddc-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 466 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 467 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 468 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 469 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 470 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 471 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 472 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 473 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 474 | "src/f32-vbinary/gen/vmul-neon-x4.c", |
| 475 | "src/f32-vbinary/gen/vmul-neon-x8.c", |
| 476 | "src/f32-vbinary/gen/vmulc-neon-x4.c", |
| 477 | "src/f32-vbinary/gen/vmulc-neon-x8.c", |
| 478 | "src/f32-vbinary/gen/vrsubc-neon-x4.c", |
| 479 | "src/f32-vbinary/gen/vrsubc-neon-x8.c", |
| 480 | "src/f32-vbinary/gen/vsub-neon-x4.c", |
| 481 | "src/f32-vbinary/gen/vsub-neon-x8.c", |
| 482 | "src/f32-vbinary/gen/vsubc-neon-x4.c", |
| 483 | "src/f32-vbinary/gen/vsubc-neon-x8.c", |
| 484 | "src/f32-vmulcaddc/gen/c4-neon-2x.c", |
| 485 | "src/f32-vmulcaddc/gen/c8-neon-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 486 | "src/q8-avgpool/mp9p8q-neon.c", |
| 487 | "src/q8-avgpool/up9-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 488 | "src/q8-dwconv/up8x9-neon.c", |
| 489 | "src/q8-gavgpool/mp7p7q-neon.c", |
| 490 | "src/q8-gavgpool/up7-neon.c", |
| 491 | "src/q8-gemm/4x8-neon.c", |
| 492 | "src/q8-gemm/8x8-neon.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 493 | "src/q8-igemm/4x8-neon.c", |
| 494 | "src/q8-igemm/8x8-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 495 | "src/q8-vadd/neon.c", |
| 496 | "src/u8-clamp/neon.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 497 | "src/u8-maxpool/9p8x-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 498 | "src/u8-rmax/neon.c", |
| 499 | "src/x32-packx/x4-neon-st4.c", |
| 500 | "src/x32-pad/x2-neon.c", |
| 501 | "src/x32-zip/x2-neon.c", |
| 502 | "src/x32-zip/x3-neon.c", |
| 503 | "src/x32-zip/x4-neon.c", |
| 504 | "src/x32-zip/xm-neon.c", |
| 505 | "src/x8-zip/x2-neon.c", |
| 506 | "src/x8-zip/x3-neon.c", |
| 507 | "src/x8-zip/x4-neon.c", |
| 508 | "src/x8-zip/xm-neon.c", |
| 509 | ] |
| 510 | |
| 511 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 512 | "src/f32-bilinear/gen/neonfma-c4.c", |
| 513 | "src/f32-bilinear/gen/neonfma-c8.c", |
| 514 | "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c", |
| 515 | "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c", |
| 516 | "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c", |
| 517 | "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 518 | "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 519 | "src/f32-igemm/gen/1x8s4-neonfma.c", |
| 520 | "src/f32-igemm/gen/4x8s4-neonfma.c", |
| 521 | "src/f32-igemm/gen/6x8s4-neonfma.c", |
| 522 | "src/f32-igemm/gen/8x8s4-neonfma.c", |
| 523 | "src/f32-dwconv/gen/up4x9-neonfma.c", |
| 524 | "src/f32-dwconv/gen/up4x9-neonfma-acc2.c", |
| 525 | "src/f32-dwconv/gen/up8x9-neonfma.c", |
| 526 | "src/f32-dwconv/gen/up8x9-neonfma-acc2.c", |
| 527 | "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c", |
| 528 | "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c", |
| 529 | "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c", |
| 530 | "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 531 | "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 532 | "src/f32-gemm/gen/1x8s4-neonfma.c", |
| 533 | "src/f32-gemm/gen/4x8s4-neonfma.c", |
| 534 | "src/f32-gemm/gen/6x8s4-neonfma.c", |
| 535 | "src/f32-gemm/gen/8x8s4-neonfma.c", |
| 536 | "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c", |
| 537 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c", |
| 538 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c", |
| 539 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 540 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 541 | "src/f32-gemm/gen-inc/1x8s4-neonfma.c", |
| 542 | "src/f32-gemm/gen-inc/4x8s4-neonfma.c", |
| 543 | "src/f32-gemm/gen-inc/6x8s4-neonfma.c", |
| 544 | "src/f32-gemm/gen-inc/8x8s4-neonfma.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 545 | "src/f32-hswish/neonfma.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 546 | "src/f32-ppmm/gen/4x8-neonfma.c", |
| 547 | "src/f32-ppmm/gen/8x8-neonfma.c", |
| 548 | "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c", |
| 549 | "src/f32-vmulcaddc/gen/c4-neonfma-2x.c", |
| 550 | "src/f32-vmulcaddc/gen/c8-neonfma-2x.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 551 | "src/math/exp-neonfma-lut64-p2.c", |
| 552 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 553 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 554 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 555 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 91f8d86 | 2019-11-27 12:25:42 -0800 | [diff] [blame] | 556 | "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c", |
| 557 | "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c", |
| 558 | "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c", |
Marat Dukhan | 22aae13 | 2019-11-22 17:10:29 -0800 | [diff] [blame] | 559 | "src/math/sigmoid-neonfma-p5-nr1recps1fma.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 560 | "src/math/sigmoid-neonfma-p5-nr2fma.c", |
Marat Dukhan | 22aae13 | 2019-11-22 17:10:29 -0800 | [diff] [blame] | 561 | "src/math/sigmoid-neonfma-p5-nr2recps.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 562 | ] |
| 563 | |
| 564 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 565 | "src/f32-vbinary/gen/vdiv-neon-x4.c", |
| 566 | "src/f32-vbinary/gen/vdiv-neon-x8.c", |
| 567 | "src/f32-vbinary/gen/vdivc-neon-x4.c", |
| 568 | "src/f32-vbinary/gen/vdivc-neon-x8.c", |
| 569 | "src/f32-vbinary/gen/vrdivc-neon-x4.c", |
| 570 | "src/f32-vbinary/gen/vrdivc-neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 571 | "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c", |
| 572 | "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c", |
| 573 | "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c", |
| 574 | "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c", |
| 575 | "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c", |
| 576 | "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 577 | "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 578 | "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c", |
| 579 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c", |
| 580 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c", |
| 581 | "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c", |
| 582 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 583 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 584 | "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c", |
| 585 | "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c", |
| 586 | "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c", |
| 587 | "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c", |
| 588 | "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c", |
| 589 | "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 590 | "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 591 | "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c", |
| 592 | "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c", |
| 593 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c", |
| 594 | "src/f32-dwconv-spchw/3x3p1-neonfma.c", |
| 595 | "src/f32-dwconv-spchw/5x5p2-neonfma.c", |
| 596 | "src/f32-dwconv-spchw/3x3s2p1-neonfma.c", |
| 597 | "src/f32-dwconv-spchw/5x5s2p2-neonfma.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 598 | "src/f32-spmm/gen/12x1-neonfma.c", |
| 599 | "src/f32-spmm/gen/12x2-neonfma.c", |
| 600 | "src/f32-spmm/gen/12x4-neonfma.c", |
| 601 | "src/f32-spmm/gen/16x1-neonfma-pipelined.c", |
| 602 | "src/f32-spmm/gen/16x1-neonfma-unroll2.c", |
| 603 | "src/f32-spmm/gen/16x1-neonfma.c", |
| 604 | "src/f32-spmm/gen/16x2-neonfma.c", |
| 605 | "src/f32-spmm/gen/16x4-neonfma.c", |
| 606 | "src/f32-spmm/gen/4x1-neonfma-pipelined.c", |
| 607 | "src/f32-spmm/gen/4x1-neonfma-unroll2.c", |
| 608 | "src/f32-spmm/gen/4x1-neonfma.c", |
| 609 | "src/f32-spmm/gen/4x2-neonfma.c", |
| 610 | "src/f32-spmm/gen/4x4-neonfma.c", |
| 611 | "src/f32-spmm/gen/8x1-neonfma-pipelined.c", |
| 612 | "src/f32-spmm/gen/8x1-neonfma-unroll2.c", |
| 613 | "src/f32-spmm/gen/8x1-neonfma.c", |
| 614 | "src/f32-spmm/gen/8x2-neonfma.c", |
| 615 | "src/f32-spmm/gen/8x4-neonfma.c", |
Marat Dukhan | 22aae13 | 2019-11-22 17:10:29 -0800 | [diff] [blame] | 616 | "src/math/sigmoid-neonfma-p5-div.c", |
Marat Dukhan | 91f8d86 | 2019-11-27 12:25:42 -0800 | [diff] [blame] | 617 | "src/math/sigmoid-neonfma-lut2048-p1-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 618 | ] |
| 619 | |
| 620 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 621 | "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c", |
| 622 | "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c", |
| 623 | "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 624 | ] |
| 625 | |
| 626 | SSE_UKERNELS = [ |
| 627 | "src/f32-avgpool/mp9p8q-sse.c", |
| 628 | "src/f32-avgpool/up9-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 629 | "src/f32-bilinear/gen/sse-c4.c", |
| 630 | "src/f32-bilinear/gen/sse-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 631 | "src/f32-clamp/sse.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 632 | "src/f32-dwconv-spchw/3x3p1-sse.c", |
| 633 | "src/f32-dwconv-spchw/3x3s2p1-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 634 | "src/f32-dwconv/gen/up4x25-sse-acc2.c", |
| 635 | "src/f32-dwconv/gen/up4x25-sse.c", |
| 636 | "src/f32-dwconv/gen/up4x4-sse-acc2.c", |
| 637 | "src/f32-dwconv/gen/up4x4-sse.c", |
| 638 | "src/f32-dwconv/gen/up4x9-sse-acc2.c", |
| 639 | "src/f32-dwconv/gen/up4x9-sse.c", |
| 640 | "src/f32-dwconv/gen/up8x25-sse-acc2.c", |
| 641 | "src/f32-dwconv/gen/up8x25-sse.c", |
| 642 | "src/f32-dwconv/gen/up8x4-sse-acc2.c", |
| 643 | "src/f32-dwconv/gen/up8x4-sse.c", |
| 644 | "src/f32-dwconv/gen/up8x9-sse-acc2.c", |
| 645 | "src/f32-dwconv/gen/up8x9-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 646 | "src/f32-gavgpool-spchw/sse-x4.c", |
| 647 | "src/f32-gavgpool/mp7p7q-sse.c", |
| 648 | "src/f32-gavgpool/up7-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 649 | "src/f32-gemm/gen/1x8-sse-dup.c", |
| 650 | "src/f32-gemm/gen/1x8-sse-load1.c", |
| 651 | "src/f32-gemm/gen/1x8s4-sse.c", |
| 652 | "src/f32-gemm/gen/4x8-sse-dup.c", |
| 653 | "src/f32-gemm/gen/4x8-sse-load1.c", |
| 654 | "src/f32-gemm/gen/4x8s4-sse.c", |
| 655 | "src/f32-gemm/gen-inc/1x8-sse-dup.c", |
| 656 | "src/f32-gemm/gen-inc/1x8-sse-load1.c", |
| 657 | "src/f32-gemm/gen-inc/1x8s4-sse.c", |
| 658 | "src/f32-gemm/gen-inc/4x8-sse-dup.c", |
| 659 | "src/f32-gemm/gen-inc/4x8-sse-load1.c", |
| 660 | "src/f32-gemm/gen-inc/4x8s4-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 661 | "src/f32-hswish/sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 662 | "src/f32-igemm/gen/1x8-sse-dup.c", |
| 663 | "src/f32-igemm/gen/1x8-sse-load1.c", |
| 664 | "src/f32-igemm/gen/1x8s4-sse.c", |
| 665 | "src/f32-igemm/gen/4x2c4-sse.c", |
| 666 | "src/f32-igemm/gen/4x8-sse-dup.c", |
| 667 | "src/f32-igemm/gen/4x8-sse-load1.c", |
| 668 | "src/f32-igemm/gen/4x8s4-sse.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 669 | "src/f32-maxpool/9p8x-sse-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 670 | "src/f32-pavgpool/mp9p8q-sse.c", |
| 671 | "src/f32-pavgpool/up9-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 672 | "src/f32-ppmm/gen/4x8-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 673 | "src/f32-rmax/sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 674 | "src/f32-spmm/gen/4x1-sse.c", |
| 675 | "src/f32-spmm/gen/8x1-sse.c", |
| 676 | "src/f32-vbinary/gen/vadd-sse-x4.c", |
| 677 | "src/f32-vbinary/gen/vadd-sse-x8.c", |
| 678 | "src/f32-vbinary/gen/vaddc-sse-x4.c", |
| 679 | "src/f32-vbinary/gen/vaddc-sse-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 680 | "src/f32-vbinary/gen/vdiv-sse-x4.c", |
| 681 | "src/f32-vbinary/gen/vdiv-sse-x8.c", |
| 682 | "src/f32-vbinary/gen/vdivc-sse-x4.c", |
| 683 | "src/f32-vbinary/gen/vdivc-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 684 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 685 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 686 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 687 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 688 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 689 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 690 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 691 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 692 | "src/f32-vbinary/gen/vmul-sse-x4.c", |
| 693 | "src/f32-vbinary/gen/vmul-sse-x8.c", |
| 694 | "src/f32-vbinary/gen/vmulc-sse-x4.c", |
| 695 | "src/f32-vbinary/gen/vmulc-sse-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 696 | "src/f32-vbinary/gen/vrdivc-sse-x4.c", |
| 697 | "src/f32-vbinary/gen/vrdivc-sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 698 | "src/f32-vbinary/gen/vrsubc-sse-x4.c", |
| 699 | "src/f32-vbinary/gen/vrsubc-sse-x8.c", |
| 700 | "src/f32-vbinary/gen/vsub-sse-x4.c", |
| 701 | "src/f32-vbinary/gen/vsub-sse-x8.c", |
| 702 | "src/f32-vbinary/gen/vsubc-sse-x4.c", |
| 703 | "src/f32-vbinary/gen/vsubc-sse-x8.c", |
| 704 | "src/f32-vmulcaddc/gen/c4-sse-2x.c", |
| 705 | "src/f32-vmulcaddc/gen/c8-sse-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 706 | "src/x32-packx/x4-sse.c", |
| 707 | ] |
| 708 | |
| 709 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 710 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 711 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 712 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 713 | "src/f32-prelu/gen/sse2-2x4.c", |
| 714 | "src/f32-prelu/gen/sse2-2x8.c", |
| 715 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
| 716 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 717 | "src/q8-avgpool/mp9p8q-sse2.c", |
| 718 | "src/q8-avgpool/up9-sse2.c", |
| 719 | "src/q8-igemm/4x4c2-sse2.c", |
| 720 | "src/q8-dwconv/up8x9-sse2.c", |
| 721 | "src/q8-gavgpool/mp7p7q-sse2.c", |
| 722 | "src/q8-gavgpool/up7-sse2.c", |
| 723 | "src/q8-gemm/2x4c8-sse2.c", |
| 724 | "src/q8-gemm/4x4c2-sse2.c", |
| 725 | "src/q8-vadd/sse2.c", |
| 726 | "src/u8-clamp/sse2.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 727 | "src/u8-maxpool/9p8x-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 728 | "src/u8-rmax/sse2.c", |
| 729 | "src/x32-pad/x2-sse2.c", |
| 730 | "src/x32-zip/x2-sse2.c", |
| 731 | "src/x32-zip/x3-sse2.c", |
| 732 | "src/x32-zip/x4-sse2.c", |
| 733 | "src/x32-zip/xm-sse2.c", |
| 734 | "src/x8-zip/x2-sse2.c", |
| 735 | "src/x8-zip/x3-sse2.c", |
| 736 | "src/x8-zip/x4-sse2.c", |
| 737 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 738 | "src/math/exp-sse2-p5.c", |
| 739 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 740 | "src/math/sigmoid-sse2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 741 | ] |
| 742 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 743 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 744 | "src/f32-prelu/gen/sse41-2x4.c", |
| 745 | "src/f32-prelu/gen/sse41-2x8.c", |
| 746 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
| 747 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 748 | ] |
| 749 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 750 | AVX_UKERNELS = [ |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 751 | "src/f32-clamp/avx.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 752 | "src/f32-dwconv/gen/up16x4-avx-acc2.c", |
| 753 | "src/f32-dwconv/gen/up16x4-avx.c", |
| 754 | "src/f32-dwconv/gen/up8x4-avx-acc2.c", |
| 755 | "src/f32-dwconv/gen/up8x4-avx.c", |
| 756 | "src/f32-dwconv/gen/up16x9-avx-acc2.c", |
| 757 | "src/f32-dwconv/gen/up16x9-avx.c", |
| 758 | "src/f32-dwconv/gen/up8x9-avx-acc2.c", |
| 759 | "src/f32-dwconv/gen/up8x9-avx.c", |
| 760 | "src/f32-dwconv/gen/up16x25-avx-acc2.c", |
| 761 | "src/f32-dwconv/gen/up16x25-avx.c", |
| 762 | "src/f32-dwconv/gen/up8x25-avx-acc2.c", |
| 763 | "src/f32-dwconv/gen/up8x25-avx.c", |
| 764 | "src/f32-gemm/gen/1x8-avx-broadcast.c", |
| 765 | "src/f32-gemm/gen/4x8-avx-broadcast.c", |
| 766 | "src/f32-gemm/gen/5x8-avx-broadcast.c", |
| 767 | "src/f32-gemm/gen/6x8-avx-broadcast.c", |
| 768 | "src/f32-gemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 769 | "src/f32-gemm/gen/1x16-avx-broadcast.c", |
| 770 | "src/f32-gemm/gen/3x16-avx-broadcast.c", |
| 771 | "src/f32-gemm/gen/4x16-avx-broadcast.c", |
| 772 | "src/f32-gemm/gen/5x16-avx-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 773 | "src/f32-gemm/gen-inc/1x8-avx-broadcast.c", |
| 774 | "src/f32-gemm/gen-inc/4x8-avx-broadcast.c", |
| 775 | "src/f32-gemm/gen-inc/5x8-avx-broadcast.c", |
| 776 | "src/f32-gemm/gen-inc/6x8-avx-broadcast.c", |
| 777 | "src/f32-gemm/gen-inc/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 778 | "src/f32-gemm/gen-inc/1x16-avx-broadcast.c", |
| 779 | "src/f32-gemm/gen-inc/3x16-avx-broadcast.c", |
| 780 | "src/f32-gemm/gen-inc/4x16-avx-broadcast.c", |
| 781 | "src/f32-gemm/gen-inc/5x16-avx-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 782 | "src/f32-igemm/gen/1x8-avx-broadcast.c", |
| 783 | "src/f32-igemm/gen/4x8-avx-broadcast.c", |
| 784 | "src/f32-igemm/gen/5x8-avx-broadcast.c", |
| 785 | "src/f32-igemm/gen/6x8-avx-broadcast.c", |
| 786 | "src/f32-igemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 787 | "src/f32-igemm/gen/1x16-avx-broadcast.c", |
| 788 | "src/f32-igemm/gen/3x16-avx-broadcast.c", |
| 789 | "src/f32-igemm/gen/4x16-avx-broadcast.c", |
| 790 | "src/f32-igemm/gen/5x16-avx-broadcast.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 791 | "src/f32-rmax/avx.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 792 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 793 | ] |
| 794 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 795 | FMA3_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 796 | "src/f32-dwconv/gen/up16x4-fma3-acc2.c", |
| 797 | "src/f32-dwconv/gen/up16x4-fma3.c", |
| 798 | "src/f32-dwconv/gen/up8x4-fma3-acc2.c", |
| 799 | "src/f32-dwconv/gen/up8x4-fma3.c", |
| 800 | "src/f32-dwconv/gen/up16x9-fma3-acc2.c", |
| 801 | "src/f32-dwconv/gen/up16x9-fma3.c", |
| 802 | "src/f32-dwconv/gen/up8x9-fma3-acc2.c", |
| 803 | "src/f32-dwconv/gen/up8x9-fma3.c", |
| 804 | "src/f32-dwconv/gen/up16x25-fma3-acc2.c", |
| 805 | "src/f32-dwconv/gen/up16x25-fma3.c", |
| 806 | "src/f32-dwconv/gen/up8x25-fma3-acc2.c", |
| 807 | "src/f32-dwconv/gen/up8x25-fma3.c", |
| 808 | "src/f32-gemm/gen/1x8-fma3-broadcast.c", |
| 809 | "src/f32-gemm/gen/4x8-fma3-broadcast.c", |
| 810 | "src/f32-gemm/gen/5x8-fma3-broadcast.c", |
| 811 | "src/f32-gemm/gen/6x8-fma3-broadcast.c", |
| 812 | "src/f32-gemm/gen/7x8-fma3-broadcast.c", |
| 813 | "src/f32-gemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 814 | "src/f32-gemm/gen/1x16-fma3-broadcast.c", |
| 815 | "src/f32-gemm/gen/3x16-fma3-broadcast.c", |
| 816 | "src/f32-gemm/gen/4x16-fma3-broadcast.c", |
| 817 | "src/f32-gemm/gen/5x16-fma3-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 818 | "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c", |
| 819 | "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c", |
| 820 | "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c", |
| 821 | "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c", |
| 822 | "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c", |
| 823 | "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 824 | "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c", |
| 825 | "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c", |
| 826 | "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c", |
| 827 | "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 828 | "src/f32-igemm/gen/1x8-fma3-broadcast.c", |
| 829 | "src/f32-igemm/gen/4x8-fma3-broadcast.c", |
| 830 | "src/f32-igemm/gen/5x8-fma3-broadcast.c", |
| 831 | "src/f32-igemm/gen/6x8-fma3-broadcast.c", |
| 832 | "src/f32-igemm/gen/7x8-fma3-broadcast.c", |
| 833 | "src/f32-igemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 834 | "src/f32-igemm/gen/1x16-fma3-broadcast.c", |
| 835 | "src/f32-igemm/gen/3x16-fma3-broadcast.c", |
| 836 | "src/f32-igemm/gen/4x16-fma3-broadcast.c", |
| 837 | "src/f32-igemm/gen/5x16-fma3-broadcast.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 838 | ] |
| 839 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 840 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame^] | 841 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 842 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 843 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 844 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 845 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 846 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 847 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 848 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 849 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 850 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 851 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 852 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 853 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 854 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 855 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 856 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 857 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 858 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 859 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 860 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 861 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 862 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 863 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 864 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 865 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 866 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 867 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 868 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 869 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 870 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 871 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 872 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 873 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 874 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 875 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 876 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 877 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 878 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 879 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 880 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 881 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 882 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 883 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 884 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 885 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 886 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 887 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 888 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 889 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 890 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 891 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 892 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 893 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 894 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 895 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 896 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 897 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 898 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 899 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 900 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 901 | "src/math/exp-avx2-p5.c", |
| 902 | "src/math/exp-avx2-perm-p3.c", |
| 903 | "src/math/exp-avx2-perm-p4.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 904 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 905 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 906 | ] |
| 907 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 908 | AVX512F_UKERNELS = [ |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 909 | "src/f32-clamp/avx512f.c", |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 910 | "src/f32-dwconv/gen/up32x4-avx512f-acc2.c", |
| 911 | "src/f32-dwconv/gen/up32x4-avx512f.c", |
| 912 | "src/f32-dwconv/gen/up16x4-avx512f-acc2.c", |
| 913 | "src/f32-dwconv/gen/up16x4-avx512f.c", |
| 914 | "src/f32-dwconv/gen/up32x9-avx512f-acc2.c", |
| 915 | "src/f32-dwconv/gen/up32x9-avx512f.c", |
| 916 | "src/f32-dwconv/gen/up16x9-avx512f-acc2.c", |
| 917 | "src/f32-dwconv/gen/up16x9-avx512f.c", |
| 918 | "src/f32-dwconv/gen/up32x25-avx512f-acc2.c", |
| 919 | "src/f32-dwconv/gen/up32x25-avx512f.c", |
| 920 | "src/f32-dwconv/gen/up16x25-avx512f-acc2.c", |
| 921 | "src/f32-dwconv/gen/up16x25-avx512f.c", |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 922 | "src/f32-gemm/gen/1x16-avx512f-broadcast.c", |
| 923 | "src/f32-gemm/gen/4x16-avx512f-broadcast.c", |
| 924 | "src/f32-gemm/gen/5x16-avx512f-broadcast.c", |
| 925 | "src/f32-gemm/gen/6x16-avx512f-broadcast.c", |
| 926 | "src/f32-gemm/gen/7x16-avx512f-broadcast.c", |
| 927 | "src/f32-gemm/gen/8x16-avx512f-broadcast.c", |
| 928 | "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c", |
| 929 | "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c", |
| 930 | "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c", |
| 931 | "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c", |
| 932 | "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c", |
| 933 | "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c", |
| 934 | "src/f32-igemm/gen/1x16-avx512f-broadcast.c", |
| 935 | "src/f32-igemm/gen/4x16-avx512f-broadcast.c", |
| 936 | "src/f32-igemm/gen/5x16-avx512f-broadcast.c", |
| 937 | "src/f32-igemm/gen/6x16-avx512f-broadcast.c", |
| 938 | "src/f32-igemm/gen/7x16-avx512f-broadcast.c", |
| 939 | "src/f32-igemm/gen/8x16-avx512f-broadcast.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame^] | 940 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 941 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 942 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 943 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 944 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 945 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 946 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 947 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 948 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 949 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 950 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 951 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 952 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 953 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 954 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 955 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 956 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 957 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 958 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 959 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 960 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 961 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 962 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 963 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 964 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 965 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 966 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 967 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 968 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 969 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 970 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 971 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 972 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 973 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 974 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 975 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 976 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 977 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame^] | 978 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 979 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 980 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 981 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 982 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 983 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 984 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 985 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 986 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 987 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 988 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 989 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 990 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 991 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 992 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 993 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 994 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 995 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 996 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 997 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 998 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 999 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 1000 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 1001 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1002 | "src/math/exp-avx512f-p5-scalef.c", |
| 1003 | "src/math/exp-avx512f-p5.c", |
| 1004 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | feb4923 | 2019-10-28 11:03:31 -0700 | [diff] [blame] | 1005 | "src/math/exp-avx512f-perm2-p2.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1006 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1007 | ] |
| 1008 | |
| 1009 | AARCH32_ASM_UKERNELS = [ |
| 1010 | "src/q8-dwconv/up8x9-aarch32-neon.S", |
Frank Barchard | 3e237f2 | 2019-12-04 23:08:51 -0800 | [diff] [blame] | 1011 | "src/f32-gemm/4x8-aarch32-neon-cortex-a75.S", |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 1012 | "src/f32-gemm/4x8-aarch32-neon-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1013 | ] |
| 1014 | |
| 1015 | AARCH64_ASM_UKERNELS = [ |
| 1016 | "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S", |
| 1017 | "src/f32-dwconv/up4x9-aarch64-neonfma.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1018 | "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S", |
| 1019 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S", |
| 1020 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1021 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
| 1022 | "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S", |
| 1023 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S", |
| 1024 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1025 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1026 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S", |
| 1027 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S", |
| 1028 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
| 1029 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S", |
| 1030 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
| 1031 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S", |
| 1032 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
| 1033 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S", |
| 1034 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S", |
| 1035 | "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S", |
| 1036 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S", |
| 1037 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S", |
| 1038 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S", |
| 1039 | "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S", |
| 1040 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S", |
| 1041 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S", |
| 1042 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S", |
| 1043 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S", |
| 1044 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S", |
| 1045 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S", |
| 1046 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S", |
| 1047 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S", |
| 1048 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S", |
| 1049 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S", |
| 1050 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S", |
| 1051 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1052 | "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 1053 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1054 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S", |
| 1055 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S", |
| 1056 | "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 1057 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1058 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S", |
| 1059 | "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 1060 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1061 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S", |
| 1062 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S", |
| 1063 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S", |
| 1064 | ] |
| 1065 | |
| 1066 | INTERNAL_MICROKERNEL_HDRS = [ |
| 1067 | "src/xnnpack/argmaxpool.h", |
| 1068 | "src/xnnpack/avgpool.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 1069 | "src/xnnpack/bilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1070 | "src/xnnpack/clamp.h", |
| 1071 | "src/xnnpack/common.h", |
| 1072 | "src/xnnpack/conv.h", |
| 1073 | "src/xnnpack/dwconv.h", |
| 1074 | "src/xnnpack/gavgpool.h", |
| 1075 | "src/xnnpack/gemm.h", |
| 1076 | "src/xnnpack/hswish.h", |
| 1077 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 1078 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1079 | "src/xnnpack/lut.h", |
| 1080 | "src/xnnpack/math.h", |
| 1081 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1082 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1083 | "src/xnnpack/packx.h", |
| 1084 | "src/xnnpack/pad.h", |
| 1085 | "src/xnnpack/params.h", |
| 1086 | "src/xnnpack/pavgpool.h", |
| 1087 | "src/xnnpack/ppmm.h", |
| 1088 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1089 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1090 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1091 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1092 | "src/xnnpack/rmax.h", |
| 1093 | "src/xnnpack/scalar-utils.h", |
| 1094 | "src/xnnpack/spmm.h", |
| 1095 | "src/xnnpack/unpool.h", |
| 1096 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1097 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1098 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1099 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1100 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1101 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1102 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1103 | "src/xnnpack/zip.h", |
| 1104 | ] |
| 1105 | |
| 1106 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1107 | "include/xnnpack.h", |
| 1108 | "src/xnnpack/allocator.h", |
| 1109 | "src/xnnpack/compute.h", |
| 1110 | "src/xnnpack/im2col.h", |
| 1111 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1112 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1113 | "src/xnnpack/operator.h", |
| 1114 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1115 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1116 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1117 | "src/xnnpack/requantization.h", |
| 1118 | ] |
| 1119 | |
| 1120 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1121 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1122 | ] |
| 1123 | |
| 1124 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1125 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1126 | "include/xnnpack.h", |
| 1127 | ] |
| 1128 | |
| 1129 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1130 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1131 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1132 | "src/xnnpack/requantization.h", |
| 1133 | "include/xnnpack.h", |
| 1134 | ] |
| 1135 | |
| 1136 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 1137 | "src/xnnpack/params.h", |
| 1138 | "src/xnnpack/common.h", |
| 1139 | ] |
| 1140 | |
| 1141 | WEIGHTS_PACK_HDRS = [ |
| 1142 | "src/xnnpack/pack.h", |
| 1143 | "src/xnnpack/operator.h", |
| 1144 | "src/xnnpack/compute.h", |
| 1145 | ] |
| 1146 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1147 | LOGGING_COPTS = select({ |
| 1148 | # No logging in optimized mode |
| 1149 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 1150 | # Full logging in debug mode |
| 1151 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 1152 | # Error-only logging in default (fastbuild) mode |
| 1153 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 1154 | }) |
| 1155 | |
| 1156 | LOGGING_HDRS = [ |
| 1157 | "src/xnnpack/log.h", |
| 1158 | ] |
| 1159 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1160 | xnnpack_cc_library( |
| 1161 | name = "scalar_ukernels", |
| 1162 | srcs = SCALAR_UKERNELS, |
| 1163 | hdrs = INTERNAL_HDRS, |
| 1164 | aarch32_copts = ["-marm"], |
| 1165 | copts = xnnpack_std_copts(), |
| 1166 | deps = [ |
| 1167 | "@FP16", |
| 1168 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1169 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1170 | ], |
| 1171 | ) |
| 1172 | |
| 1173 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1174 | name = "wasm_ukernels", |
| 1175 | hdrs = INTERNAL_HDRS, |
| 1176 | copts = xnnpack_std_copts(), |
| 1177 | wasm_srcs = WASM_UKERNELS, |
| 1178 | deps = [ |
| 1179 | "@FP16", |
| 1180 | "@FXdiv", |
| 1181 | "@pthreadpool", |
| 1182 | ], |
| 1183 | ) |
| 1184 | |
| 1185 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1186 | name = "psimd_ukernels", |
| 1187 | srcs = PSIMD_UKERNELS, |
| 1188 | hdrs = INTERNAL_HDRS, |
| 1189 | aarch32_copts = [ |
| 1190 | "-marm", |
| 1191 | "-mfpu=neon", |
| 1192 | ], |
| 1193 | copts = xnnpack_std_copts(), |
| 1194 | optimized_copts = [ |
| 1195 | "-O3", |
| 1196 | "-ffast-math", |
| 1197 | ], |
| 1198 | deps = [ |
| 1199 | "@FP16", |
| 1200 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1201 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1202 | ], |
| 1203 | ) |
| 1204 | |
| 1205 | xnnpack_cc_library( |
| 1206 | name = "neon_ukernels", |
| 1207 | hdrs = INTERNAL_HDRS, |
| 1208 | aarch32_copts = [ |
| 1209 | "-marm", |
| 1210 | "-mfpu=neon", |
| 1211 | ], |
| 1212 | aarch32_srcs = NEON_UKERNELS, |
| 1213 | aarch64_srcs = NEON_UKERNELS, |
| 1214 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1215 | deps = [ |
| 1216 | "@FP16", |
| 1217 | "@pthreadpool", |
| 1218 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1219 | ) |
| 1220 | |
| 1221 | xnnpack_cc_library( |
| 1222 | name = "neonfma_ukernels", |
| 1223 | hdrs = INTERNAL_HDRS, |
| 1224 | aarch32_copts = [ |
| 1225 | "-marm", |
| 1226 | "-mfpu=neon-vfpv4", |
| 1227 | ], |
| 1228 | aarch32_srcs = NEONFMA_UKERNELS, |
| 1229 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 1230 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1231 | deps = [ |
| 1232 | "@FP16", |
| 1233 | "@pthreadpool", |
| 1234 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1235 | ) |
| 1236 | |
| 1237 | xnnpack_cc_library( |
| 1238 | name = "neonfp16arith_ukernels", |
| 1239 | hdrs = INTERNAL_HDRS, |
| 1240 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 1241 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 1242 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1243 | deps = [ |
| 1244 | "@FP16", |
| 1245 | "@pthreadpool", |
| 1246 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1247 | ) |
| 1248 | |
| 1249 | xnnpack_cc_library( |
| 1250 | name = "sse2_ukernels", |
| 1251 | hdrs = INTERNAL_HDRS, |
| 1252 | copts = xnnpack_std_copts(), |
| 1253 | x86_copts = ["-msse2"], |
| 1254 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1255 | deps = [ |
| 1256 | "@FP16", |
| 1257 | "@pthreadpool", |
| 1258 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1259 | ) |
| 1260 | |
| 1261 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1262 | name = "sse41_ukernels", |
| 1263 | hdrs = INTERNAL_HDRS, |
| 1264 | copts = xnnpack_std_copts(), |
| 1265 | x86_copts = ["-msse4.1"], |
| 1266 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1267 | deps = [ |
| 1268 | "@FP16", |
| 1269 | "@pthreadpool", |
| 1270 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1271 | ) |
| 1272 | |
| 1273 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1274 | name = "avx_ukernels", |
| 1275 | hdrs = INTERNAL_HDRS, |
| 1276 | copts = xnnpack_std_copts(), |
| 1277 | x86_copts = ["-mavx"], |
| 1278 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1279 | deps = [ |
| 1280 | "@FP16", |
| 1281 | "@pthreadpool", |
| 1282 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1283 | ) |
| 1284 | |
| 1285 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1286 | name = "fma3_ukernels", |
| 1287 | hdrs = INTERNAL_HDRS, |
| 1288 | copts = xnnpack_std_copts(), |
| 1289 | x86_copts = [ |
| 1290 | "-mfma", |
| 1291 | ], |
| 1292 | x86_srcs = FMA3_UKERNELS, |
| 1293 | deps = [ |
| 1294 | "@FP16", |
| 1295 | "@pthreadpool", |
| 1296 | ], |
| 1297 | ) |
| 1298 | |
| 1299 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1300 | name = "avx2_ukernels", |
| 1301 | hdrs = INTERNAL_HDRS, |
| 1302 | copts = xnnpack_std_copts(), |
| 1303 | x86_copts = [ |
| 1304 | "-mfma", |
| 1305 | "-mavx2", |
| 1306 | ], |
| 1307 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1308 | deps = [ |
| 1309 | "@FP16", |
| 1310 | "@pthreadpool", |
| 1311 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1312 | ) |
| 1313 | |
| 1314 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1315 | name = "avx512f_ukernels", |
| 1316 | hdrs = INTERNAL_HDRS, |
| 1317 | copts = xnnpack_std_copts(), |
| 1318 | x86_copts = ["-mavx512f"], |
| 1319 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1320 | deps = [ |
| 1321 | "@FP16", |
| 1322 | "@pthreadpool", |
| 1323 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1324 | ) |
| 1325 | |
| 1326 | xnnpack_cc_library( |
| 1327 | name = "asm_ukernels", |
| 1328 | hdrs = ["src/xnnpack/assembly.h"], |
| 1329 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
| 1330 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 1331 | ) |
| 1332 | |
| 1333 | xnnpack_aggregate_library( |
| 1334 | name = "ukernels", |
| 1335 | aarch32_deps = [ |
| 1336 | ":psimd_ukernels", |
| 1337 | ":neon_ukernels", |
| 1338 | ":neonfma_ukernels", |
| 1339 | ":asm_ukernels", |
| 1340 | ], |
| 1341 | aarch64_deps = [ |
| 1342 | ":psimd_ukernels", |
| 1343 | ":neon_ukernels", |
| 1344 | ":neonfma_ukernels", |
| 1345 | ":neonfp16arith_ukernels", |
| 1346 | ":asm_ukernels", |
| 1347 | ], |
| 1348 | generic_deps = [":scalar_ukernels"], |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1349 | wasm_deps = [ |
| 1350 | ":wasm_ukernels", |
| 1351 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1352 | wasmsimd_deps = [ |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1353 | ":wasm_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1354 | ":psimd_ukernels", |
| 1355 | ], |
| 1356 | x86_deps = [ |
| 1357 | ":psimd_ukernels", |
| 1358 | ":sse2_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1359 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1360 | ":avx_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1361 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1362 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1363 | ":avx512f_ukernels", |
| 1364 | ], |
| 1365 | ) |
| 1366 | |
| 1367 | xnnpack_cc_library( |
| 1368 | name = "im2col", |
| 1369 | srcs = ["src/im2col.c"], |
| 1370 | hdrs = [ |
| 1371 | "src/xnnpack/common.h", |
| 1372 | "src/xnnpack/im2col.h", |
| 1373 | ], |
| 1374 | copts = xnnpack_std_copts(), |
| 1375 | ) |
| 1376 | |
| 1377 | xnnpack_cc_library( |
| 1378 | name = "indirection", |
| 1379 | srcs = ["src/indirection.c"], |
| 1380 | hdrs = INTERNAL_HDRS, |
| 1381 | copts = xnnpack_std_copts(), |
| 1382 | deps = [ |
| 1383 | "@FP16", |
| 1384 | "@FXdiv", |
| 1385 | "@pthreadpool", |
| 1386 | ], |
| 1387 | ) |
| 1388 | |
| 1389 | xnnpack_cc_library( |
| 1390 | name = "operator_run", |
| 1391 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1392 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 1393 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1394 | # Wrappers for multi-pass microkernels use VLAs for temporary buffers. |
| 1395 | "-Wno-vla", |
| 1396 | ], |
| 1397 | deps = [ |
| 1398 | "@FP16", |
| 1399 | "@FXdiv", |
| 1400 | "@clog", |
| 1401 | "@pthreadpool", |
| 1402 | ], |
| 1403 | ) |
| 1404 | |
| 1405 | cc_library( |
| 1406 | name = "enable_assembly", |
| 1407 | defines = select({ |
| 1408 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 1409 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 1410 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1411 | }), |
| 1412 | ) |
| 1413 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1414 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1415 | name = "operators", |
| 1416 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1417 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1418 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1419 | ], |
| 1420 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1421 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1422 | "-Isrc", |
| 1423 | "-Iinclude", |
| 1424 | ] + select({ |
| 1425 | ":debug_build": [], |
| 1426 | "//conditions:default": xnnpack_min_size_copts(), |
| 1427 | }), |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1428 | wasm_srcs = ["src/wasm-stubs.c"], |
| 1429 | wasmsimd_srcs = ["src/wasm-stubs.c"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1430 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1431 | ":indirection", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1432 | "@FP16", |
| 1433 | "@FXdiv", |
| 1434 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1435 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1436 | ], |
| 1437 | ) |
| 1438 | |
| 1439 | cc_library( |
| 1440 | name = "XNNPACK", |
| 1441 | srcs = [ |
| 1442 | "src/init.c", |
| 1443 | ], |
| 1444 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1445 | "-Isrc", |
| 1446 | "-Iinclude", |
| 1447 | ] + select({ |
| 1448 | ":debug_build": [], |
| 1449 | "//conditions:default": xnnpack_min_size_copts(), |
| 1450 | }), |
| 1451 | includes = ["include"], |
| 1452 | linkstatic = True, |
| 1453 | textual_hdrs = ["include/xnnpack.h"], |
| 1454 | visibility = xnnpack_visibility(), |
| 1455 | deps = [ |
| 1456 | ":enable_assembly", |
| 1457 | ":ukernels", |
| 1458 | ":operator_run", |
| 1459 | ":operators", |
| 1460 | "@clog", |
| 1461 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 1462 | ] + select({ |
| 1463 | ":emscripten": [], |
| 1464 | "//conditions:default": ["@cpuinfo"], |
| 1465 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1466 | ) |
| 1467 | |
| 1468 | cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1469 | name = "xnnpack_operators_nhwc_f32", |
| 1470 | srcs = [ |
| 1471 | "src/init.c", |
| 1472 | ], |
| 1473 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1474 | "-Isrc", |
| 1475 | "-Iinclude", |
| 1476 | ] + select({ |
| 1477 | ":debug_build": [], |
| 1478 | "//conditions:default": xnnpack_min_size_copts(), |
| 1479 | }), |
| 1480 | defines = [ |
| 1481 | "XNN_NO_Q8_OPERATORS", |
| 1482 | "XNN_NO_U8_OPERATORS", |
| 1483 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1484 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1485 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1486 | includes = ["include"], |
| 1487 | linkstatic = True, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1488 | textual_hdrs = ["include/xnnpack.h"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1489 | visibility = xnnpack_visibility(), |
| 1490 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1491 | ":enable_assembly", |
| 1492 | ":ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1493 | ":operator_run", |
| 1494 | ":operators", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1495 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1496 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1497 | ] + select({ |
| 1498 | ":emscripten": [], |
| 1499 | "//conditions:default": ["@cpuinfo"], |
| 1500 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1501 | ) |
| 1502 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1503 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1504 | name = "bench_utils", |
| 1505 | srcs = ["bench/utils.cc"], |
| 1506 | hdrs = ["bench/utils.h"], |
| 1507 | copts = ["-Wno-unused-result"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 1508 | deps = [ |
| 1509 | "@com_google_benchmark//:benchmark", |
| 1510 | "@cpuinfo", |
| 1511 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1512 | ) |
| 1513 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1514 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1515 | |
| 1516 | xnnpack_benchmark( |
| 1517 | name = "q8_gemm_bench", |
| 1518 | srcs = [ |
| 1519 | "bench/gemm.h", |
| 1520 | "bench/q8-gemm.cc", |
| 1521 | "src/xnnpack/AlignedAllocator.h", |
| 1522 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1523 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 1524 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
| 1525 | ) |
| 1526 | |
| 1527 | xnnpack_benchmark( |
| 1528 | name = "f16_gemm_bench", |
| 1529 | srcs = [ |
| 1530 | "bench/f16-gemm.cc", |
| 1531 | "bench/gemm.h", |
| 1532 | "src/xnnpack/AlignedAllocator.h", |
| 1533 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1534 | copts = ["-Wno-unused-function"], |
| 1535 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1536 | ) |
| 1537 | |
| 1538 | xnnpack_benchmark( |
| 1539 | name = "f32_igemm_bench", |
| 1540 | srcs = [ |
| 1541 | "bench/f32-igemm.cc", |
| 1542 | "bench/conv.h", |
| 1543 | "src/xnnpack/AlignedAllocator.h", |
| 1544 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1545 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1546 | ) |
| 1547 | |
| 1548 | xnnpack_benchmark( |
| 1549 | name = "f32_conv_hwc_bench", |
| 1550 | srcs = [ |
| 1551 | "bench/f32-conv-hwc.cc", |
| 1552 | "bench/dconv.h", |
| 1553 | "src/xnnpack/AlignedAllocator.h", |
| 1554 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1555 | copts = ["-Wno-unused-function"], |
| 1556 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1557 | ) |
| 1558 | |
| 1559 | xnnpack_benchmark( |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 1560 | name = "f32_conv_hwc2spchw_bench", |
| 1561 | srcs = [ |
| 1562 | "bench/f32-conv-hwc2spchw.cc", |
| 1563 | "bench/dconv.h", |
| 1564 | "src/xnnpack/AlignedAllocator.h", |
| 1565 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1566 | copts = ["-Wno-unused-function"], |
| 1567 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1568 | ) |
| 1569 | |
| 1570 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1571 | name = "f32_dwconv_bench", |
| 1572 | srcs = [ |
| 1573 | "bench/f32-dwconv.cc", |
| 1574 | "bench/dwconv.h", |
| 1575 | "src/xnnpack/AlignedAllocator.h", |
| 1576 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1577 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 1578 | ) |
| 1579 | |
| 1580 | xnnpack_benchmark( |
| 1581 | name = "f32_dwconv_spchw_bench", |
| 1582 | srcs = [ |
| 1583 | "bench/f32-dwconv-spchw.cc", |
| 1584 | "bench/dwconv.h", |
| 1585 | "src/xnnpack/AlignedAllocator.h", |
| 1586 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1587 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 1588 | ) |
| 1589 | |
| 1590 | xnnpack_benchmark( |
| 1591 | name = "f32_gemm_bench", |
| 1592 | srcs = [ |
| 1593 | "bench/f32-gemm.cc", |
| 1594 | "bench/gemm.h", |
| 1595 | "src/xnnpack/AlignedAllocator.h", |
| 1596 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1597 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(), |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1598 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1599 | ) |
| 1600 | |
| 1601 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame^] | 1602 | name = "f32_raddexpminusmax_bench", |
| 1603 | srcs = [ |
| 1604 | "bench/f32-raddexpminusmax.cc", |
| 1605 | "src/xnnpack/AlignedAllocator.h", |
| 1606 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1607 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1608 | ) |
| 1609 | |
| 1610 | xnnpack_benchmark( |
| 1611 | name = "f32_raddextexp_bench", |
| 1612 | srcs = [ |
| 1613 | "bench/f32-raddextexp.cc", |
| 1614 | "src/xnnpack/AlignedAllocator.h", |
| 1615 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1616 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1617 | ) |
| 1618 | |
| 1619 | xnnpack_benchmark( |
| 1620 | name = "f32_raddstoreexpminusmax_bench", |
| 1621 | srcs = [ |
| 1622 | "bench/f32-raddstoreexpminusmax.cc", |
| 1623 | "src/xnnpack/AlignedAllocator.h", |
| 1624 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1625 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1626 | ) |
| 1627 | |
| 1628 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1629 | name = "f32_rmax_bench", |
| 1630 | srcs = [ |
| 1631 | "bench/f32-rmax.cc", |
| 1632 | "src/xnnpack/AlignedAllocator.h", |
| 1633 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1634 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1635 | ) |
| 1636 | |
| 1637 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 1638 | name = "f32_sigmoid_bench", |
| 1639 | srcs = [ |
| 1640 | "bench/f32-sigmoid.cc", |
| 1641 | "src/xnnpack/AlignedAllocator.h", |
| 1642 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1643 | copts = ["-Wno-unused-function"], |
| 1644 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1645 | ) |
| 1646 | |
| 1647 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1648 | name = "f32_spmm_bench", |
| 1649 | srcs = [ |
| 1650 | "bench/f32-spmm.cc", |
| 1651 | "bench/gemm.h", |
| 1652 | "src/xnnpack/AlignedAllocator.h", |
| 1653 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1654 | copts = ["-Wno-unused-function"], |
| 1655 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1656 | ) |
| 1657 | |
| 1658 | xnnpack_benchmark( |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 1659 | name = "f32_softargmax_bench", |
| 1660 | srcs = [ |
| 1661 | "bench/f32-softargmax.cc", |
| 1662 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1663 | copts = ["-Wno-unused-function"], |
| 1664 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1665 | ) |
| 1666 | |
| 1667 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame^] | 1668 | name = "f32_vscaleexpminusmax_bench", |
| 1669 | srcs = [ |
| 1670 | "bench/f32-vscaleexpminusmax.cc", |
| 1671 | "src/xnnpack/AlignedAllocator.h", |
| 1672 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1673 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1674 | ) |
| 1675 | |
| 1676 | xnnpack_benchmark( |
| 1677 | name = "f32_vscaleextexp_bench", |
| 1678 | srcs = [ |
| 1679 | "bench/f32-vscaleextexp.cc", |
| 1680 | "src/xnnpack/AlignedAllocator.h", |
| 1681 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1682 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1683 | ) |
| 1684 | |
| 1685 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1686 | name = "f32_im2col_gemm_bench", |
| 1687 | srcs = [ |
| 1688 | "bench/f32-im2col-gemm.cc", |
| 1689 | "bench/conv.h", |
| 1690 | "src/xnnpack/AlignedAllocator.h", |
| 1691 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1692 | deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"], |
| 1693 | ) |
| 1694 | |
| 1695 | ########################### Benchmarks for operators ########################### |
| 1696 | |
| 1697 | xnnpack_benchmark( |
| 1698 | name = "add_bench", |
| 1699 | srcs = ["bench/add.cc"], |
| 1700 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1701 | ) |
| 1702 | |
| 1703 | xnnpack_benchmark( |
| 1704 | name = "average_pooling_bench", |
| 1705 | srcs = ["bench/average-pooling.cc"], |
| 1706 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1707 | ) |
| 1708 | |
| 1709 | xnnpack_benchmark( |
| 1710 | name = "channel_shuffle_bench", |
| 1711 | srcs = ["bench/channel-shuffle.cc"], |
| 1712 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1713 | ) |
| 1714 | |
| 1715 | xnnpack_benchmark( |
| 1716 | name = "convolution_bench", |
| 1717 | srcs = ["bench/convolution.cc"], |
| 1718 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
| 1719 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
| 1720 | ) |
| 1721 | |
| 1722 | xnnpack_benchmark( |
| 1723 | name = "deconvolution_bench", |
| 1724 | srcs = ["bench/deconvolution.cc"], |
| 1725 | copts = xnnpack_optional_tflite_copts(), |
| 1726 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 1727 | ) |
| 1728 | |
| 1729 | xnnpack_benchmark( |
| 1730 | name = "global_average_pooling_bench", |
| 1731 | srcs = ["bench/global-average-pooling.cc"], |
| 1732 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1733 | ) |
| 1734 | |
| 1735 | xnnpack_benchmark( |
| 1736 | name = "max_pooling_bench", |
| 1737 | srcs = ["bench/max-pooling.cc"], |
| 1738 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1739 | ) |
| 1740 | |
| 1741 | xnnpack_benchmark( |
| 1742 | name = "sigmoid_bench", |
| 1743 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 1744 | copts = xnnpack_optional_tflite_copts(), |
| 1745 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1746 | ) |
| 1747 | |
| 1748 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 1749 | name = "prelu_bench", |
| 1750 | srcs = ["bench/prelu.cc"], |
| 1751 | copts = xnnpack_optional_tflite_copts(), |
| 1752 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 1753 | ) |
| 1754 | |
| 1755 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1756 | name = "softargmax_bench", |
| 1757 | srcs = ["bench/softargmax.cc"], |
| 1758 | deps = OPERATOR_BENCHMARK_DEPS, |
| 1759 | ) |
| 1760 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1761 | ############################# End-to-end benchmarks ############################ |
| 1762 | |
| 1763 | cc_library( |
| 1764 | name = "mobilenet_v1", |
| 1765 | srcs = ["models/mobilenet-v1.cc"], |
| 1766 | hdrs = ["models/models.h"], |
| 1767 | linkstatic = True, |
| 1768 | deps = [ |
| 1769 | ":XNNPACK", |
| 1770 | "@pthreadpool", |
| 1771 | ], |
| 1772 | ) |
| 1773 | |
| 1774 | cc_library( |
| 1775 | name = "mobilenet_v2", |
| 1776 | srcs = ["models/mobilenet-v2.cc"], |
| 1777 | hdrs = ["models/models.h"], |
| 1778 | linkstatic = True, |
| 1779 | deps = [ |
| 1780 | ":XNNPACK", |
| 1781 | "@pthreadpool", |
| 1782 | ], |
| 1783 | ) |
| 1784 | |
| 1785 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 1786 | name = "f32_dwconv_e2e_bench", |
| 1787 | srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS, |
| 1788 | copts = ["-Wno-unused-function"], |
| 1789 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 1790 | ":XNNPACK", |
| 1791 | ":mobilenet_v1", |
| 1792 | ":mobilenet_v2", |
| 1793 | ], |
| 1794 | ) |
| 1795 | |
| 1796 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 1797 | name = "f32_gemm_e2e_bench", |
| 1798 | srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS, |
| 1799 | copts = ["-Wno-unused-function"], |
| 1800 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 1801 | ":XNNPACK", |
| 1802 | ":mobilenet_v1", |
| 1803 | ":mobilenet_v2", |
| 1804 | ], |
| 1805 | ) |
| 1806 | |
| 1807 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1808 | name = "end2end_bench", |
| 1809 | srcs = ["bench/end2end.cc"], |
| 1810 | deps = [ |
| 1811 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 1812 | ":bench_utils", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 1813 | ":mobilenet_v1", |
| 1814 | ":mobilenet_v2", |
| 1815 | "@pthreadpool", |
| 1816 | ], |
| 1817 | ) |
| 1818 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1819 | #################### Accuracy evaluation for math functions #################### |
| 1820 | |
| 1821 | xnnpack_benchmark( |
| 1822 | name = "f32_exp_eval", |
| 1823 | srcs = [ |
| 1824 | "eval/f32-exp.cc", |
| 1825 | "src/xnnpack/AlignedAllocator.h", |
| 1826 | ] + ACCURACY_EVAL_HDRS, |
| 1827 | deps = ACCURACY_EVAL_DEPS, |
| 1828 | ) |
| 1829 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 1830 | xnnpack_benchmark( |
| 1831 | name = "f32_expminus_eval", |
| 1832 | srcs = [ |
| 1833 | "eval/f32-expminus.cc", |
| 1834 | "src/xnnpack/AlignedAllocator.h", |
| 1835 | ] + ACCURACY_EVAL_HDRS, |
| 1836 | deps = ACCURACY_EVAL_DEPS, |
| 1837 | ) |
| 1838 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1839 | xnnpack_benchmark( |
| 1840 | name = "f32_extexp_eval", |
| 1841 | srcs = [ |
| 1842 | "eval/f32-extexp.cc", |
| 1843 | "src/xnnpack/AlignedAllocator.h", |
| 1844 | ] + ACCURACY_EVAL_HDRS, |
| 1845 | deps = ACCURACY_EVAL_DEPS, |
| 1846 | ) |
| 1847 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1848 | xnnpack_benchmark( |
| 1849 | name = "f32_sigmoid_eval", |
| 1850 | srcs = [ |
| 1851 | "eval/f32-sigmoid.cc", |
| 1852 | "src/xnnpack/AlignedAllocator.h", |
| 1853 | ] + ACCURACY_EVAL_HDRS, |
| 1854 | deps = ACCURACY_EVAL_DEPS, |
| 1855 | ) |
| 1856 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1857 | ######################### Unit tests for micro-kernels ######################### |
| 1858 | |
| 1859 | xnnpack_unit_test( |
| 1860 | name = "f16_gemm_test", |
| 1861 | srcs = [ |
| 1862 | "test/f16-gemm.cc", |
| 1863 | "test/gemm-microkernel-tester.h", |
| 1864 | "src/xnnpack/AlignedAllocator.h", |
| 1865 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1866 | deps = MICROKERNEL_TEST_DEPS, |
| 1867 | ) |
| 1868 | |
| 1869 | xnnpack_unit_test( |
| 1870 | name = "f32_argmaxpool_test", |
| 1871 | srcs = [ |
| 1872 | "test/f32-argmaxpool.cc", |
| 1873 | "test/argmaxpool-microkernel-tester.h", |
| 1874 | "src/xnnpack/AlignedAllocator.h", |
| 1875 | ] + MICROKERNEL_TEST_HDRS, |
| 1876 | deps = MICROKERNEL_TEST_DEPS, |
| 1877 | ) |
| 1878 | |
| 1879 | xnnpack_unit_test( |
| 1880 | name = "f32_avgpool_test", |
| 1881 | srcs = [ |
| 1882 | "test/f32-avgpool.cc", |
| 1883 | "test/avgpool-microkernel-tester.h", |
| 1884 | "src/xnnpack/AlignedAllocator.h", |
| 1885 | ] + MICROKERNEL_TEST_HDRS, |
| 1886 | deps = MICROKERNEL_TEST_DEPS, |
| 1887 | ) |
| 1888 | |
| 1889 | xnnpack_unit_test( |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 1890 | name = "f32_bilinear_test", |
| 1891 | srcs = [ |
| 1892 | "test/f32-bilinear.cc", |
| 1893 | "test/bilinear-microkernel-tester.h", |
| 1894 | "src/xnnpack/AlignedAllocator.h", |
| 1895 | ] + MICROKERNEL_TEST_HDRS, |
| 1896 | deps = MICROKERNEL_TEST_DEPS, |
| 1897 | ) |
| 1898 | |
| 1899 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1900 | name = "f32_clamp_test", |
| 1901 | srcs = [ |
| 1902 | "test/f32-clamp.cc", |
| 1903 | "test/clamp-microkernel-tester.h", |
| 1904 | ] + MICROKERNEL_TEST_HDRS, |
| 1905 | deps = MICROKERNEL_TEST_DEPS, |
| 1906 | ) |
| 1907 | |
| 1908 | xnnpack_unit_test( |
| 1909 | name = "f32_igemm_test", |
| 1910 | srcs = [ |
| 1911 | "test/f32-igemm.cc", |
| 1912 | "test/gemm-microkernel-tester.h", |
| 1913 | "src/xnnpack/AlignedAllocator.h", |
| 1914 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1915 | deps = MICROKERNEL_TEST_DEPS, |
| 1916 | ) |
| 1917 | |
| 1918 | xnnpack_unit_test( |
| 1919 | name = "f32_conv_hwc_test", |
| 1920 | srcs = [ |
| 1921 | "test/f32-conv-hwc.cc", |
| 1922 | "test/conv-hwc-microkernel-tester.h", |
| 1923 | "src/xnnpack/AlignedAllocator.h", |
| 1924 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1925 | deps = MICROKERNEL_TEST_DEPS, |
| 1926 | ) |
| 1927 | |
| 1928 | xnnpack_unit_test( |
| 1929 | name = "f32_conv_hwc2spchw_test", |
| 1930 | srcs = [ |
| 1931 | "test/f32-conv-hwc2spchw.cc", |
| 1932 | "test/conv-hwc2spchw-microkernel-tester.h", |
| 1933 | "src/xnnpack/AlignedAllocator.h", |
| 1934 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1935 | deps = MICROKERNEL_TEST_DEPS, |
| 1936 | ) |
| 1937 | |
| 1938 | xnnpack_unit_test( |
| 1939 | name = "f32_dwconv_test", |
| 1940 | srcs = [ |
| 1941 | "test/f32-dwconv.cc", |
| 1942 | "test/dwconv-microkernel-tester.h", |
| 1943 | "src/xnnpack/AlignedAllocator.h", |
| 1944 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1945 | deps = MICROKERNEL_TEST_DEPS, |
| 1946 | ) |
| 1947 | |
| 1948 | xnnpack_unit_test( |
| 1949 | name = "f32_dwconv_spchw_test", |
| 1950 | srcs = [ |
| 1951 | "test/f32-dwconv-spchw.cc", |
| 1952 | "test/dwconv-spchw-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 = "f32_gavgpool_test", |
| 1960 | srcs = [ |
| 1961 | "test/f32-gavgpool.cc", |
| 1962 | "test/gavgpool-microkernel-tester.h", |
| 1963 | "src/xnnpack/AlignedAllocator.h", |
| 1964 | ] + MICROKERNEL_TEST_HDRS, |
| 1965 | deps = MICROKERNEL_TEST_DEPS, |
| 1966 | ) |
| 1967 | |
| 1968 | xnnpack_unit_test( |
| 1969 | name = "f32_gavgpool_spchw_test", |
| 1970 | srcs = [ |
| 1971 | "test/f32-gavgpool-spchw.cc", |
| 1972 | "test/gavgpool-spchw-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 = "f32_gemm_test", |
| 1980 | srcs = [ |
| 1981 | "test/f32-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 = "f32_gemminc_test", |
| 1990 | srcs = [ |
| 1991 | "test/f32-gemminc.cc", |
| 1992 | "test/gemm-microkernel-tester.h", |
| 1993 | "src/xnnpack/AlignedAllocator.h", |
| 1994 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 1995 | deps = MICROKERNEL_TEST_DEPS, |
| 1996 | ) |
| 1997 | |
| 1998 | xnnpack_unit_test( |
| 1999 | name = "f32_hswish_test", |
| 2000 | srcs = [ |
| 2001 | "test/f32-hswish.cc", |
| 2002 | "test/hswish-microkernel-tester.h", |
| 2003 | ] + MICROKERNEL_TEST_HDRS, |
| 2004 | deps = MICROKERNEL_TEST_DEPS, |
| 2005 | ) |
| 2006 | |
| 2007 | xnnpack_unit_test( |
| 2008 | name = "f32_maxpool_test", |
| 2009 | srcs = [ |
| 2010 | "test/f32-maxpool.cc", |
| 2011 | "test/maxpool-microkernel-tester.h", |
| 2012 | ] + MICROKERNEL_TEST_HDRS, |
| 2013 | deps = MICROKERNEL_TEST_DEPS, |
| 2014 | ) |
| 2015 | |
| 2016 | xnnpack_unit_test( |
| 2017 | name = "f32_pavgpool_test", |
| 2018 | srcs = [ |
| 2019 | "test/f32-pavgpool.cc", |
| 2020 | "test/avgpool-microkernel-tester.h", |
| 2021 | "src/xnnpack/AlignedAllocator.h", |
| 2022 | ] + MICROKERNEL_TEST_HDRS, |
| 2023 | deps = MICROKERNEL_TEST_DEPS, |
| 2024 | ) |
| 2025 | |
| 2026 | xnnpack_unit_test( |
| 2027 | name = "f32_ppmm_test", |
| 2028 | srcs = [ |
| 2029 | "test/f32-ppmm.cc", |
| 2030 | "test/gemm-microkernel-tester.h", |
| 2031 | "src/xnnpack/AlignedAllocator.h", |
| 2032 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2033 | deps = MICROKERNEL_TEST_DEPS, |
| 2034 | ) |
| 2035 | |
| 2036 | xnnpack_unit_test( |
| 2037 | name = "f32_prelu_test", |
| 2038 | srcs = [ |
| 2039 | "test/f32-prelu.cc", |
| 2040 | "test/prelu-microkernel-tester.h", |
| 2041 | "src/xnnpack/AlignedAllocator.h", |
| 2042 | ] + MICROKERNEL_TEST_HDRS, |
| 2043 | deps = MICROKERNEL_TEST_DEPS, |
| 2044 | ) |
| 2045 | |
| 2046 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2047 | name = "f32_raddexpminusmax_test", |
| 2048 | srcs = [ |
| 2049 | "test/f32-raddexpminusmax.cc", |
| 2050 | "test/raddexpminusmax-microkernel-tester.h", |
| 2051 | ] + MICROKERNEL_TEST_HDRS, |
| 2052 | deps = MICROKERNEL_TEST_DEPS, |
| 2053 | ) |
| 2054 | |
| 2055 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2056 | name = "f32_raddextexp_test", |
| 2057 | srcs = [ |
| 2058 | "test/f32-raddextexp.cc", |
| 2059 | "test/raddextexp-microkernel-tester.h", |
| 2060 | ] + MICROKERNEL_TEST_HDRS, |
| 2061 | deps = MICROKERNEL_TEST_DEPS, |
| 2062 | ) |
| 2063 | |
| 2064 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2065 | name = "f32_raddstoreexpminusmax_test", |
| 2066 | srcs = [ |
| 2067 | "test/f32-raddstoreexpminusmax.cc", |
| 2068 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 2069 | ] + MICROKERNEL_TEST_HDRS, |
| 2070 | deps = MICROKERNEL_TEST_DEPS, |
| 2071 | ) |
| 2072 | |
| 2073 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2074 | name = "f32_rmax_test", |
| 2075 | srcs = [ |
| 2076 | "test/f32-rmax.cc", |
| 2077 | "test/rmax-microkernel-tester.h", |
| 2078 | ] + MICROKERNEL_TEST_HDRS, |
| 2079 | deps = MICROKERNEL_TEST_DEPS, |
| 2080 | ) |
| 2081 | |
| 2082 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2083 | name = "f32_sigmoid_test", |
| 2084 | srcs = [ |
| 2085 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2086 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2087 | ] + MICROKERNEL_TEST_HDRS, |
| 2088 | deps = MICROKERNEL_TEST_DEPS, |
| 2089 | ) |
| 2090 | |
| 2091 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2092 | name = "f32_spmm_test", |
| 2093 | srcs = [ |
| 2094 | "test/f32-spmm.cc", |
| 2095 | "test/spmm-microkernel-tester.h", |
| 2096 | "src/xnnpack/AlignedAllocator.h", |
| 2097 | ] + MICROKERNEL_TEST_HDRS, |
| 2098 | deps = MICROKERNEL_TEST_DEPS, |
| 2099 | ) |
| 2100 | |
| 2101 | xnnpack_unit_test( |
| 2102 | name = "f32_vadd_test", |
| 2103 | srcs = [ |
| 2104 | "test/f32-vadd.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2105 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2106 | ] + MICROKERNEL_TEST_HDRS, |
| 2107 | deps = MICROKERNEL_TEST_DEPS, |
| 2108 | ) |
| 2109 | |
| 2110 | xnnpack_unit_test( |
| 2111 | name = "f32_vaddc_test", |
| 2112 | srcs = [ |
| 2113 | "test/f32-vaddc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2114 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2115 | ] + MICROKERNEL_TEST_HDRS, |
| 2116 | deps = MICROKERNEL_TEST_DEPS, |
| 2117 | ) |
| 2118 | |
| 2119 | xnnpack_unit_test( |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 2120 | name = "f32_vdiv_test", |
| 2121 | srcs = [ |
| 2122 | "test/f32-vdiv.cc", |
| 2123 | "test/vbinary-microkernel-tester.h", |
| 2124 | ] + MICROKERNEL_TEST_HDRS, |
| 2125 | deps = MICROKERNEL_TEST_DEPS, |
| 2126 | ) |
| 2127 | |
| 2128 | xnnpack_unit_test( |
| 2129 | name = "f32_vdivc_test", |
| 2130 | srcs = [ |
| 2131 | "test/f32-vdivc.cc", |
| 2132 | "test/vbinaryc-microkernel-tester.h", |
| 2133 | ] + MICROKERNEL_TEST_HDRS, |
| 2134 | deps = MICROKERNEL_TEST_DEPS, |
| 2135 | ) |
| 2136 | |
| 2137 | xnnpack_unit_test( |
| 2138 | name = "f32_vrdivc_test", |
| 2139 | srcs = [ |
| 2140 | "test/f32-vrdivc.cc", |
| 2141 | "test/vbinaryc-microkernel-tester.h", |
| 2142 | ] + MICROKERNEL_TEST_HDRS, |
| 2143 | deps = MICROKERNEL_TEST_DEPS, |
| 2144 | ) |
| 2145 | |
| 2146 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 2147 | name = "f32_vmax_test", |
| 2148 | srcs = [ |
| 2149 | "test/f32-vmax.cc", |
| 2150 | "test/vbinary-microkernel-tester.h", |
| 2151 | ] + MICROKERNEL_TEST_HDRS, |
| 2152 | deps = MICROKERNEL_TEST_DEPS, |
| 2153 | ) |
| 2154 | |
| 2155 | xnnpack_unit_test( |
| 2156 | name = "f32_vmaxc_test", |
| 2157 | srcs = [ |
| 2158 | "test/f32-vmaxc.cc", |
| 2159 | "test/vbinaryc-microkernel-tester.h", |
| 2160 | ] + MICROKERNEL_TEST_HDRS, |
| 2161 | deps = MICROKERNEL_TEST_DEPS, |
| 2162 | ) |
| 2163 | |
| 2164 | xnnpack_unit_test( |
| 2165 | name = "f32_vmin_test", |
| 2166 | srcs = [ |
| 2167 | "test/f32-vmin.cc", |
| 2168 | "test/vbinary-microkernel-tester.h", |
| 2169 | ] + MICROKERNEL_TEST_HDRS, |
| 2170 | deps = MICROKERNEL_TEST_DEPS, |
| 2171 | ) |
| 2172 | |
| 2173 | xnnpack_unit_test( |
| 2174 | name = "f32_vminc_test", |
| 2175 | srcs = [ |
| 2176 | "test/f32-vminc.cc", |
| 2177 | "test/vbinaryc-microkernel-tester.h", |
| 2178 | ] + MICROKERNEL_TEST_HDRS, |
| 2179 | deps = MICROKERNEL_TEST_DEPS, |
| 2180 | ) |
| 2181 | |
| 2182 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2183 | name = "f32_vmul_test", |
| 2184 | srcs = [ |
| 2185 | "test/f32-vmul.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2186 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2187 | ] + MICROKERNEL_TEST_HDRS, |
| 2188 | deps = MICROKERNEL_TEST_DEPS, |
| 2189 | ) |
| 2190 | |
| 2191 | xnnpack_unit_test( |
| 2192 | name = "f32_vmulc_test", |
| 2193 | srcs = [ |
| 2194 | "test/f32-vmulc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2195 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2196 | ] + MICROKERNEL_TEST_HDRS, |
| 2197 | deps = MICROKERNEL_TEST_DEPS, |
| 2198 | ) |
| 2199 | |
| 2200 | xnnpack_unit_test( |
| 2201 | name = "f32_vmulcaddc_test", |
| 2202 | srcs = [ |
| 2203 | "test/f32-vmulcaddc.cc", |
| 2204 | "test/vmulcaddc-microkernel-tester.h", |
| 2205 | "src/xnnpack/AlignedAllocator.h", |
| 2206 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2207 | deps = MICROKERNEL_TEST_DEPS, |
| 2208 | ) |
| 2209 | |
| 2210 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2211 | name = "f32_vscale_test", |
| 2212 | srcs = [ |
| 2213 | "test/f32-vscale.cc", |
| 2214 | "test/vscale-microkernel-tester.h", |
| 2215 | ] + MICROKERNEL_TEST_HDRS, |
| 2216 | deps = MICROKERNEL_TEST_DEPS, |
| 2217 | ) |
| 2218 | |
| 2219 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2220 | name = "f32_vscaleexpminusmax_test", |
| 2221 | srcs = [ |
| 2222 | "test/f32-vscaleexpminusmax.cc", |
| 2223 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 2224 | ] + MICROKERNEL_TEST_HDRS, |
| 2225 | deps = MICROKERNEL_TEST_DEPS, |
| 2226 | ) |
| 2227 | |
| 2228 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2229 | name = "f32_vscaleextexp_test", |
| 2230 | srcs = [ |
| 2231 | "test/f32-vscaleextexp.cc", |
| 2232 | "test/vscaleextexp-microkernel-tester.h", |
| 2233 | ] + MICROKERNEL_TEST_HDRS, |
| 2234 | deps = MICROKERNEL_TEST_DEPS, |
| 2235 | ) |
| 2236 | |
| 2237 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2238 | name = "f32_vsub_test", |
| 2239 | srcs = [ |
| 2240 | "test/f32-vsub.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2241 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2242 | ] + MICROKERNEL_TEST_HDRS, |
| 2243 | deps = MICROKERNEL_TEST_DEPS, |
| 2244 | ) |
| 2245 | |
| 2246 | xnnpack_unit_test( |
| 2247 | name = "f32_vsubc_test", |
| 2248 | srcs = [ |
| 2249 | "test/f32-vsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2250 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2251 | ] + MICROKERNEL_TEST_HDRS, |
| 2252 | deps = MICROKERNEL_TEST_DEPS, |
| 2253 | ) |
| 2254 | |
| 2255 | xnnpack_unit_test( |
| 2256 | name = "f32_vrsubc_test", |
| 2257 | srcs = [ |
| 2258 | "test/f32-vrsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2259 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2260 | ] + MICROKERNEL_TEST_HDRS, |
| 2261 | deps = MICROKERNEL_TEST_DEPS, |
| 2262 | ) |
| 2263 | |
| 2264 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2265 | name = "q8_avgpool_test", |
| 2266 | srcs = [ |
| 2267 | "test/q8-avgpool.cc", |
| 2268 | "test/avgpool-microkernel-tester.h", |
| 2269 | "src/xnnpack/AlignedAllocator.h", |
| 2270 | ] + MICROKERNEL_TEST_HDRS, |
| 2271 | deps = MICROKERNEL_TEST_DEPS, |
| 2272 | ) |
| 2273 | |
| 2274 | xnnpack_unit_test( |
| 2275 | name = "q8_igemm_test", |
| 2276 | srcs = [ |
| 2277 | "test/q8-igemm.cc", |
| 2278 | "test/gemm-microkernel-tester.h", |
| 2279 | "src/xnnpack/AlignedAllocator.h", |
| 2280 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2281 | deps = MICROKERNEL_TEST_DEPS, |
| 2282 | ) |
| 2283 | |
| 2284 | xnnpack_unit_test( |
| 2285 | name = "q8_dwconv_test", |
| 2286 | srcs = [ |
| 2287 | "test/q8-dwconv.cc", |
| 2288 | "test/dwconv-microkernel-tester.h", |
| 2289 | "src/xnnpack/AlignedAllocator.h", |
| 2290 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2291 | deps = MICROKERNEL_TEST_DEPS, |
| 2292 | ) |
| 2293 | |
| 2294 | xnnpack_unit_test( |
| 2295 | name = "q8_gavgpool_test", |
| 2296 | srcs = [ |
| 2297 | "test/q8-gavgpool.cc", |
| 2298 | "test/gavgpool-microkernel-tester.h", |
| 2299 | "src/xnnpack/AlignedAllocator.h", |
| 2300 | ] + MICROKERNEL_TEST_HDRS, |
| 2301 | deps = MICROKERNEL_TEST_DEPS, |
| 2302 | ) |
| 2303 | |
| 2304 | xnnpack_unit_test( |
| 2305 | name = "q8_gemm_test", |
| 2306 | srcs = [ |
| 2307 | "test/q8-gemm.cc", |
| 2308 | "test/gemm-microkernel-tester.h", |
| 2309 | "src/xnnpack/AlignedAllocator.h", |
| 2310 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2311 | deps = MICROKERNEL_TEST_DEPS, |
| 2312 | ) |
| 2313 | |
| 2314 | xnnpack_unit_test( |
| 2315 | name = "q8_vadd_test", |
| 2316 | srcs = [ |
| 2317 | "test/q8-vadd.cc", |
| 2318 | "test/vadd-microkernel-tester.h", |
| 2319 | ] + MICROKERNEL_TEST_HDRS, |
| 2320 | deps = MICROKERNEL_TEST_DEPS, |
| 2321 | ) |
| 2322 | |
| 2323 | xnnpack_unit_test( |
| 2324 | name = "u8_clamp_test", |
| 2325 | srcs = [ |
| 2326 | "test/u8-clamp.cc", |
| 2327 | "test/clamp-microkernel-tester.h", |
| 2328 | ] + MICROKERNEL_TEST_HDRS, |
| 2329 | deps = MICROKERNEL_TEST_DEPS, |
| 2330 | ) |
| 2331 | |
| 2332 | xnnpack_unit_test( |
| 2333 | name = "u8_lut32norm_test", |
| 2334 | srcs = [ |
| 2335 | "test/u8-lut32norm.cc", |
| 2336 | "test/lut-norm-microkernel-tester.h", |
| 2337 | ] + MICROKERNEL_TEST_HDRS, |
| 2338 | deps = MICROKERNEL_TEST_DEPS, |
| 2339 | ) |
| 2340 | |
| 2341 | xnnpack_unit_test( |
| 2342 | name = "u8_maxpool_test", |
| 2343 | srcs = [ |
| 2344 | "test/u8-maxpool.cc", |
| 2345 | "test/maxpool-microkernel-tester.h", |
| 2346 | ] + MICROKERNEL_TEST_HDRS, |
| 2347 | deps = MICROKERNEL_TEST_DEPS, |
| 2348 | ) |
| 2349 | |
| 2350 | xnnpack_unit_test( |
| 2351 | name = "u8_rmax_test", |
| 2352 | srcs = [ |
| 2353 | "test/u8-rmax.cc", |
| 2354 | "test/rmax-microkernel-tester.h", |
| 2355 | ] + MICROKERNEL_TEST_HDRS, |
| 2356 | deps = MICROKERNEL_TEST_DEPS, |
| 2357 | ) |
| 2358 | |
| 2359 | xnnpack_unit_test( |
| 2360 | name = "x32_packx_test", |
| 2361 | srcs = [ |
| 2362 | "test/x32-packx.cc", |
| 2363 | "test/pack-microkernel-tester.h", |
| 2364 | "src/xnnpack/AlignedAllocator.h", |
| 2365 | ] + MICROKERNEL_TEST_HDRS, |
| 2366 | deps = MICROKERNEL_TEST_DEPS, |
| 2367 | ) |
| 2368 | |
| 2369 | xnnpack_unit_test( |
| 2370 | name = "x32_pad_test", |
| 2371 | srcs = [ |
| 2372 | "test/x32-pad.cc", |
| 2373 | "test/pad-microkernel-tester.h", |
| 2374 | ] + MICROKERNEL_TEST_HDRS, |
| 2375 | deps = MICROKERNEL_TEST_DEPS, |
| 2376 | ) |
| 2377 | |
| 2378 | xnnpack_unit_test( |
| 2379 | name = "x32_unpool_test", |
| 2380 | srcs = [ |
| 2381 | "test/x32-unpool.cc", |
| 2382 | "test/unpool-microkernel-tester.h", |
| 2383 | ] + MICROKERNEL_TEST_HDRS, |
| 2384 | deps = MICROKERNEL_TEST_DEPS, |
| 2385 | ) |
| 2386 | |
| 2387 | xnnpack_unit_test( |
| 2388 | name = "x32_zip_test", |
| 2389 | srcs = [ |
| 2390 | "test/x32-zip.cc", |
| 2391 | "test/zip-microkernel-tester.h", |
| 2392 | ] + MICROKERNEL_TEST_HDRS, |
| 2393 | deps = MICROKERNEL_TEST_DEPS, |
| 2394 | ) |
| 2395 | |
| 2396 | xnnpack_unit_test( |
| 2397 | name = "x8_lut_test", |
| 2398 | srcs = [ |
| 2399 | "test/x8-lut.cc", |
| 2400 | "test/lut-microkernel-tester.h", |
| 2401 | ] + MICROKERNEL_TEST_HDRS, |
| 2402 | deps = MICROKERNEL_TEST_DEPS, |
| 2403 | ) |
| 2404 | |
| 2405 | xnnpack_unit_test( |
| 2406 | name = "x8_zip_test", |
| 2407 | srcs = [ |
| 2408 | "test/x8-zip.cc", |
| 2409 | "test/zip-microkernel-tester.h", |
| 2410 | ] + MICROKERNEL_TEST_HDRS, |
| 2411 | deps = MICROKERNEL_TEST_DEPS, |
| 2412 | ) |
| 2413 | |
| 2414 | ########################### Size test for the library ########################## |
| 2415 | |
| 2416 | xnnpack_binary( |
| 2417 | name = "size_test", |
| 2418 | srcs = ["test/size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2419 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2420 | ) |
| 2421 | |
| 2422 | ########################### Unit tests for operators ########################### |
| 2423 | |
| 2424 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2425 | name = "add_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2426 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2427 | "test/add-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2428 | "test/add-operator-tester.h", |
| 2429 | ], |
| 2430 | deps = OPERATOR_TEST_DEPS, |
| 2431 | ) |
| 2432 | |
| 2433 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2434 | name = "add_nd_test", |
| 2435 | srcs = [ |
| 2436 | "test/add-nd.cc", |
| 2437 | "test/binary-elementwise-operator-tester.h", |
| 2438 | ], |
| 2439 | deps = OPERATOR_TEST_DEPS, |
| 2440 | ) |
| 2441 | |
| 2442 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2443 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2444 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2445 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2446 | "test/argmax-pooling-operator-tester.h", |
| 2447 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2448 | deps = OPERATOR_TEST_DEPS, |
| 2449 | ) |
| 2450 | |
| 2451 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2452 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2453 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2454 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2455 | "test/average-pooling-operator-tester.h", |
| 2456 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2457 | deps = OPERATOR_TEST_DEPS, |
| 2458 | ) |
| 2459 | |
| 2460 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2461 | name = "channel_pad_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2462 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2463 | "test/channel-pad-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2464 | "test/channel-pad-operator-tester.h", |
| 2465 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2466 | deps = OPERATOR_TEST_DEPS, |
| 2467 | ) |
| 2468 | |
| 2469 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2470 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2471 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2472 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2473 | "test/channel-shuffle-operator-tester.h", |
| 2474 | ], |
| 2475 | deps = OPERATOR_TEST_DEPS, |
| 2476 | ) |
| 2477 | |
| 2478 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2479 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2480 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2481 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2482 | "test/clamp-operator-tester.h", |
| 2483 | ], |
| 2484 | deps = OPERATOR_TEST_DEPS, |
| 2485 | ) |
| 2486 | |
| 2487 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2488 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2489 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2490 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2491 | "test/convolution-operator-tester.h", |
| 2492 | ], |
| 2493 | deps = OPERATOR_TEST_DEPS, |
| 2494 | ) |
| 2495 | |
| 2496 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2497 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2498 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2499 | "test/convolution-nchw.cc", |
| 2500 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2501 | ], |
| 2502 | deps = OPERATOR_TEST_DEPS, |
| 2503 | ) |
| 2504 | |
| 2505 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2506 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2507 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2508 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2509 | "test/deconvolution-operator-tester.h", |
| 2510 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2511 | deps = OPERATOR_TEST_DEPS, |
| 2512 | ) |
| 2513 | |
| 2514 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 2515 | name = "divide_nd_test", |
| 2516 | srcs = [ |
| 2517 | "test/binary-elementwise-operator-tester.h", |
| 2518 | "test/divide-nd.cc", |
| 2519 | ], |
| 2520 | deps = OPERATOR_TEST_DEPS, |
| 2521 | ) |
| 2522 | |
| 2523 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2524 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2525 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2526 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2527 | "test/fully-connected-operator-tester.h", |
| 2528 | ], |
| 2529 | deps = OPERATOR_TEST_DEPS, |
| 2530 | ) |
| 2531 | |
| 2532 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2533 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2534 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2535 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2536 | "test/global-average-pooling-operator-tester.h", |
| 2537 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2538 | deps = OPERATOR_TEST_DEPS, |
| 2539 | ) |
| 2540 | |
| 2541 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2542 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2543 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2544 | "test/global-average-pooling-ncw.cc", |
| 2545 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2546 | ], |
| 2547 | deps = OPERATOR_TEST_DEPS, |
| 2548 | ) |
| 2549 | |
| 2550 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2551 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2552 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2553 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2554 | "test/hardswish-operator-tester.h", |
| 2555 | ], |
| 2556 | deps = OPERATOR_TEST_DEPS, |
| 2557 | ) |
| 2558 | |
| 2559 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2560 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2561 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2562 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2563 | "test/leaky-relu-operator-tester.h", |
| 2564 | ], |
| 2565 | deps = OPERATOR_TEST_DEPS, |
| 2566 | ) |
| 2567 | |
| 2568 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2569 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2570 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2571 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2572 | "test/max-pooling-operator-tester.h", |
| 2573 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2574 | deps = OPERATOR_TEST_DEPS, |
| 2575 | ) |
| 2576 | |
| 2577 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 2578 | name = "maximum_nd_test", |
| 2579 | srcs = [ |
| 2580 | "test/binary-elementwise-operator-tester.h", |
| 2581 | "test/maximum-nd.cc", |
| 2582 | ], |
| 2583 | deps = OPERATOR_TEST_DEPS, |
| 2584 | ) |
| 2585 | |
| 2586 | xnnpack_unit_test( |
| 2587 | name = "minimum_nd_test", |
| 2588 | srcs = [ |
| 2589 | "test/binary-elementwise-operator-tester.h", |
| 2590 | "test/minimum-nd.cc", |
| 2591 | ], |
| 2592 | deps = OPERATOR_TEST_DEPS, |
| 2593 | ) |
| 2594 | |
| 2595 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2596 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 2597 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2598 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2599 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 2600 | ], |
| 2601 | deps = OPERATOR_TEST_DEPS, |
| 2602 | ) |
| 2603 | |
| 2604 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2605 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2606 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2607 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2608 | "test/prelu-operator-tester.h", |
| 2609 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2610 | deps = OPERATOR_TEST_DEPS, |
| 2611 | ) |
| 2612 | |
| 2613 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2614 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 2615 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2616 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 2617 | "test/resize-bilinear-operator-tester.h", |
| 2618 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2619 | deps = OPERATOR_TEST_DEPS, |
| 2620 | ) |
| 2621 | |
| 2622 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2623 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2624 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2625 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2626 | "test/sigmoid-operator-tester.h", |
| 2627 | ], |
| 2628 | deps = OPERATOR_TEST_DEPS, |
| 2629 | ) |
| 2630 | |
| 2631 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2632 | name = "softargmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2633 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2634 | "test/softargmax-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2635 | "test/softargmax-operator-tester.h", |
| 2636 | ], |
| 2637 | deps = OPERATOR_TEST_DEPS, |
| 2638 | ) |
| 2639 | |
| 2640 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 2641 | name = "subtract_nd_test", |
| 2642 | srcs = [ |
| 2643 | "test/binary-elementwise-operator-tester.h", |
| 2644 | "test/subtract-nd.cc", |
| 2645 | ], |
| 2646 | deps = OPERATOR_TEST_DEPS, |
| 2647 | ) |
| 2648 | |
| 2649 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2650 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2651 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2652 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2653 | "test/unpooling-operator-tester.h", |
| 2654 | ], |
| 2655 | deps = OPERATOR_TEST_DEPS, |
| 2656 | ) |
| 2657 | |
| 2658 | ############################# Build configurations ############################# |
| 2659 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2660 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2661 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2662 | name = "xnn_enable_assembly_explicit_true", |
| 2663 | define_values = {"xnn_enable_assembly": "true"}, |
| 2664 | ) |
| 2665 | |
| 2666 | # Disables usage of assembly kernels. |
| 2667 | config_setting( |
| 2668 | name = "xnn_enable_assembly_explicit_false", |
| 2669 | define_values = {"xnn_enable_assembly": "false"}, |
| 2670 | ) |
| 2671 | |
| 2672 | # Builds with -c dbg |
| 2673 | config_setting( |
| 2674 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2675 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2676 | "compilation_mode": "dbg", |
| 2677 | }, |
| 2678 | ) |
| 2679 | |
| 2680 | # Builds with -c opt |
| 2681 | config_setting( |
| 2682 | name = "optimized_build", |
| 2683 | values = { |
| 2684 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2685 | }, |
| 2686 | ) |
| 2687 | |
| 2688 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2689 | name = "linux_k8", |
| 2690 | values = {"cpu": "k8"}, |
| 2691 | ) |
| 2692 | |
| 2693 | config_setting( |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 2694 | name = "linux_aarch64", |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 2695 | values = {"cpu": "aarch64"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 2696 | ) |
| 2697 | |
| 2698 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2699 | name = "android", |
| 2700 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 2701 | ) |
| 2702 | |
| 2703 | config_setting( |
| 2704 | name = "android_armv7", |
| 2705 | values = { |
| 2706 | "crosstool_top": "//external:android/crosstool", |
| 2707 | "cpu": "armeabi-v7a", |
| 2708 | }, |
| 2709 | ) |
| 2710 | |
| 2711 | config_setting( |
| 2712 | name = "android_arm64", |
| 2713 | values = { |
| 2714 | "crosstool_top": "//external:android/crosstool", |
| 2715 | "cpu": "arm64-v8a", |
| 2716 | }, |
| 2717 | ) |
| 2718 | |
| 2719 | config_setting( |
| 2720 | name = "android_x86", |
| 2721 | values = { |
| 2722 | "crosstool_top": "//external:android/crosstool", |
| 2723 | "cpu": "x86", |
| 2724 | }, |
| 2725 | ) |
| 2726 | |
| 2727 | config_setting( |
| 2728 | name = "android_x86_64", |
| 2729 | values = { |
| 2730 | "crosstool_top": "//external:android/crosstool", |
| 2731 | "cpu": "x86_64", |
| 2732 | }, |
| 2733 | ) |
| 2734 | |
| 2735 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 2736 | name = "macos_x86_64", |
| 2737 | values = { |
| 2738 | "apple_platform_type": "macos", |
| 2739 | "cpu": "darwin", |
| 2740 | }, |
| 2741 | ) |
| 2742 | |
| 2743 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2744 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2745 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2746 | ) |
| 2747 | |
| 2748 | config_setting( |
| 2749 | name = "emscripten_wasm", |
| 2750 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2751 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2752 | "cpu": "wasm", |
| 2753 | }, |
| 2754 | ) |
| 2755 | |
| 2756 | config_setting( |
| 2757 | name = "emscripten_wasmsimd", |
| 2758 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2759 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2760 | "cpu": "wasm", |
Marat Dukhan | 8c19e3c | 2019-10-30 12:14:58 -0700 | [diff] [blame] | 2761 | "features": "wasm_simd", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2762 | }, |
| 2763 | ) |
| 2764 | |
| 2765 | config_setting( |
| 2766 | name = "emscripten_asmjs", |
| 2767 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 2768 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2769 | "cpu": "asmjs", |
| 2770 | }, |
| 2771 | ) |