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 | a84e40b | 2019-12-11 15:38:03 -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_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility") |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 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", |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 72 | "src/softmax-nc.c", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 73 | "src/unpooling-nhwc.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 74 | ] |
| 75 | |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 76 | TABLE_SRCS = [ |
| 77 | "src/tables/exp2-k-over-64.c", |
| 78 | "src/tables/exp2-k-over-2048.c", |
| 79 | ] |
| 80 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 81 | SCALAR_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 82 | "src/f32-argmaxpool/4x-scalar-c1.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 83 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 84 | "src/f32-argmaxpool/9x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 85 | "src/f32-avgpool/mp9p8q-scalar.c", |
| 86 | "src/f32-avgpool/up9-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 87 | "src/f32-bilinear/gen/scalar-c1.c", |
| 88 | "src/f32-bilinear/gen/scalar-c2.c", |
| 89 | "src/f32-bilinear/gen/scalar-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 90 | "src/f32-clamp/scalar.c", |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 91 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
Marat Dukhan | 6b7dfae | 2019-12-04 16:00:52 -0800 | [diff] [blame] | 92 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 93 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c", |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 94 | "src/f32-dwconv-spchw/3x3p1-scalar.c", |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 95 | "src/f32-dwconv-spchw/3x3s2p1-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 96 | "src/f32-dwconv-spchw/5x5p2-scalar.c", |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 97 | "src/f32-dwconv-spchw/5x5s2p2-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 98 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 99 | "src/f32-dwconv/gen/up1x25-scalar.c", |
| 100 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 101 | "src/f32-dwconv/gen/up1x4-scalar.c", |
| 102 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 103 | "src/f32-dwconv/gen/up1x9-scalar.c", |
| 104 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 105 | "src/f32-dwconv/gen/up2x25-scalar.c", |
| 106 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 107 | "src/f32-dwconv/gen/up2x4-scalar.c", |
| 108 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 109 | "src/f32-dwconv/gen/up2x9-scalar.c", |
Erich Elsen | 34dc2c0 | 2019-10-16 05:11:41 -0700 | [diff] [blame] | 110 | "src/f32-gavgpool-spchw/scalar-x1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 111 | "src/f32-gavgpool/mp7p7q-scalar.c", |
| 112 | "src/f32-gavgpool/up7-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 113 | "src/f32-gemm/gen-inc/1x4-scalar.c", |
| 114 | "src/f32-gemm/gen-inc/2x4-scalar.c", |
| 115 | "src/f32-gemm/gen-inc/4x4-scalar.c", |
| 116 | "src/f32-gemm/gen/1x4-scalar.c", |
| 117 | "src/f32-gemm/gen/2x4-scalar.c", |
| 118 | "src/f32-gemm/gen/4x2-scalar.c", |
| 119 | "src/f32-gemm/gen/4x4-scalar.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 120 | "src/f32-hswish/gen/scalar-x1.c", |
| 121 | "src/f32-hswish/gen/scalar-x2.c", |
| 122 | "src/f32-hswish/gen/scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 123 | "src/f32-igemm/gen/1x4-scalar.c", |
| 124 | "src/f32-igemm/gen/2x4-scalar.c", |
| 125 | "src/f32-igemm/gen/4x2-scalar.c", |
| 126 | "src/f32-igemm/gen/4x4-scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 127 | "src/f32-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 128 | "src/f32-pavgpool/mp9p8q-scalar.c", |
| 129 | "src/f32-pavgpool/up9-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 130 | "src/f32-ppmm/gen/2x4-scalar.c", |
| 131 | "src/f32-ppmm/gen/3x3-scalar.c", |
| 132 | "src/f32-ppmm/gen/4x2-scalar.c", |
| 133 | "src/f32-ppmm/gen/4x4-scalar.c", |
| 134 | "src/f32-prelu/gen/scalar-2x1.c", |
| 135 | "src/f32-prelu/gen/scalar-2x4.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 136 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
| 137 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
| 138 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
| 139 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
| 140 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 141 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
| 142 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
| 143 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
| 144 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
| 145 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
| 146 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 147 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 148 | "src/f32-rmax/scalar.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 149 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 150 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 151 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
| 152 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 153 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 154 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
| 155 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 156 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 157 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 158 | "src/f32-spmm/gen/1x1-scalar-pipelined.c", |
| 159 | "src/f32-spmm/gen/1x1-scalar.c", |
| 160 | "src/f32-spmm/gen/2x1-scalar-pipelined.c", |
| 161 | "src/f32-spmm/gen/2x1-scalar.c", |
| 162 | "src/f32-spmm/gen/4x1-scalar-pipelined.c", |
| 163 | "src/f32-spmm/gen/4x1-scalar.c", |
| 164 | "src/f32-spmm/gen/8x1-scalar-pipelined.c", |
| 165 | "src/f32-spmm/gen/8x1-scalar.c", |
| 166 | "src/f32-spmm/gen/8x2-scalar.c", |
| 167 | "src/f32-spmm/gen/8x4-scalar.c", |
| 168 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 169 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 170 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
| 171 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 172 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 173 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 174 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 175 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 176 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
| 177 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 178 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 179 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 180 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 181 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 182 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| 183 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 184 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 185 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| 186 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 187 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 188 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| 189 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 190 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 191 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 192 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 193 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 194 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
| 195 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 196 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 197 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 198 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 199 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 200 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 201 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 202 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 203 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
| 204 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 205 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 206 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
| 207 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 208 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 209 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
| 210 | "src/f32-vmulcaddc/gen/c1-scalar-2x.c", |
| 211 | "src/f32-vmulcaddc/gen/c2-scalar-2x.c", |
| 212 | "src/f32-vmulcaddc/gen/c4-scalar-2x.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 213 | "src/math/expminus-scalar-lut2048-p1.c", |
| 214 | "src/math/expminus-scalar-lut64-p2.c", |
| 215 | "src/math/expminus-scalar-p5.c", |
| 216 | "src/math/sigmoid-scalar-lut2048-p1-div.c", |
| 217 | "src/math/sigmoid-scalar-lut64-p2-div.c", |
| 218 | "src/math/sigmoid-scalar-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 219 | "src/q8-avgpool/mp9p8q-scalar.c", |
| 220 | "src/q8-avgpool/up9-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 221 | "src/q8-dwconv/up1x9-scalar.c", |
| 222 | "src/q8-gavgpool/mp7p7q-scalar.c", |
| 223 | "src/q8-gavgpool/up7-scalar.c", |
| 224 | "src/q8-gemm/2x2-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 225 | "src/q8-igemm/2x2-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 226 | "src/q8-vadd/scalar.c", |
| 227 | "src/u8-clamp/scalar.c", |
| 228 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 229 | "src/u8-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 230 | "src/u8-rmax/scalar.c", |
| 231 | "src/x32-packx/x2-scalar.c", |
| 232 | "src/x32-packx/x3-scalar.c", |
| 233 | "src/x32-packx/x4-scalar.c", |
| 234 | "src/x32-pad/x2-scalar.c", |
| 235 | "src/x32-unpool/scalar.c", |
| 236 | "src/x32-zip/x2-scalar.c", |
| 237 | "src/x32-zip/x3-scalar.c", |
| 238 | "src/x32-zip/x4-scalar.c", |
| 239 | "src/x32-zip/xm-scalar.c", |
| 240 | "src/x8-lut/scalar.c", |
| 241 | "src/x8-zip/x2-scalar.c", |
| 242 | "src/x8-zip/x3-scalar.c", |
| 243 | "src/x8-zip/x4-scalar.c", |
| 244 | "src/x8-zip/xm-scalar.c", |
| 245 | ] |
| 246 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 247 | WASM_UKERNELS = [ |
| 248 | "src/f32-avgpool/mp9p8q-wasm.c", |
| 249 | "src/f32-avgpool/up9-wasm.c", |
| 250 | "src/f32-clamp/wasm.c", |
| 251 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 252 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 253 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 254 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 255 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 256 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 257 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 258 | "src/f32-dwconv/gen/up2x25-wasm.c", |
| 259 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 260 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 261 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 262 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| 263 | "src/f32-gavgpool/mp7p7q-wasm.c", |
| 264 | "src/f32-gavgpool/up7-wasm.c", |
| 265 | "src/f32-gemm/gen-inc/1x4-wasm.c", |
| 266 | "src/f32-gemm/gen-inc/2x4-wasm.c", |
| 267 | "src/f32-gemm/gen-inc/4x4-wasm.c", |
| 268 | "src/f32-gemm/gen/1x4-wasm.c", |
| 269 | "src/f32-gemm/gen/2x4-wasm.c", |
| 270 | "src/f32-gemm/gen/4x2-wasm.c", |
| 271 | "src/f32-gemm/gen/4x4-wasm.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 272 | "src/f32-hswish/gen/wasm-x1.c", |
| 273 | "src/f32-hswish/gen/wasm-x2.c", |
| 274 | "src/f32-hswish/gen/wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 275 | "src/f32-igemm/gen/1x4-wasm.c", |
| 276 | "src/f32-igemm/gen/2x4-wasm.c", |
| 277 | "src/f32-igemm/gen/4x2-wasm.c", |
| 278 | "src/f32-igemm/gen/4x4-wasm.c", |
| 279 | "src/f32-maxpool/9p8x-wasm-c1.c", |
| 280 | "src/f32-pavgpool/mp9p8q-wasm.c", |
| 281 | "src/f32-pavgpool/up9-wasm.c", |
| 282 | "src/f32-prelu/gen/wasm-2x1.c", |
| 283 | "src/f32-prelu/gen/wasm-2x4.c", |
| 284 | "src/f32-vbinary/gen/vadd-wasm-x1.c", |
| 285 | "src/f32-vbinary/gen/vadd-wasm-x2.c", |
| 286 | "src/f32-vbinary/gen/vadd-wasm-x4.c", |
| 287 | "src/f32-vbinary/gen/vaddc-wasm-x1.c", |
| 288 | "src/f32-vbinary/gen/vaddc-wasm-x2.c", |
| 289 | "src/f32-vbinary/gen/vaddc-wasm-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 290 | "src/f32-vbinary/gen/vdiv-wasm-x1.c", |
| 291 | "src/f32-vbinary/gen/vdiv-wasm-x2.c", |
| 292 | "src/f32-vbinary/gen/vdiv-wasm-x4.c", |
| 293 | "src/f32-vbinary/gen/vdivc-wasm-x1.c", |
| 294 | "src/f32-vbinary/gen/vdivc-wasm-x2.c", |
| 295 | "src/f32-vbinary/gen/vdivc-wasm-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 296 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 297 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 298 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
| 299 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 300 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 301 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
| 302 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 303 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 304 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
| 305 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 306 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 307 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 308 | "src/f32-vbinary/gen/vmul-wasm-x1.c", |
| 309 | "src/f32-vbinary/gen/vmul-wasm-x2.c", |
| 310 | "src/f32-vbinary/gen/vmul-wasm-x4.c", |
| 311 | "src/f32-vbinary/gen/vmulc-wasm-x1.c", |
| 312 | "src/f32-vbinary/gen/vmulc-wasm-x2.c", |
| 313 | "src/f32-vbinary/gen/vmulc-wasm-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 314 | "src/f32-vbinary/gen/vrdivc-wasm-x1.c", |
| 315 | "src/f32-vbinary/gen/vrdivc-wasm-x2.c", |
| 316 | "src/f32-vbinary/gen/vrdivc-wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 317 | "src/f32-vbinary/gen/vrsubc-wasm-x1.c", |
| 318 | "src/f32-vbinary/gen/vrsubc-wasm-x2.c", |
| 319 | "src/f32-vbinary/gen/vrsubc-wasm-x4.c", |
| 320 | "src/f32-vbinary/gen/vsub-wasm-x1.c", |
| 321 | "src/f32-vbinary/gen/vsub-wasm-x2.c", |
| 322 | "src/f32-vbinary/gen/vsub-wasm-x4.c", |
| 323 | "src/f32-vbinary/gen/vsubc-wasm-x1.c", |
| 324 | "src/f32-vbinary/gen/vsubc-wasm-x2.c", |
| 325 | "src/f32-vbinary/gen/vsubc-wasm-x4.c", |
| 326 | "src/f32-vmulcaddc/gen/c1-wasm-2x.c", |
| 327 | "src/f32-vmulcaddc/gen/c2-wasm-2x.c", |
| 328 | "src/f32-vmulcaddc/gen/c4-wasm-2x.c", |
| 329 | ] |
| 330 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 331 | PSIMD_FASTMATH_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 332 | "src/f32-argmaxpool/4x-psimd-c4.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 333 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 334 | "src/f32-argmaxpool/9x-psimd-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 335 | "src/f32-avgpool/mp9p8q-psimd.c", |
| 336 | "src/f32-avgpool/up9-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 337 | "src/f32-bilinear/gen/psimd-c4.c", |
| 338 | "src/f32-bilinear/gen/psimd-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 339 | "src/f32-clamp/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 340 | "src/f32-dwconv/gen/up4x25-psimd-acc2.c", |
| 341 | "src/f32-dwconv/gen/up4x25-psimd.c", |
| 342 | "src/f32-dwconv/gen/up4x4-psimd-acc2.c", |
| 343 | "src/f32-dwconv/gen/up4x4-psimd.c", |
| 344 | "src/f32-dwconv/gen/up4x9-psimd-acc2.c", |
| 345 | "src/f32-dwconv/gen/up4x9-psimd.c", |
| 346 | "src/f32-dwconv/gen/up8x25-psimd-acc2.c", |
| 347 | "src/f32-dwconv/gen/up8x25-psimd.c", |
| 348 | "src/f32-dwconv/gen/up8x4-psimd-acc2.c", |
| 349 | "src/f32-dwconv/gen/up8x4-psimd.c", |
| 350 | "src/f32-dwconv/gen/up8x9-psimd-acc2.c", |
| 351 | "src/f32-dwconv/gen/up8x9-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 352 | "src/f32-gavgpool/mp7p7q-psimd.c", |
| 353 | "src/f32-gavgpool/up7-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 354 | "src/f32-gemm/gen/1x8-psimd-loadsplat.c", |
| 355 | "src/f32-gemm/gen/1x8-psimd-splat.c", |
| 356 | "src/f32-gemm/gen/1x8s4-psimd.c", |
| 357 | "src/f32-gemm/gen/4x8-psimd-loadsplat.c", |
| 358 | "src/f32-gemm/gen/4x8-psimd-splat.c", |
| 359 | "src/f32-gemm/gen/4x8s4-psimd.c", |
| 360 | "src/f32-gemm/gen/6x8-psimd-loadsplat.c", |
| 361 | "src/f32-gemm/gen/6x8-psimd-splat.c", |
| 362 | "src/f32-gemm/gen/6x8s4-psimd.c", |
| 363 | "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c", |
| 364 | "src/f32-gemm/gen-inc/1x8-psimd-splat.c", |
| 365 | "src/f32-gemm/gen-inc/1x8s4-psimd.c", |
| 366 | "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c", |
| 367 | "src/f32-gemm/gen-inc/4x8-psimd-splat.c", |
| 368 | "src/f32-gemm/gen-inc/4x8s4-psimd.c", |
| 369 | "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c", |
| 370 | "src/f32-gemm/gen-inc/6x8-psimd-splat.c", |
| 371 | "src/f32-gemm/gen-inc/6x8s4-psimd.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 372 | "src/f32-hswish/gen/psimd-x4.c", |
| 373 | "src/f32-hswish/gen/psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 374 | "src/f32-igemm/gen/1x8-psimd-loadsplat.c", |
| 375 | "src/f32-igemm/gen/1x8-psimd-splat.c", |
| 376 | "src/f32-igemm/gen/1x8s4-psimd.c", |
| 377 | "src/f32-igemm/gen/4x2c4-psimd.c", |
| 378 | "src/f32-igemm/gen/4x8-psimd-loadsplat.c", |
| 379 | "src/f32-igemm/gen/4x8-psimd-splat.c", |
| 380 | "src/f32-igemm/gen/4x8s4-psimd.c", |
| 381 | "src/f32-igemm/gen/6x8-psimd-loadsplat.c", |
| 382 | "src/f32-igemm/gen/6x8-psimd-splat.c", |
| 383 | "src/f32-igemm/gen/6x8s4-psimd.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 384 | "src/f32-maxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 385 | "src/f32-pavgpool/mp9p8q-psimd.c", |
| 386 | "src/f32-pavgpool/up9-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 387 | "src/f32-ppmm/gen/4x8-psimd.c", |
| 388 | "src/f32-prelu/gen/psimd-2x4.c", |
| 389 | "src/f32-prelu/gen/psimd-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 390 | "src/f32-rmax/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 391 | "src/f32-vbinary/gen/vadd-psimd-x4.c", |
| 392 | "src/f32-vbinary/gen/vadd-psimd-x8.c", |
| 393 | "src/f32-vbinary/gen/vaddc-psimd-x4.c", |
| 394 | "src/f32-vbinary/gen/vaddc-psimd-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 395 | "src/f32-vbinary/gen/vdiv-psimd-x4.c", |
| 396 | "src/f32-vbinary/gen/vdiv-psimd-x8.c", |
| 397 | "src/f32-vbinary/gen/vdivc-psimd-x4.c", |
| 398 | "src/f32-vbinary/gen/vdivc-psimd-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 399 | "src/f32-vbinary/gen/vmax-psimd-x4.c", |
| 400 | "src/f32-vbinary/gen/vmax-psimd-x8.c", |
| 401 | "src/f32-vbinary/gen/vmaxc-psimd-x4.c", |
| 402 | "src/f32-vbinary/gen/vmaxc-psimd-x8.c", |
| 403 | "src/f32-vbinary/gen/vmin-psimd-x4.c", |
| 404 | "src/f32-vbinary/gen/vmin-psimd-x8.c", |
| 405 | "src/f32-vbinary/gen/vminc-psimd-x4.c", |
| 406 | "src/f32-vbinary/gen/vminc-psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 407 | "src/f32-vbinary/gen/vmul-psimd-x4.c", |
| 408 | "src/f32-vbinary/gen/vmul-psimd-x8.c", |
| 409 | "src/f32-vbinary/gen/vmulc-psimd-x4.c", |
| 410 | "src/f32-vbinary/gen/vmulc-psimd-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 411 | "src/f32-vbinary/gen/vrdivc-psimd-x4.c", |
| 412 | "src/f32-vbinary/gen/vrdivc-psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 413 | "src/f32-vbinary/gen/vrsubc-psimd-x4.c", |
| 414 | "src/f32-vbinary/gen/vrsubc-psimd-x8.c", |
| 415 | "src/f32-vbinary/gen/vsub-psimd-x4.c", |
| 416 | "src/f32-vbinary/gen/vsub-psimd-x8.c", |
| 417 | "src/f32-vbinary/gen/vsubc-psimd-x4.c", |
| 418 | "src/f32-vbinary/gen/vsubc-psimd-x8.c", |
| 419 | "src/f32-vmulcaddc/gen/c4-psimd-2x.c", |
| 420 | "src/f32-vmulcaddc/gen/c8-psimd-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 421 | "src/x32-packx/x4-psimd.c", |
| 422 | "src/x32-pad/x2-psimd.c", |
| 423 | "src/x32-unpool/psimd.c", |
| 424 | "src/x32-zip/x2-psimd.c", |
| 425 | "src/x32-zip/x3-psimd.c", |
| 426 | "src/x32-zip/x4-psimd.c", |
| 427 | "src/x32-zip/xm-psimd.c", |
| 428 | ] |
| 429 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 430 | PSIMD_ACCMATH_UKERNELS = [ |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 431 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c", |
| 432 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c", |
| 433 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c", |
| 434 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c", |
| 435 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c", |
| 436 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c", |
| 437 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c", |
| 438 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c", |
| 439 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c", |
| 440 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c", |
| 441 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c", |
| 442 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 443 | "src/f32-sigmoid/gen/psimd-p5-div-x4.c", |
| 444 | "src/f32-sigmoid/gen/psimd-p5-div-x8.c", |
| 445 | "src/f32-sigmoid/gen/psimd-p5-div-x12.c", |
| 446 | "src/f32-sigmoid/gen/psimd-p5-div-x16.c", |
| 447 | "src/f32-sigmoid/gen/psimd-p5-div-x20.c", |
| 448 | "src/f32-sigmoid/gen/psimd-p5-div-x24.c", |
| 449 | "src/math/sigmoid-psimd-p5-div.c", |
| 450 | ] |
| 451 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 452 | # ISA-specific micro-kernels |
| 453 | NEON_UKERNELS = [ |
| 454 | "src/f32-avgpool/mp9p8q-neon.c", |
| 455 | "src/f32-avgpool/up9-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 456 | "src/f32-bilinear/gen/neon-c4.c", |
| 457 | "src/f32-bilinear/gen/neon-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 458 | "src/f32-clamp/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 459 | "src/f32-dwconv/gen/up4x9-neon.c", |
| 460 | "src/f32-dwconv/gen/up4x9-neon-acc2.c", |
| 461 | "src/f32-dwconv/gen/up8x9-neon.c", |
| 462 | "src/f32-dwconv/gen/up8x9-neon-acc2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 463 | "src/f32-gavgpool-spchw/neon-x4.c", |
| 464 | "src/f32-gavgpool/mp7p7q-neon.c", |
| 465 | "src/f32-gavgpool/up7-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 466 | "src/f32-gemm/gen/1x8-neon-lane-ld64.c", |
| 467 | "src/f32-gemm/gen/4x2-neon-lane-ld64.c", |
| 468 | "src/f32-gemm/gen/4x8-neon-lane-ld128.c", |
| 469 | "src/f32-gemm/gen/4x8-neon-lane-ld64.c", |
| 470 | "src/f32-gemm/gen/5x8-neon-lane-ld64.c", |
| 471 | "src/f32-gemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 472 | "src/f32-gemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 473 | "src/f32-gemm/gen/1x8-neon-dup-ld64.c", |
| 474 | "src/f32-gemm/gen/4x8-neon-dup-ld128.c", |
| 475 | "src/f32-gemm/gen/4x8-neon-dup-ld64.c", |
| 476 | "src/f32-gemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 477 | "src/f32-gemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 478 | "src/f32-gemm/gen/1x8s4-neon.c", |
| 479 | "src/f32-gemm/gen/4x8s4-neon.c", |
| 480 | "src/f32-gemm/gen/6x8s4-neon.c", |
| 481 | "src/f32-gemm/gen/8x8s4-neon.c", |
| 482 | "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c", |
| 483 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c", |
| 484 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c", |
| 485 | "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c", |
| 486 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 487 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 488 | "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c", |
| 489 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c", |
| 490 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c", |
| 491 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 492 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 493 | "src/f32-gemm/gen-inc/1x8s4-neon.c", |
| 494 | "src/f32-gemm/gen-inc/4x8s4-neon.c", |
| 495 | "src/f32-gemm/gen-inc/6x8s4-neon.c", |
| 496 | "src/f32-gemm/gen-inc/8x8s4-neon.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 497 | "src/f32-hswish/gen/neon-x4.c", |
| 498 | "src/f32-hswish/gen/neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 499 | "src/f32-igemm/gen/1x8-neon-lane-ld64.c", |
| 500 | "src/f32-igemm/gen/4x2-neon-lane-ld64.c", |
| 501 | "src/f32-igemm/gen/4x4-neon-lane-ld64.c", |
| 502 | "src/f32-igemm/gen/4x8-neon-lane-ld128.c", |
| 503 | "src/f32-igemm/gen/4x8-neon-lane-ld64.c", |
| 504 | "src/f32-igemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 505 | "src/f32-igemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 506 | "src/f32-igemm/gen/1x8-neon-dup-ld64.c", |
| 507 | "src/f32-igemm/gen/4x8-neon-dup-ld128.c", |
| 508 | "src/f32-igemm/gen/4x8-neon-dup-ld64.c", |
| 509 | "src/f32-igemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 510 | "src/f32-igemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 511 | "src/f32-igemm/gen/1x8s4-neon.c", |
| 512 | "src/f32-igemm/gen/4x8s4-neon.c", |
| 513 | "src/f32-igemm/gen/6x8s4-neon.c", |
| 514 | "src/f32-igemm/gen/8x8s4-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 515 | "src/f32-pavgpool/mp9p8q-neon.c", |
| 516 | "src/f32-pavgpool/up9-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 517 | "src/f32-ppmm/gen/4x8-neon.c", |
| 518 | "src/f32-ppmm/gen/8x8-neon.c", |
| 519 | "src/f32-prelu/gen/neon-2x4.c", |
| 520 | "src/f32-prelu/gen/neon-2x8.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 521 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 522 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 523 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 524 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 525 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 526 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 527 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 528 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 529 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 530 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| 531 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 532 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 533 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| 534 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| 535 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| 536 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| 537 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 538 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| 539 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| 540 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 541 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| 542 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 543 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 544 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 545 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 546 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 547 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 548 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 549 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 550 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 551 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 552 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| 553 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 554 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 555 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 556 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 557 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 558 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 559 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 560 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 561 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 562 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 563 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 564 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 565 | "src/f32-vbinary/gen/vadd-neon-x4.c", |
| 566 | "src/f32-vbinary/gen/vadd-neon-x8.c", |
| 567 | "src/f32-vbinary/gen/vaddc-neon-x4.c", |
| 568 | "src/f32-vbinary/gen/vaddc-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 569 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 570 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 571 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 572 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 573 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 574 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 575 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 576 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 577 | "src/f32-vbinary/gen/vmul-neon-x4.c", |
| 578 | "src/f32-vbinary/gen/vmul-neon-x8.c", |
| 579 | "src/f32-vbinary/gen/vmulc-neon-x4.c", |
| 580 | "src/f32-vbinary/gen/vmulc-neon-x8.c", |
| 581 | "src/f32-vbinary/gen/vrsubc-neon-x4.c", |
| 582 | "src/f32-vbinary/gen/vrsubc-neon-x8.c", |
| 583 | "src/f32-vbinary/gen/vsub-neon-x4.c", |
| 584 | "src/f32-vbinary/gen/vsub-neon-x8.c", |
| 585 | "src/f32-vbinary/gen/vsubc-neon-x4.c", |
| 586 | "src/f32-vbinary/gen/vsubc-neon-x8.c", |
| 587 | "src/f32-vmulcaddc/gen/c4-neon-2x.c", |
| 588 | "src/f32-vmulcaddc/gen/c8-neon-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 589 | "src/q8-avgpool/mp9p8q-neon.c", |
| 590 | "src/q8-avgpool/up9-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 591 | "src/q8-dwconv/up8x9-neon.c", |
| 592 | "src/q8-gavgpool/mp7p7q-neon.c", |
| 593 | "src/q8-gavgpool/up7-neon.c", |
| 594 | "src/q8-gemm/4x8-neon.c", |
| 595 | "src/q8-gemm/8x8-neon.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 596 | "src/q8-igemm/4x8-neon.c", |
| 597 | "src/q8-igemm/8x8-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 598 | "src/q8-vadd/neon.c", |
| 599 | "src/u8-clamp/neon.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 600 | "src/u8-maxpool/9p8x-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 601 | "src/u8-rmax/neon.c", |
| 602 | "src/x32-packx/x4-neon-st4.c", |
| 603 | "src/x32-pad/x2-neon.c", |
| 604 | "src/x32-zip/x2-neon.c", |
| 605 | "src/x32-zip/x3-neon.c", |
| 606 | "src/x32-zip/x4-neon.c", |
| 607 | "src/x32-zip/xm-neon.c", |
| 608 | "src/x8-zip/x2-neon.c", |
| 609 | "src/x8-zip/x3-neon.c", |
| 610 | "src/x8-zip/x4-neon.c", |
| 611 | "src/x8-zip/xm-neon.c", |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 612 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 613 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 614 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 615 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 616 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 617 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 618 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 619 | ] |
| 620 | |
| 621 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 622 | "src/f32-bilinear/gen/neonfma-c4.c", |
| 623 | "src/f32-bilinear/gen/neonfma-c8.c", |
| 624 | "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c", |
| 625 | "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c", |
| 626 | "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c", |
| 627 | "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 628 | "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 629 | "src/f32-igemm/gen/1x8s4-neonfma.c", |
| 630 | "src/f32-igemm/gen/4x8s4-neonfma.c", |
| 631 | "src/f32-igemm/gen/6x8s4-neonfma.c", |
| 632 | "src/f32-igemm/gen/8x8s4-neonfma.c", |
| 633 | "src/f32-dwconv/gen/up4x9-neonfma.c", |
| 634 | "src/f32-dwconv/gen/up4x9-neonfma-acc2.c", |
| 635 | "src/f32-dwconv/gen/up8x9-neonfma.c", |
| 636 | "src/f32-dwconv/gen/up8x9-neonfma-acc2.c", |
| 637 | "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c", |
| 638 | "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c", |
| 639 | "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c", |
| 640 | "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 641 | "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 642 | "src/f32-gemm/gen/1x8s4-neonfma.c", |
| 643 | "src/f32-gemm/gen/4x8s4-neonfma.c", |
| 644 | "src/f32-gemm/gen/6x8s4-neonfma.c", |
| 645 | "src/f32-gemm/gen/8x8s4-neonfma.c", |
| 646 | "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c", |
| 647 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c", |
| 648 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c", |
| 649 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 650 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 651 | "src/f32-gemm/gen-inc/1x8s4-neonfma.c", |
| 652 | "src/f32-gemm/gen-inc/4x8s4-neonfma.c", |
| 653 | "src/f32-gemm/gen-inc/6x8s4-neonfma.c", |
| 654 | "src/f32-gemm/gen-inc/8x8s4-neonfma.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 655 | "src/f32-hswish/gen/neonfma-x4.c", |
| 656 | "src/f32-hswish/gen/neonfma-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 657 | "src/f32-ppmm/gen/4x8-neonfma.c", |
| 658 | "src/f32-ppmm/gen/8x8-neonfma.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 659 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 660 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 661 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 662 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 663 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 664 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 665 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 666 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 667 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 668 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| 669 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 670 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 671 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| 672 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| 673 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| 674 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| 675 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 676 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| 677 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| 678 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 679 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| 680 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 681 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 682 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 683 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 684 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 685 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 686 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 687 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 688 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 689 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 690 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 691 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 692 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 693 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 694 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 695 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 696 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 697 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 698 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 699 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 700 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| 701 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 702 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 703 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 704 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 705 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 706 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| 707 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 708 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 709 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 710 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 711 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 712 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| 713 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 714 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 715 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 716 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 717 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 718 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| 719 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 720 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 721 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 722 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 723 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 724 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| 725 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 726 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 727 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 728 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 729 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 730 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| 731 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 732 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 733 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 734 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 735 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 736 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 737 | "src/f32-vmulcaddc/gen/c4-neonfma-2x.c", |
| 738 | "src/f32-vmulcaddc/gen/c8-neonfma-2x.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 739 | "src/math/exp-neonfma-lut64-p2.c", |
| 740 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 741 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 742 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 743 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 744 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 745 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 746 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| 747 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 748 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 749 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| 750 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 751 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 752 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| 753 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 754 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 755 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| 756 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 757 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 758 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| 759 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 760 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 761 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 762 | ] |
| 763 | |
| 764 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 765 | "src/f32-vbinary/gen/vdiv-neon-x4.c", |
| 766 | "src/f32-vbinary/gen/vdiv-neon-x8.c", |
| 767 | "src/f32-vbinary/gen/vdivc-neon-x4.c", |
| 768 | "src/f32-vbinary/gen/vdivc-neon-x8.c", |
| 769 | "src/f32-vbinary/gen/vrdivc-neon-x4.c", |
| 770 | "src/f32-vbinary/gen/vrdivc-neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 771 | "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c", |
| 772 | "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c", |
| 773 | "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c", |
| 774 | "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c", |
| 775 | "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c", |
| 776 | "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 777 | "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 778 | "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c", |
| 779 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c", |
| 780 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c", |
| 781 | "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c", |
| 782 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 783 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 784 | "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c", |
| 785 | "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c", |
| 786 | "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c", |
| 787 | "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c", |
| 788 | "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c", |
| 789 | "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 790 | "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 791 | "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c", |
| 792 | "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c", |
| 793 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c", |
| 794 | "src/f32-dwconv-spchw/3x3p1-neonfma.c", |
| 795 | "src/f32-dwconv-spchw/5x5p2-neonfma.c", |
| 796 | "src/f32-dwconv-spchw/3x3s2p1-neonfma.c", |
| 797 | "src/f32-dwconv-spchw/5x5s2p2-neonfma.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 798 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 799 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 800 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 801 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 802 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 803 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| 804 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 805 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 806 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 807 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 808 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 809 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 810 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 811 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 812 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 813 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 814 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 815 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 816 | "src/f32-spmm/gen/12x1-neonfma.c", |
| 817 | "src/f32-spmm/gen/12x2-neonfma.c", |
| 818 | "src/f32-spmm/gen/12x4-neonfma.c", |
| 819 | "src/f32-spmm/gen/16x1-neonfma-pipelined.c", |
| 820 | "src/f32-spmm/gen/16x1-neonfma-unroll2.c", |
| 821 | "src/f32-spmm/gen/16x1-neonfma.c", |
| 822 | "src/f32-spmm/gen/16x2-neonfma.c", |
| 823 | "src/f32-spmm/gen/16x4-neonfma.c", |
| 824 | "src/f32-spmm/gen/4x1-neonfma-pipelined.c", |
| 825 | "src/f32-spmm/gen/4x1-neonfma-unroll2.c", |
| 826 | "src/f32-spmm/gen/4x1-neonfma.c", |
| 827 | "src/f32-spmm/gen/4x2-neonfma.c", |
| 828 | "src/f32-spmm/gen/4x4-neonfma.c", |
| 829 | "src/f32-spmm/gen/8x1-neonfma-pipelined.c", |
| 830 | "src/f32-spmm/gen/8x1-neonfma-unroll2.c", |
| 831 | "src/f32-spmm/gen/8x1-neonfma.c", |
| 832 | "src/f32-spmm/gen/8x2-neonfma.c", |
| 833 | "src/f32-spmm/gen/8x4-neonfma.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 834 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| 835 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| 836 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| 837 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| 838 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| 839 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 840 | ] |
| 841 | |
| 842 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 843 | "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c", |
| 844 | "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c", |
| 845 | "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 846 | ] |
| 847 | |
| 848 | SSE_UKERNELS = [ |
| 849 | "src/f32-avgpool/mp9p8q-sse.c", |
| 850 | "src/f32-avgpool/up9-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 851 | "src/f32-bilinear/gen/sse-c4.c", |
| 852 | "src/f32-bilinear/gen/sse-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 853 | "src/f32-clamp/sse.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 854 | "src/f32-dwconv-spchw/3x3p1-sse.c", |
| 855 | "src/f32-dwconv-spchw/3x3s2p1-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 856 | "src/f32-dwconv/gen/up4x25-sse-acc2.c", |
| 857 | "src/f32-dwconv/gen/up4x25-sse.c", |
| 858 | "src/f32-dwconv/gen/up4x4-sse-acc2.c", |
| 859 | "src/f32-dwconv/gen/up4x4-sse.c", |
| 860 | "src/f32-dwconv/gen/up4x9-sse-acc2.c", |
| 861 | "src/f32-dwconv/gen/up4x9-sse.c", |
| 862 | "src/f32-dwconv/gen/up8x25-sse-acc2.c", |
| 863 | "src/f32-dwconv/gen/up8x25-sse.c", |
| 864 | "src/f32-dwconv/gen/up8x4-sse-acc2.c", |
| 865 | "src/f32-dwconv/gen/up8x4-sse.c", |
| 866 | "src/f32-dwconv/gen/up8x9-sse-acc2.c", |
| 867 | "src/f32-dwconv/gen/up8x9-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 868 | "src/f32-gavgpool-spchw/sse-x4.c", |
| 869 | "src/f32-gavgpool/mp7p7q-sse.c", |
| 870 | "src/f32-gavgpool/up7-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 871 | "src/f32-gemm/gen/1x8-sse-dup.c", |
| 872 | "src/f32-gemm/gen/1x8-sse-load1.c", |
| 873 | "src/f32-gemm/gen/1x8s4-sse.c", |
| 874 | "src/f32-gemm/gen/4x8-sse-dup.c", |
| 875 | "src/f32-gemm/gen/4x8-sse-load1.c", |
| 876 | "src/f32-gemm/gen/4x8s4-sse.c", |
| 877 | "src/f32-gemm/gen-inc/1x8-sse-dup.c", |
| 878 | "src/f32-gemm/gen-inc/1x8-sse-load1.c", |
| 879 | "src/f32-gemm/gen-inc/1x8s4-sse.c", |
| 880 | "src/f32-gemm/gen-inc/4x8-sse-dup.c", |
| 881 | "src/f32-gemm/gen-inc/4x8-sse-load1.c", |
| 882 | "src/f32-gemm/gen-inc/4x8s4-sse.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 883 | "src/f32-hswish/gen/sse-x4.c", |
| 884 | "src/f32-hswish/gen/sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 885 | "src/f32-igemm/gen/1x8-sse-dup.c", |
| 886 | "src/f32-igemm/gen/1x8-sse-load1.c", |
| 887 | "src/f32-igemm/gen/1x8s4-sse.c", |
| 888 | "src/f32-igemm/gen/4x2c4-sse.c", |
| 889 | "src/f32-igemm/gen/4x8-sse-dup.c", |
| 890 | "src/f32-igemm/gen/4x8-sse-load1.c", |
| 891 | "src/f32-igemm/gen/4x8s4-sse.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 892 | "src/f32-maxpool/9p8x-sse-c4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 893 | "src/f32-pavgpool/mp9p8q-sse.c", |
| 894 | "src/f32-pavgpool/up9-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 895 | "src/f32-ppmm/gen/4x8-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 896 | "src/f32-rmax/sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 897 | "src/f32-spmm/gen/4x1-sse.c", |
| 898 | "src/f32-spmm/gen/8x1-sse.c", |
| 899 | "src/f32-vbinary/gen/vadd-sse-x4.c", |
| 900 | "src/f32-vbinary/gen/vadd-sse-x8.c", |
| 901 | "src/f32-vbinary/gen/vaddc-sse-x4.c", |
| 902 | "src/f32-vbinary/gen/vaddc-sse-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 903 | "src/f32-vbinary/gen/vdiv-sse-x4.c", |
| 904 | "src/f32-vbinary/gen/vdiv-sse-x8.c", |
| 905 | "src/f32-vbinary/gen/vdivc-sse-x4.c", |
| 906 | "src/f32-vbinary/gen/vdivc-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 907 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 908 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 909 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 910 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 911 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 912 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 913 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 914 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 915 | "src/f32-vbinary/gen/vmul-sse-x4.c", |
| 916 | "src/f32-vbinary/gen/vmul-sse-x8.c", |
| 917 | "src/f32-vbinary/gen/vmulc-sse-x4.c", |
| 918 | "src/f32-vbinary/gen/vmulc-sse-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 919 | "src/f32-vbinary/gen/vrdivc-sse-x4.c", |
| 920 | "src/f32-vbinary/gen/vrdivc-sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 921 | "src/f32-vbinary/gen/vrsubc-sse-x4.c", |
| 922 | "src/f32-vbinary/gen/vrsubc-sse-x8.c", |
| 923 | "src/f32-vbinary/gen/vsub-sse-x4.c", |
| 924 | "src/f32-vbinary/gen/vsub-sse-x8.c", |
| 925 | "src/f32-vbinary/gen/vsubc-sse-x4.c", |
| 926 | "src/f32-vbinary/gen/vsubc-sse-x8.c", |
| 927 | "src/f32-vmulcaddc/gen/c4-sse-2x.c", |
| 928 | "src/f32-vmulcaddc/gen/c8-sse-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 929 | "src/x32-packx/x4-sse.c", |
| 930 | ] |
| 931 | |
| 932 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 933 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 934 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 935 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 936 | "src/f32-prelu/gen/sse2-2x4.c", |
| 937 | "src/f32-prelu/gen/sse2-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 938 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| 939 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| 940 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| 941 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| 942 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 943 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| 944 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| 945 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 946 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| 947 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 948 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 949 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 950 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 951 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 952 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 953 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 954 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 955 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 956 | "src/q8-avgpool/mp9p8q-sse2.c", |
| 957 | "src/q8-avgpool/up9-sse2.c", |
| 958 | "src/q8-igemm/4x4c2-sse2.c", |
| 959 | "src/q8-dwconv/up8x9-sse2.c", |
| 960 | "src/q8-gavgpool/mp7p7q-sse2.c", |
| 961 | "src/q8-gavgpool/up7-sse2.c", |
| 962 | "src/q8-gemm/2x4c8-sse2.c", |
| 963 | "src/q8-gemm/4x4c2-sse2.c", |
| 964 | "src/q8-vadd/sse2.c", |
| 965 | "src/u8-clamp/sse2.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 966 | "src/u8-maxpool/9p8x-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 967 | "src/u8-rmax/sse2.c", |
| 968 | "src/x32-pad/x2-sse2.c", |
| 969 | "src/x32-zip/x2-sse2.c", |
| 970 | "src/x32-zip/x3-sse2.c", |
| 971 | "src/x32-zip/x4-sse2.c", |
| 972 | "src/x32-zip/xm-sse2.c", |
| 973 | "src/x8-zip/x2-sse2.c", |
| 974 | "src/x8-zip/x3-sse2.c", |
| 975 | "src/x8-zip/x4-sse2.c", |
| 976 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 977 | "src/math/exp-sse2-p5.c", |
| 978 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 979 | "src/math/sigmoid-sse2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 980 | ] |
| 981 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 982 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 983 | "src/f32-prelu/gen/sse41-2x4.c", |
| 984 | "src/f32-prelu/gen/sse41-2x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 985 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 986 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 987 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 988 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 989 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 990 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 991 | ] |
| 992 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 993 | AVX_UKERNELS = [ |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 994 | "src/f32-clamp/avx.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 995 | "src/f32-dwconv/gen/up16x4-avx-acc2.c", |
| 996 | "src/f32-dwconv/gen/up16x4-avx.c", |
| 997 | "src/f32-dwconv/gen/up8x4-avx-acc2.c", |
| 998 | "src/f32-dwconv/gen/up8x4-avx.c", |
| 999 | "src/f32-dwconv/gen/up16x9-avx-acc2.c", |
| 1000 | "src/f32-dwconv/gen/up16x9-avx.c", |
| 1001 | "src/f32-dwconv/gen/up8x9-avx-acc2.c", |
| 1002 | "src/f32-dwconv/gen/up8x9-avx.c", |
| 1003 | "src/f32-dwconv/gen/up16x25-avx-acc2.c", |
| 1004 | "src/f32-dwconv/gen/up16x25-avx.c", |
| 1005 | "src/f32-dwconv/gen/up8x25-avx-acc2.c", |
| 1006 | "src/f32-dwconv/gen/up8x25-avx.c", |
| 1007 | "src/f32-gemm/gen/1x8-avx-broadcast.c", |
| 1008 | "src/f32-gemm/gen/4x8-avx-broadcast.c", |
| 1009 | "src/f32-gemm/gen/5x8-avx-broadcast.c", |
| 1010 | "src/f32-gemm/gen/6x8-avx-broadcast.c", |
| 1011 | "src/f32-gemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1012 | "src/f32-gemm/gen/1x16-avx-broadcast.c", |
| 1013 | "src/f32-gemm/gen/3x16-avx-broadcast.c", |
| 1014 | "src/f32-gemm/gen/4x16-avx-broadcast.c", |
| 1015 | "src/f32-gemm/gen/5x16-avx-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1016 | "src/f32-gemm/gen-inc/1x8-avx-broadcast.c", |
| 1017 | "src/f32-gemm/gen-inc/4x8-avx-broadcast.c", |
| 1018 | "src/f32-gemm/gen-inc/5x8-avx-broadcast.c", |
| 1019 | "src/f32-gemm/gen-inc/6x8-avx-broadcast.c", |
| 1020 | "src/f32-gemm/gen-inc/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1021 | "src/f32-gemm/gen-inc/1x16-avx-broadcast.c", |
| 1022 | "src/f32-gemm/gen-inc/3x16-avx-broadcast.c", |
| 1023 | "src/f32-gemm/gen-inc/4x16-avx-broadcast.c", |
| 1024 | "src/f32-gemm/gen-inc/5x16-avx-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1025 | "src/f32-hswish/gen/avx-x8.c", |
| 1026 | "src/f32-hswish/gen/avx-x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1027 | "src/f32-igemm/gen/1x8-avx-broadcast.c", |
| 1028 | "src/f32-igemm/gen/4x8-avx-broadcast.c", |
| 1029 | "src/f32-igemm/gen/5x8-avx-broadcast.c", |
| 1030 | "src/f32-igemm/gen/6x8-avx-broadcast.c", |
| 1031 | "src/f32-igemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1032 | "src/f32-igemm/gen/1x16-avx-broadcast.c", |
| 1033 | "src/f32-igemm/gen/3x16-avx-broadcast.c", |
| 1034 | "src/f32-igemm/gen/4x16-avx-broadcast.c", |
| 1035 | "src/f32-igemm/gen/5x16-avx-broadcast.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1036 | "src/f32-rmax/avx.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1037 | "src/f32-vbinary/gen/vadd-avx-x8.c", |
| 1038 | "src/f32-vbinary/gen/vadd-avx-x16.c", |
| 1039 | "src/f32-vbinary/gen/vaddc-avx-x8.c", |
| 1040 | "src/f32-vbinary/gen/vaddc-avx-x16.c", |
| 1041 | "src/f32-vbinary/gen/vdiv-avx-x8.c", |
| 1042 | "src/f32-vbinary/gen/vdiv-avx-x16.c", |
| 1043 | "src/f32-vbinary/gen/vdivc-avx-x8.c", |
| 1044 | "src/f32-vbinary/gen/vdivc-avx-x16.c", |
| 1045 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 1046 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 1047 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 1048 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 1049 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 1050 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 1051 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 1052 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
| 1053 | "src/f32-vbinary/gen/vmul-avx-x8.c", |
| 1054 | "src/f32-vbinary/gen/vmul-avx-x16.c", |
| 1055 | "src/f32-vbinary/gen/vmulc-avx-x8.c", |
| 1056 | "src/f32-vbinary/gen/vmulc-avx-x16.c", |
| 1057 | "src/f32-vbinary/gen/vrdivc-avx-x8.c", |
| 1058 | "src/f32-vbinary/gen/vrdivc-avx-x16.c", |
| 1059 | "src/f32-vbinary/gen/vrsubc-avx-x8.c", |
| 1060 | "src/f32-vbinary/gen/vrsubc-avx-x16.c", |
| 1061 | "src/f32-vbinary/gen/vsub-avx-x8.c", |
| 1062 | "src/f32-vbinary/gen/vsub-avx-x16.c", |
| 1063 | "src/f32-vbinary/gen/vsubc-avx-x8.c", |
| 1064 | "src/f32-vbinary/gen/vsubc-avx-x16.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1065 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1066 | ] |
| 1067 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1068 | FMA3_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1069 | "src/f32-dwconv/gen/up16x4-fma3-acc2.c", |
| 1070 | "src/f32-dwconv/gen/up16x4-fma3.c", |
| 1071 | "src/f32-dwconv/gen/up8x4-fma3-acc2.c", |
| 1072 | "src/f32-dwconv/gen/up8x4-fma3.c", |
| 1073 | "src/f32-dwconv/gen/up16x9-fma3-acc2.c", |
| 1074 | "src/f32-dwconv/gen/up16x9-fma3.c", |
| 1075 | "src/f32-dwconv/gen/up8x9-fma3-acc2.c", |
| 1076 | "src/f32-dwconv/gen/up8x9-fma3.c", |
| 1077 | "src/f32-dwconv/gen/up16x25-fma3-acc2.c", |
| 1078 | "src/f32-dwconv/gen/up16x25-fma3.c", |
| 1079 | "src/f32-dwconv/gen/up8x25-fma3-acc2.c", |
| 1080 | "src/f32-dwconv/gen/up8x25-fma3.c", |
| 1081 | "src/f32-gemm/gen/1x8-fma3-broadcast.c", |
| 1082 | "src/f32-gemm/gen/4x8-fma3-broadcast.c", |
| 1083 | "src/f32-gemm/gen/5x8-fma3-broadcast.c", |
| 1084 | "src/f32-gemm/gen/6x8-fma3-broadcast.c", |
| 1085 | "src/f32-gemm/gen/7x8-fma3-broadcast.c", |
| 1086 | "src/f32-gemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1087 | "src/f32-gemm/gen/1x16-fma3-broadcast.c", |
| 1088 | "src/f32-gemm/gen/3x16-fma3-broadcast.c", |
| 1089 | "src/f32-gemm/gen/4x16-fma3-broadcast.c", |
| 1090 | "src/f32-gemm/gen/5x16-fma3-broadcast.c", |
Ashkan Aliabadi | d94b856 | 2019-12-10 11:33:51 -0800 | [diff] [blame] | 1091 | "src/f32-gemm/gen/1x16s4-fma3-broadcast.c", |
| 1092 | "src/f32-gemm/gen/3x16s4-fma3-broadcast.c", |
| 1093 | "src/f32-gemm/gen/4x16s4-fma3-broadcast.c", |
| 1094 | "src/f32-gemm/gen/5x16s4-fma3-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1095 | "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c", |
| 1096 | "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c", |
| 1097 | "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c", |
| 1098 | "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c", |
| 1099 | "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c", |
| 1100 | "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1101 | "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c", |
| 1102 | "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c", |
| 1103 | "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c", |
| 1104 | "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c", |
Ashkan Aliabadi | d94b856 | 2019-12-10 11:33:51 -0800 | [diff] [blame] | 1105 | "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c", |
| 1106 | "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c", |
| 1107 | "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c", |
| 1108 | "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1109 | "src/f32-hswish/gen/fma3-x8.c", |
| 1110 | "src/f32-hswish/gen/fma3-x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1111 | "src/f32-igemm/gen/1x8-fma3-broadcast.c", |
| 1112 | "src/f32-igemm/gen/4x8-fma3-broadcast.c", |
| 1113 | "src/f32-igemm/gen/5x8-fma3-broadcast.c", |
| 1114 | "src/f32-igemm/gen/6x8-fma3-broadcast.c", |
| 1115 | "src/f32-igemm/gen/7x8-fma3-broadcast.c", |
| 1116 | "src/f32-igemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1117 | "src/f32-igemm/gen/1x16-fma3-broadcast.c", |
| 1118 | "src/f32-igemm/gen/3x16-fma3-broadcast.c", |
| 1119 | "src/f32-igemm/gen/4x16-fma3-broadcast.c", |
| 1120 | "src/f32-igemm/gen/5x16-fma3-broadcast.c", |
Ashkan Aliabadi | d94b856 | 2019-12-10 11:33:51 -0800 | [diff] [blame] | 1121 | "src/f32-igemm/gen/1x16s4-fma3-broadcast.c", |
| 1122 | "src/f32-igemm/gen/3x16s4-fma3-broadcast.c", |
| 1123 | "src/f32-igemm/gen/4x16s4-fma3-broadcast.c", |
| 1124 | "src/f32-igemm/gen/5x16s4-fma3-broadcast.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1125 | ] |
| 1126 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1127 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1128 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 1129 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1130 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1131 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 1132 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1133 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 1134 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1135 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1136 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 1137 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1138 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1139 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 1140 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 1141 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 1142 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 1143 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 1144 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 1145 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 1146 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 1147 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 1148 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 1149 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 1150 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 1151 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 1152 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 1153 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1154 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1155 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 1156 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1157 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 1158 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1159 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1160 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 1161 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1162 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1163 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1164 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 1165 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 1166 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 1167 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 1168 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 1169 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 1170 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 1171 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 1172 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 1173 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 1174 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 1175 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 1176 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 1177 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 1178 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 1179 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 1180 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 1181 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 1182 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 1183 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 1184 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 1185 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 1186 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 1187 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 1188 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 1189 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 1190 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 1191 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 1192 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 1193 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1194 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 1195 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 1196 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 1197 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 1198 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 1199 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 1200 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 1201 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 1202 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 1203 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 1204 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 1205 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 1206 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 1207 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 1208 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 1209 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 1210 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 1211 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 1212 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 1213 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 1214 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 1215 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 1216 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 1217 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1218 | "src/math/exp-avx2-p5.c", |
| 1219 | "src/math/exp-avx2-perm-p3.c", |
| 1220 | "src/math/exp-avx2-perm-p4.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 1221 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1222 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 1223 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 1224 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 1225 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| 1226 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 1227 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 1228 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1229 | ] |
| 1230 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1231 | AVX512F_UKERNELS = [ |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1232 | "src/f32-clamp/avx512f.c", |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 1233 | "src/f32-dwconv/gen/up32x4-avx512f-acc2.c", |
| 1234 | "src/f32-dwconv/gen/up32x4-avx512f.c", |
| 1235 | "src/f32-dwconv/gen/up16x4-avx512f-acc2.c", |
| 1236 | "src/f32-dwconv/gen/up16x4-avx512f.c", |
| 1237 | "src/f32-dwconv/gen/up32x9-avx512f-acc2.c", |
| 1238 | "src/f32-dwconv/gen/up32x9-avx512f.c", |
| 1239 | "src/f32-dwconv/gen/up16x9-avx512f-acc2.c", |
| 1240 | "src/f32-dwconv/gen/up16x9-avx512f.c", |
| 1241 | "src/f32-dwconv/gen/up32x25-avx512f-acc2.c", |
| 1242 | "src/f32-dwconv/gen/up32x25-avx512f.c", |
| 1243 | "src/f32-dwconv/gen/up16x25-avx512f-acc2.c", |
| 1244 | "src/f32-dwconv/gen/up16x25-avx512f.c", |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1245 | "src/f32-gemm/gen/1x16-avx512f-broadcast.c", |
| 1246 | "src/f32-gemm/gen/4x16-avx512f-broadcast.c", |
| 1247 | "src/f32-gemm/gen/5x16-avx512f-broadcast.c", |
| 1248 | "src/f32-gemm/gen/6x16-avx512f-broadcast.c", |
| 1249 | "src/f32-gemm/gen/7x16-avx512f-broadcast.c", |
| 1250 | "src/f32-gemm/gen/8x16-avx512f-broadcast.c", |
| 1251 | "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c", |
| 1252 | "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c", |
| 1253 | "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c", |
| 1254 | "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c", |
| 1255 | "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c", |
| 1256 | "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1257 | "src/f32-hswish/gen/avx512f-x16.c", |
| 1258 | "src/f32-hswish/gen/avx512f-x32.c", |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1259 | "src/f32-igemm/gen/1x16-avx512f-broadcast.c", |
| 1260 | "src/f32-igemm/gen/4x16-avx512f-broadcast.c", |
| 1261 | "src/f32-igemm/gen/5x16-avx512f-broadcast.c", |
| 1262 | "src/f32-igemm/gen/6x16-avx512f-broadcast.c", |
| 1263 | "src/f32-igemm/gen/7x16-avx512f-broadcast.c", |
| 1264 | "src/f32-igemm/gen/8x16-avx512f-broadcast.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1265 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1266 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1267 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1268 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1269 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1270 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1271 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1272 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1273 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1274 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1275 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1276 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1277 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 1278 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1279 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1280 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 1281 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1282 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 1283 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1284 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1285 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 1286 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1287 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1288 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1289 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1290 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1291 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1292 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1293 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1294 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1295 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1296 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1297 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1298 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1299 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1300 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1301 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1302 | "src/f32-vbinary/gen/vadd-avx512f-x16.c", |
| 1303 | "src/f32-vbinary/gen/vadd-avx512f-x32.c", |
| 1304 | "src/f32-vbinary/gen/vaddc-avx512f-x16.c", |
| 1305 | "src/f32-vbinary/gen/vaddc-avx512f-x32.c", |
| 1306 | "src/f32-vbinary/gen/vdiv-avx512f-x16.c", |
| 1307 | "src/f32-vbinary/gen/vdiv-avx512f-x32.c", |
| 1308 | "src/f32-vbinary/gen/vdivc-avx512f-x16.c", |
| 1309 | "src/f32-vbinary/gen/vdivc-avx512f-x32.c", |
| 1310 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 1311 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 1312 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 1313 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 1314 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 1315 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 1316 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 1317 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
| 1318 | "src/f32-vbinary/gen/vmul-avx512f-x16.c", |
| 1319 | "src/f32-vbinary/gen/vmul-avx512f-x32.c", |
| 1320 | "src/f32-vbinary/gen/vmulc-avx512f-x16.c", |
| 1321 | "src/f32-vbinary/gen/vmulc-avx512f-x32.c", |
| 1322 | "src/f32-vbinary/gen/vrdivc-avx512f-x16.c", |
| 1323 | "src/f32-vbinary/gen/vrdivc-avx512f-x32.c", |
| 1324 | "src/f32-vbinary/gen/vrsubc-avx512f-x16.c", |
| 1325 | "src/f32-vbinary/gen/vrsubc-avx512f-x32.c", |
| 1326 | "src/f32-vbinary/gen/vsub-avx512f-x16.c", |
| 1327 | "src/f32-vbinary/gen/vsub-avx512f-x32.c", |
| 1328 | "src/f32-vbinary/gen/vsubc-avx512f-x16.c", |
| 1329 | "src/f32-vbinary/gen/vsubc-avx512f-x32.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1330 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1331 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 1332 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 1333 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 1334 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 1335 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 1336 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 1337 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 1338 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1339 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1340 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1341 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 1342 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1343 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 1344 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 1345 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 1346 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 1347 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 1348 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 1349 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 1350 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 1351 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 1352 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 1353 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 1354 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1355 | "src/math/exp-avx512f-p5-scalef.c", |
| 1356 | "src/math/exp-avx512f-p5.c", |
| 1357 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | feb4923 | 2019-10-28 11:03:31 -0700 | [diff] [blame] | 1358 | "src/math/exp-avx512f-perm2-p2.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1359 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1360 | ] |
| 1361 | |
| 1362 | AARCH32_ASM_UKERNELS = [ |
| 1363 | "src/q8-dwconv/up8x9-aarch32-neon.S", |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 1364 | "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S", |
Frank Barchard | abf8154 | 2019-12-13 16:18:30 -0800 | [diff] [blame] | 1365 | "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S", |
| 1366 | "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S", |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 1367 | "src/f32-gemm/4x8-aarch32-neon-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1368 | ] |
| 1369 | |
| 1370 | AARCH64_ASM_UKERNELS = [ |
| 1371 | "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S", |
| 1372 | "src/f32-dwconv/up4x9-aarch64-neonfma.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1373 | "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S", |
| 1374 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S", |
| 1375 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1376 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
| 1377 | "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S", |
| 1378 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S", |
| 1379 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1380 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1381 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S", |
| 1382 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1383 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1384 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
| 1385 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1386 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1387 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1388 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
| 1389 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S", |
| 1390 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S", |
| 1391 | "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S", |
| 1392 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S", |
| 1393 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S", |
| 1394 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S", |
| 1395 | "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S", |
| 1396 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S", |
| 1397 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S", |
| 1398 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S", |
| 1399 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S", |
| 1400 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1401 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1402 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S", |
| 1403 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1404 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1405 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1406 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S", |
| 1407 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S", |
| 1408 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1409 | "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 1410 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1411 | "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1412 | "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1413 | "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 1414 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1415 | "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1416 | "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1417 | "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S", |
| 1418 | "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 1419 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1420 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1421 | "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
| 1422 | "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1423 | ] |
| 1424 | |
| 1425 | INTERNAL_MICROKERNEL_HDRS = [ |
| 1426 | "src/xnnpack/argmaxpool.h", |
| 1427 | "src/xnnpack/avgpool.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 1428 | "src/xnnpack/bilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1429 | "src/xnnpack/clamp.h", |
| 1430 | "src/xnnpack/common.h", |
| 1431 | "src/xnnpack/conv.h", |
| 1432 | "src/xnnpack/dwconv.h", |
| 1433 | "src/xnnpack/gavgpool.h", |
| 1434 | "src/xnnpack/gemm.h", |
| 1435 | "src/xnnpack/hswish.h", |
| 1436 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 1437 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1438 | "src/xnnpack/lut.h", |
| 1439 | "src/xnnpack/math.h", |
| 1440 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1441 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1442 | "src/xnnpack/packx.h", |
| 1443 | "src/xnnpack/pad.h", |
| 1444 | "src/xnnpack/params.h", |
| 1445 | "src/xnnpack/pavgpool.h", |
| 1446 | "src/xnnpack/ppmm.h", |
| 1447 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1448 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1449 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1450 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1451 | "src/xnnpack/rmax.h", |
| 1452 | "src/xnnpack/scalar-utils.h", |
| 1453 | "src/xnnpack/spmm.h", |
| 1454 | "src/xnnpack/unpool.h", |
| 1455 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1456 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1457 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1458 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1459 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1460 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1461 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1462 | "src/xnnpack/zip.h", |
| 1463 | ] |
| 1464 | |
| 1465 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1466 | "include/xnnpack.h", |
| 1467 | "src/xnnpack/allocator.h", |
| 1468 | "src/xnnpack/compute.h", |
| 1469 | "src/xnnpack/im2col.h", |
| 1470 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1471 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1472 | "src/xnnpack/operator.h", |
| 1473 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1474 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1475 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1476 | "src/xnnpack/requantization.h", |
| 1477 | ] |
| 1478 | |
| 1479 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1480 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1481 | ] |
| 1482 | |
| 1483 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1484 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1485 | "include/xnnpack.h", |
| 1486 | ] |
| 1487 | |
| 1488 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1489 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1490 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1491 | "src/xnnpack/requantization.h", |
| 1492 | "include/xnnpack.h", |
| 1493 | ] |
| 1494 | |
| 1495 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 1496 | "src/xnnpack/params.h", |
| 1497 | "src/xnnpack/common.h", |
| 1498 | ] |
| 1499 | |
| 1500 | WEIGHTS_PACK_HDRS = [ |
| 1501 | "src/xnnpack/pack.h", |
| 1502 | "src/xnnpack/operator.h", |
| 1503 | "src/xnnpack/compute.h", |
| 1504 | ] |
| 1505 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1506 | LOGGING_COPTS = select({ |
| 1507 | # No logging in optimized mode |
| 1508 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 1509 | # Full logging in debug mode |
| 1510 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 1511 | # Error-only logging in default (fastbuild) mode |
| 1512 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 1513 | }) |
| 1514 | |
| 1515 | LOGGING_HDRS = [ |
| 1516 | "src/xnnpack/log.h", |
| 1517 | ] |
| 1518 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1519 | xnnpack_cc_library( |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1520 | name = "tables", |
| 1521 | srcs = TABLE_SRCS, |
| 1522 | hdrs = INTERNAL_HDRS, |
| 1523 | copts = xnnpack_std_copts(), |
| 1524 | ) |
| 1525 | |
| 1526 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1527 | name = "scalar_ukernels", |
| 1528 | srcs = SCALAR_UKERNELS, |
| 1529 | hdrs = INTERNAL_HDRS, |
| 1530 | aarch32_copts = ["-marm"], |
| 1531 | copts = xnnpack_std_copts(), |
| 1532 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1533 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1534 | "@FP16", |
| 1535 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1536 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1537 | ], |
| 1538 | ) |
| 1539 | |
| 1540 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1541 | name = "wasm_ukernels", |
| 1542 | hdrs = INTERNAL_HDRS, |
| 1543 | copts = xnnpack_std_copts(), |
| 1544 | wasm_srcs = WASM_UKERNELS, |
| 1545 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1546 | ":tables", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1547 | "@FP16", |
| 1548 | "@FXdiv", |
| 1549 | "@pthreadpool", |
| 1550 | ], |
| 1551 | ) |
| 1552 | |
| 1553 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1554 | name = "psimd_fastmath_ukernels", |
| 1555 | srcs = PSIMD_FASTMATH_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1556 | hdrs = INTERNAL_HDRS, |
| 1557 | aarch32_copts = [ |
| 1558 | "-marm", |
| 1559 | "-mfpu=neon", |
| 1560 | ], |
| 1561 | copts = xnnpack_std_copts(), |
| 1562 | optimized_copts = [ |
| 1563 | "-O3", |
| 1564 | "-ffast-math", |
| 1565 | ], |
| 1566 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1567 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1568 | "@FP16", |
| 1569 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1570 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1571 | ], |
| 1572 | ) |
| 1573 | |
| 1574 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1575 | name = "psimd_accmath_ukernels", |
| 1576 | srcs = PSIMD_ACCMATH_UKERNELS, |
| 1577 | hdrs = INTERNAL_HDRS, |
| 1578 | aarch32_copts = [ |
| 1579 | "-marm", |
| 1580 | "-mfpu=neon", |
| 1581 | ], |
| 1582 | copts = xnnpack_std_copts(), |
| 1583 | optimized_copts = [ |
| 1584 | "-O3", |
| 1585 | ], |
| 1586 | deps = [ |
| 1587 | ":tables", |
| 1588 | "@FP16", |
| 1589 | "@psimd", |
| 1590 | "@pthreadpool", |
| 1591 | ], |
| 1592 | ) |
| 1593 | |
| 1594 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1595 | name = "neon_ukernels", |
| 1596 | hdrs = INTERNAL_HDRS, |
| 1597 | aarch32_copts = [ |
| 1598 | "-marm", |
| 1599 | "-mfpu=neon", |
| 1600 | ], |
| 1601 | aarch32_srcs = NEON_UKERNELS, |
| 1602 | aarch64_srcs = NEON_UKERNELS, |
| 1603 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1604 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1605 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1606 | "@FP16", |
| 1607 | "@pthreadpool", |
| 1608 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1609 | ) |
| 1610 | |
| 1611 | xnnpack_cc_library( |
| 1612 | name = "neonfma_ukernels", |
| 1613 | hdrs = INTERNAL_HDRS, |
| 1614 | aarch32_copts = [ |
| 1615 | "-marm", |
| 1616 | "-mfpu=neon-vfpv4", |
| 1617 | ], |
| 1618 | aarch32_srcs = NEONFMA_UKERNELS, |
| 1619 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 1620 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1621 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1622 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1623 | "@FP16", |
| 1624 | "@pthreadpool", |
| 1625 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1626 | ) |
| 1627 | |
| 1628 | xnnpack_cc_library( |
| 1629 | name = "neonfp16arith_ukernels", |
| 1630 | hdrs = INTERNAL_HDRS, |
| 1631 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 1632 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 1633 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1634 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1635 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1636 | "@FP16", |
| 1637 | "@pthreadpool", |
| 1638 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1639 | ) |
| 1640 | |
| 1641 | xnnpack_cc_library( |
| 1642 | name = "sse2_ukernels", |
| 1643 | hdrs = INTERNAL_HDRS, |
| 1644 | copts = xnnpack_std_copts(), |
| 1645 | x86_copts = ["-msse2"], |
| 1646 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1647 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1648 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1649 | "@FP16", |
| 1650 | "@pthreadpool", |
| 1651 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1652 | ) |
| 1653 | |
| 1654 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1655 | name = "sse41_ukernels", |
| 1656 | hdrs = INTERNAL_HDRS, |
| 1657 | copts = xnnpack_std_copts(), |
| 1658 | x86_copts = ["-msse4.1"], |
| 1659 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1660 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1661 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1662 | "@FP16", |
| 1663 | "@pthreadpool", |
| 1664 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1665 | ) |
| 1666 | |
| 1667 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1668 | name = "avx_ukernels", |
| 1669 | hdrs = INTERNAL_HDRS, |
| 1670 | copts = xnnpack_std_copts(), |
| 1671 | x86_copts = ["-mavx"], |
| 1672 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1673 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1674 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1675 | "@FP16", |
| 1676 | "@pthreadpool", |
| 1677 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1678 | ) |
| 1679 | |
| 1680 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1681 | name = "fma3_ukernels", |
| 1682 | hdrs = INTERNAL_HDRS, |
| 1683 | copts = xnnpack_std_copts(), |
| 1684 | x86_copts = [ |
| 1685 | "-mfma", |
| 1686 | ], |
| 1687 | x86_srcs = FMA3_UKERNELS, |
| 1688 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1689 | ":tables", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1690 | "@FP16", |
| 1691 | "@pthreadpool", |
| 1692 | ], |
| 1693 | ) |
| 1694 | |
| 1695 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1696 | name = "avx2_ukernels", |
| 1697 | hdrs = INTERNAL_HDRS, |
| 1698 | copts = xnnpack_std_copts(), |
| 1699 | x86_copts = [ |
| 1700 | "-mfma", |
| 1701 | "-mavx2", |
| 1702 | ], |
| 1703 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1704 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1705 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1706 | "@FP16", |
| 1707 | "@pthreadpool", |
| 1708 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1709 | ) |
| 1710 | |
| 1711 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1712 | name = "avx512f_ukernels", |
| 1713 | hdrs = INTERNAL_HDRS, |
| 1714 | copts = xnnpack_std_copts(), |
| 1715 | x86_copts = ["-mavx512f"], |
| 1716 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1717 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1718 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1719 | "@FP16", |
| 1720 | "@pthreadpool", |
| 1721 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1722 | ) |
| 1723 | |
| 1724 | xnnpack_cc_library( |
| 1725 | name = "asm_ukernels", |
| 1726 | hdrs = ["src/xnnpack/assembly.h"], |
| 1727 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
| 1728 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 1729 | ) |
| 1730 | |
| 1731 | xnnpack_aggregate_library( |
| 1732 | name = "ukernels", |
| 1733 | aarch32_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1734 | ":psimd_fastmath_ukernels", |
| 1735 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1736 | ":neon_ukernels", |
| 1737 | ":neonfma_ukernels", |
| 1738 | ":asm_ukernels", |
| 1739 | ], |
| 1740 | aarch64_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1741 | ":psimd_fastmath_ukernels", |
| 1742 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1743 | ":neon_ukernels", |
| 1744 | ":neonfma_ukernels", |
| 1745 | ":neonfp16arith_ukernels", |
| 1746 | ":asm_ukernels", |
| 1747 | ], |
| 1748 | generic_deps = [":scalar_ukernels"], |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1749 | wasm_deps = [ |
| 1750 | ":wasm_ukernels", |
| 1751 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1752 | wasmsimd_deps = [ |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1753 | ":wasm_ukernels", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1754 | ":psimd_fastmath_ukernels", |
| 1755 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1756 | ], |
| 1757 | x86_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1758 | ":psimd_fastmath_ukernels", |
| 1759 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1760 | ":sse2_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1761 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1762 | ":avx_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1763 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1764 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1765 | ":avx512f_ukernels", |
| 1766 | ], |
| 1767 | ) |
| 1768 | |
| 1769 | xnnpack_cc_library( |
| 1770 | name = "im2col", |
| 1771 | srcs = ["src/im2col.c"], |
| 1772 | hdrs = [ |
| 1773 | "src/xnnpack/common.h", |
| 1774 | "src/xnnpack/im2col.h", |
| 1775 | ], |
| 1776 | copts = xnnpack_std_copts(), |
| 1777 | ) |
| 1778 | |
| 1779 | xnnpack_cc_library( |
| 1780 | name = "indirection", |
| 1781 | srcs = ["src/indirection.c"], |
| 1782 | hdrs = INTERNAL_HDRS, |
| 1783 | copts = xnnpack_std_copts(), |
| 1784 | deps = [ |
| 1785 | "@FP16", |
| 1786 | "@FXdiv", |
| 1787 | "@pthreadpool", |
| 1788 | ], |
| 1789 | ) |
| 1790 | |
| 1791 | xnnpack_cc_library( |
| 1792 | name = "operator_run", |
| 1793 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1794 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 1795 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1796 | # Wrappers for multi-pass microkernels use VLAs for temporary buffers. |
| 1797 | "-Wno-vla", |
| 1798 | ], |
| 1799 | deps = [ |
| 1800 | "@FP16", |
| 1801 | "@FXdiv", |
| 1802 | "@clog", |
| 1803 | "@pthreadpool", |
| 1804 | ], |
| 1805 | ) |
| 1806 | |
| 1807 | cc_library( |
| 1808 | name = "enable_assembly", |
| 1809 | defines = select({ |
| 1810 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 1811 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 1812 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1813 | }), |
| 1814 | ) |
| 1815 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1816 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1817 | name = "operators", |
| 1818 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1819 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1820 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1821 | ], |
| 1822 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1823 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1824 | "-Isrc", |
| 1825 | "-Iinclude", |
| 1826 | ] + select({ |
| 1827 | ":debug_build": [], |
| 1828 | "//conditions:default": xnnpack_min_size_copts(), |
| 1829 | }), |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1830 | wasm_srcs = ["src/wasm-stubs.c"], |
| 1831 | wasmsimd_srcs = ["src/wasm-stubs.c"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1832 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1833 | ":indirection", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1834 | "@FP16", |
| 1835 | "@FXdiv", |
| 1836 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1837 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1838 | ], |
| 1839 | ) |
| 1840 | |
| 1841 | cc_library( |
| 1842 | name = "XNNPACK", |
| 1843 | srcs = [ |
| 1844 | "src/init.c", |
| 1845 | ], |
| 1846 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1847 | "-Isrc", |
| 1848 | "-Iinclude", |
| 1849 | ] + select({ |
| 1850 | ":debug_build": [], |
| 1851 | "//conditions:default": xnnpack_min_size_copts(), |
| 1852 | }), |
| 1853 | includes = ["include"], |
| 1854 | linkstatic = True, |
| 1855 | textual_hdrs = ["include/xnnpack.h"], |
| 1856 | visibility = xnnpack_visibility(), |
| 1857 | deps = [ |
| 1858 | ":enable_assembly", |
| 1859 | ":ukernels", |
| 1860 | ":operator_run", |
| 1861 | ":operators", |
| 1862 | "@clog", |
| 1863 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 1864 | ] + select({ |
| 1865 | ":emscripten": [], |
| 1866 | "//conditions:default": ["@cpuinfo"], |
| 1867 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1868 | ) |
| 1869 | |
| 1870 | cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1871 | name = "xnnpack_operators_nhwc_f32", |
| 1872 | srcs = [ |
| 1873 | "src/init.c", |
| 1874 | ], |
| 1875 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1876 | "-Isrc", |
| 1877 | "-Iinclude", |
| 1878 | ] + select({ |
| 1879 | ":debug_build": [], |
| 1880 | "//conditions:default": xnnpack_min_size_copts(), |
| 1881 | }), |
| 1882 | defines = [ |
| 1883 | "XNN_NO_Q8_OPERATORS", |
| 1884 | "XNN_NO_U8_OPERATORS", |
| 1885 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1886 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1887 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1888 | includes = ["include"], |
| 1889 | linkstatic = True, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1890 | textual_hdrs = ["include/xnnpack.h"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1891 | visibility = xnnpack_visibility(), |
| 1892 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1893 | ":enable_assembly", |
| 1894 | ":ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1895 | ":operator_run", |
| 1896 | ":operators", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1897 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1898 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1899 | ] + select({ |
| 1900 | ":emscripten": [], |
| 1901 | "//conditions:default": ["@cpuinfo"], |
| 1902 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1903 | ) |
| 1904 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1905 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1906 | name = "bench_utils", |
| 1907 | srcs = ["bench/utils.cc"], |
| 1908 | hdrs = ["bench/utils.h"], |
| 1909 | copts = ["-Wno-unused-result"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 1910 | deps = [ |
| 1911 | "@com_google_benchmark//:benchmark", |
| 1912 | "@cpuinfo", |
| 1913 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1914 | ) |
| 1915 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1916 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1917 | |
| 1918 | xnnpack_benchmark( |
| 1919 | name = "q8_gemm_bench", |
| 1920 | srcs = [ |
| 1921 | "bench/gemm.h", |
| 1922 | "bench/q8-gemm.cc", |
| 1923 | "src/xnnpack/AlignedAllocator.h", |
| 1924 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1925 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 1926 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
| 1927 | ) |
| 1928 | |
| 1929 | xnnpack_benchmark( |
| 1930 | name = "f16_gemm_bench", |
| 1931 | srcs = [ |
| 1932 | "bench/f16-gemm.cc", |
| 1933 | "bench/gemm.h", |
| 1934 | "src/xnnpack/AlignedAllocator.h", |
| 1935 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1936 | copts = ["-Wno-unused-function"], |
| 1937 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1938 | ) |
| 1939 | |
| 1940 | xnnpack_benchmark( |
| 1941 | name = "f32_igemm_bench", |
| 1942 | srcs = [ |
| 1943 | "bench/f32-igemm.cc", |
| 1944 | "bench/conv.h", |
| 1945 | "src/xnnpack/AlignedAllocator.h", |
| 1946 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1947 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1948 | ) |
| 1949 | |
| 1950 | xnnpack_benchmark( |
| 1951 | name = "f32_conv_hwc_bench", |
| 1952 | srcs = [ |
| 1953 | "bench/f32-conv-hwc.cc", |
| 1954 | "bench/dconv.h", |
| 1955 | "src/xnnpack/AlignedAllocator.h", |
| 1956 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1957 | copts = ["-Wno-unused-function"], |
| 1958 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1959 | ) |
| 1960 | |
| 1961 | xnnpack_benchmark( |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 1962 | name = "f32_conv_hwc2spchw_bench", |
| 1963 | srcs = [ |
| 1964 | "bench/f32-conv-hwc2spchw.cc", |
| 1965 | "bench/dconv.h", |
| 1966 | "src/xnnpack/AlignedAllocator.h", |
| 1967 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1968 | copts = ["-Wno-unused-function"], |
| 1969 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 1970 | ) |
| 1971 | |
| 1972 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1973 | name = "f32_dwconv_bench", |
| 1974 | srcs = [ |
| 1975 | "bench/f32-dwconv.cc", |
| 1976 | "bench/dwconv.h", |
| 1977 | "src/xnnpack/AlignedAllocator.h", |
| 1978 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 1979 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 1980 | ) |
| 1981 | |
| 1982 | xnnpack_benchmark( |
| 1983 | name = "f32_dwconv_spchw_bench", |
| 1984 | srcs = [ |
| 1985 | "bench/f32-dwconv-spchw.cc", |
| 1986 | "bench/dwconv.h", |
| 1987 | "src/xnnpack/AlignedAllocator.h", |
| 1988 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1989 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 1990 | ) |
| 1991 | |
| 1992 | xnnpack_benchmark( |
| 1993 | name = "f32_gemm_bench", |
| 1994 | srcs = [ |
| 1995 | "bench/f32-gemm.cc", |
| 1996 | "bench/gemm.h", |
| 1997 | "src/xnnpack/AlignedAllocator.h", |
| 1998 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1999 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(), |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 2000 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2001 | ) |
| 2002 | |
| 2003 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2004 | name = "f32_raddexpminusmax_bench", |
| 2005 | srcs = [ |
| 2006 | "bench/f32-raddexpminusmax.cc", |
| 2007 | "src/xnnpack/AlignedAllocator.h", |
| 2008 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2009 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2010 | ) |
| 2011 | |
| 2012 | xnnpack_benchmark( |
| 2013 | name = "f32_raddextexp_bench", |
| 2014 | srcs = [ |
| 2015 | "bench/f32-raddextexp.cc", |
| 2016 | "src/xnnpack/AlignedAllocator.h", |
| 2017 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2018 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2019 | ) |
| 2020 | |
| 2021 | xnnpack_benchmark( |
| 2022 | name = "f32_raddstoreexpminusmax_bench", |
| 2023 | srcs = [ |
| 2024 | "bench/f32-raddstoreexpminusmax.cc", |
| 2025 | "src/xnnpack/AlignedAllocator.h", |
| 2026 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2027 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2028 | ) |
| 2029 | |
| 2030 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2031 | name = "f32_rmax_bench", |
| 2032 | srcs = [ |
| 2033 | "bench/f32-rmax.cc", |
| 2034 | "src/xnnpack/AlignedAllocator.h", |
| 2035 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2036 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2037 | ) |
| 2038 | |
| 2039 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 2040 | name = "f32_sigmoid_bench", |
| 2041 | srcs = [ |
| 2042 | "bench/f32-sigmoid.cc", |
| 2043 | "src/xnnpack/AlignedAllocator.h", |
| 2044 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2045 | copts = ["-Wno-unused-function"], |
| 2046 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2047 | ) |
| 2048 | |
| 2049 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2050 | name = "f32_spmm_bench", |
| 2051 | srcs = [ |
| 2052 | "bench/f32-spmm.cc", |
| 2053 | "bench/gemm.h", |
| 2054 | "src/xnnpack/AlignedAllocator.h", |
| 2055 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2056 | copts = ["-Wno-unused-function"], |
| 2057 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2058 | ) |
| 2059 | |
| 2060 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 2061 | name = "f32_softmax_bench", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 2062 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 2063 | "bench/f32-softmax.cc", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 2064 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2065 | copts = ["-Wno-unused-function"], |
| 2066 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2067 | ) |
| 2068 | |
| 2069 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2070 | name = "f32_vscaleexpminusmax_bench", |
| 2071 | srcs = [ |
| 2072 | "bench/f32-vscaleexpminusmax.cc", |
| 2073 | "src/xnnpack/AlignedAllocator.h", |
| 2074 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2075 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2076 | ) |
| 2077 | |
| 2078 | xnnpack_benchmark( |
| 2079 | name = "f32_vscaleextexp_bench", |
| 2080 | srcs = [ |
| 2081 | "bench/f32-vscaleextexp.cc", |
| 2082 | "src/xnnpack/AlignedAllocator.h", |
| 2083 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2084 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2085 | ) |
| 2086 | |
| 2087 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2088 | name = "f32_im2col_gemm_bench", |
| 2089 | srcs = [ |
| 2090 | "bench/f32-im2col-gemm.cc", |
| 2091 | "bench/conv.h", |
| 2092 | "src/xnnpack/AlignedAllocator.h", |
| 2093 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2094 | deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"], |
| 2095 | ) |
| 2096 | |
| 2097 | ########################### Benchmarks for operators ########################### |
| 2098 | |
| 2099 | xnnpack_benchmark( |
| 2100 | name = "add_bench", |
| 2101 | srcs = ["bench/add.cc"], |
| 2102 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2103 | ) |
| 2104 | |
| 2105 | xnnpack_benchmark( |
| 2106 | name = "average_pooling_bench", |
| 2107 | srcs = ["bench/average-pooling.cc"], |
| 2108 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2109 | ) |
| 2110 | |
| 2111 | xnnpack_benchmark( |
| 2112 | name = "channel_shuffle_bench", |
| 2113 | srcs = ["bench/channel-shuffle.cc"], |
| 2114 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2115 | ) |
| 2116 | |
| 2117 | xnnpack_benchmark( |
| 2118 | name = "convolution_bench", |
| 2119 | srcs = ["bench/convolution.cc"], |
| 2120 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
| 2121 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
| 2122 | ) |
| 2123 | |
| 2124 | xnnpack_benchmark( |
| 2125 | name = "deconvolution_bench", |
| 2126 | srcs = ["bench/deconvolution.cc"], |
| 2127 | copts = xnnpack_optional_tflite_copts(), |
| 2128 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 2129 | ) |
| 2130 | |
| 2131 | xnnpack_benchmark( |
| 2132 | name = "global_average_pooling_bench", |
| 2133 | srcs = ["bench/global-average-pooling.cc"], |
| 2134 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2135 | ) |
| 2136 | |
| 2137 | xnnpack_benchmark( |
| 2138 | name = "max_pooling_bench", |
| 2139 | srcs = ["bench/max-pooling.cc"], |
| 2140 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2141 | ) |
| 2142 | |
| 2143 | xnnpack_benchmark( |
| 2144 | name = "sigmoid_bench", |
| 2145 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 2146 | copts = xnnpack_optional_tflite_copts(), |
| 2147 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2148 | ) |
| 2149 | |
| 2150 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 2151 | name = "prelu_bench", |
| 2152 | srcs = ["bench/prelu.cc"], |
| 2153 | copts = xnnpack_optional_tflite_copts(), |
| 2154 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 2155 | ) |
| 2156 | |
| 2157 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 2158 | name = "softmax_bench", |
| 2159 | srcs = ["bench/softmax.cc"], |
Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame^] | 2160 | copts = xnnpack_optional_tflite_copts(), |
| 2161 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2162 | ) |
| 2163 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2164 | ############################# End-to-end benchmarks ############################ |
| 2165 | |
| 2166 | cc_library( |
| 2167 | name = "mobilenet_v1", |
| 2168 | srcs = ["models/mobilenet-v1.cc"], |
| 2169 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2170 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2171 | linkstatic = True, |
| 2172 | deps = [ |
| 2173 | ":XNNPACK", |
| 2174 | "@pthreadpool", |
| 2175 | ], |
| 2176 | ) |
| 2177 | |
| 2178 | cc_library( |
| 2179 | name = "mobilenet_v2", |
| 2180 | srcs = ["models/mobilenet-v2.cc"], |
| 2181 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2182 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2183 | linkstatic = True, |
| 2184 | deps = [ |
| 2185 | ":XNNPACK", |
| 2186 | "@pthreadpool", |
| 2187 | ], |
| 2188 | ) |
| 2189 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2190 | cc_library( |
| 2191 | name = "mobilenet_v3_large", |
| 2192 | srcs = ["models/mobilenet-v3-large.cc"], |
| 2193 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2194 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2195 | linkstatic = True, |
| 2196 | deps = [ |
| 2197 | ":XNNPACK", |
| 2198 | "@pthreadpool", |
| 2199 | ], |
| 2200 | ) |
| 2201 | |
| 2202 | cc_library( |
| 2203 | name = "mobilenet_v3_small", |
| 2204 | srcs = ["models/mobilenet-v3-small.cc"], |
| 2205 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2206 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2207 | linkstatic = True, |
| 2208 | deps = [ |
| 2209 | ":XNNPACK", |
| 2210 | "@pthreadpool", |
| 2211 | ], |
| 2212 | ) |
| 2213 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2214 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 2215 | name = "f32_dwconv_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2216 | srcs = [ |
| 2217 | "bench/f32-dwconv-e2e.cc", |
| 2218 | "bench/end2end.h", |
| 2219 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 2220 | copts = ["-Wno-unused-function"], |
| 2221 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 2222 | ":XNNPACK", |
| 2223 | ":mobilenet_v1", |
| 2224 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2225 | ":mobilenet_v3_large", |
| 2226 | ":mobilenet_v3_small", |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 2227 | ], |
| 2228 | ) |
| 2229 | |
| 2230 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 2231 | name = "f32_gemm_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2232 | srcs = [ |
| 2233 | "bench/f32-gemm-e2e.cc", |
| 2234 | "bench/end2end.h", |
| 2235 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 2236 | copts = ["-Wno-unused-function"], |
| 2237 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 2238 | ":XNNPACK", |
| 2239 | ":mobilenet_v1", |
| 2240 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2241 | ":mobilenet_v3_large", |
| 2242 | ":mobilenet_v3_small", |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 2243 | ], |
| 2244 | ) |
| 2245 | |
| 2246 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2247 | name = "end2end_bench", |
| 2248 | srcs = ["bench/end2end.cc"], |
| 2249 | deps = [ |
| 2250 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 2251 | ":bench_utils", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2252 | ":mobilenet_v1", |
| 2253 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2254 | ":mobilenet_v3_large", |
| 2255 | ":mobilenet_v3_small", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2256 | "@pthreadpool", |
| 2257 | ], |
| 2258 | ) |
| 2259 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2260 | #################### Accuracy evaluation for math functions #################### |
| 2261 | |
| 2262 | xnnpack_benchmark( |
| 2263 | name = "f32_exp_eval", |
| 2264 | srcs = [ |
| 2265 | "eval/f32-exp.cc", |
| 2266 | "src/xnnpack/AlignedAllocator.h", |
| 2267 | ] + ACCURACY_EVAL_HDRS, |
| 2268 | deps = ACCURACY_EVAL_DEPS, |
| 2269 | ) |
| 2270 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 2271 | xnnpack_benchmark( |
| 2272 | name = "f32_expminus_eval", |
| 2273 | srcs = [ |
| 2274 | "eval/f32-expminus.cc", |
| 2275 | "src/xnnpack/AlignedAllocator.h", |
| 2276 | ] + ACCURACY_EVAL_HDRS, |
| 2277 | deps = ACCURACY_EVAL_DEPS, |
| 2278 | ) |
| 2279 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2280 | xnnpack_benchmark( |
| 2281 | name = "f32_extexp_eval", |
| 2282 | srcs = [ |
| 2283 | "eval/f32-extexp.cc", |
| 2284 | "src/xnnpack/AlignedAllocator.h", |
| 2285 | ] + ACCURACY_EVAL_HDRS, |
| 2286 | deps = ACCURACY_EVAL_DEPS, |
| 2287 | ) |
| 2288 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2289 | xnnpack_benchmark( |
| 2290 | name = "f32_sigmoid_eval", |
| 2291 | srcs = [ |
| 2292 | "eval/f32-sigmoid.cc", |
| 2293 | "src/xnnpack/AlignedAllocator.h", |
| 2294 | ] + ACCURACY_EVAL_HDRS, |
| 2295 | deps = ACCURACY_EVAL_DEPS, |
| 2296 | ) |
| 2297 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2298 | ######################### Unit tests for micro-kernels ######################### |
| 2299 | |
| 2300 | xnnpack_unit_test( |
| 2301 | name = "f16_gemm_test", |
| 2302 | srcs = [ |
| 2303 | "test/f16-gemm.cc", |
| 2304 | "test/gemm-microkernel-tester.h", |
| 2305 | "src/xnnpack/AlignedAllocator.h", |
| 2306 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2307 | deps = MICROKERNEL_TEST_DEPS, |
| 2308 | ) |
| 2309 | |
| 2310 | xnnpack_unit_test( |
| 2311 | name = "f32_argmaxpool_test", |
| 2312 | srcs = [ |
| 2313 | "test/f32-argmaxpool.cc", |
| 2314 | "test/argmaxpool-microkernel-tester.h", |
| 2315 | "src/xnnpack/AlignedAllocator.h", |
| 2316 | ] + MICROKERNEL_TEST_HDRS, |
| 2317 | deps = MICROKERNEL_TEST_DEPS, |
| 2318 | ) |
| 2319 | |
| 2320 | xnnpack_unit_test( |
| 2321 | name = "f32_avgpool_test", |
| 2322 | srcs = [ |
| 2323 | "test/f32-avgpool.cc", |
| 2324 | "test/avgpool-microkernel-tester.h", |
| 2325 | "src/xnnpack/AlignedAllocator.h", |
| 2326 | ] + MICROKERNEL_TEST_HDRS, |
| 2327 | deps = MICROKERNEL_TEST_DEPS, |
| 2328 | ) |
| 2329 | |
| 2330 | xnnpack_unit_test( |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 2331 | name = "f32_bilinear_test", |
| 2332 | srcs = [ |
| 2333 | "test/f32-bilinear.cc", |
| 2334 | "test/bilinear-microkernel-tester.h", |
| 2335 | "src/xnnpack/AlignedAllocator.h", |
| 2336 | ] + MICROKERNEL_TEST_HDRS, |
| 2337 | deps = MICROKERNEL_TEST_DEPS, |
| 2338 | ) |
| 2339 | |
| 2340 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2341 | name = "f32_clamp_test", |
| 2342 | srcs = [ |
| 2343 | "test/f32-clamp.cc", |
| 2344 | "test/clamp-microkernel-tester.h", |
| 2345 | ] + MICROKERNEL_TEST_HDRS, |
| 2346 | deps = MICROKERNEL_TEST_DEPS, |
| 2347 | ) |
| 2348 | |
| 2349 | xnnpack_unit_test( |
| 2350 | name = "f32_igemm_test", |
| 2351 | srcs = [ |
| 2352 | "test/f32-igemm.cc", |
| 2353 | "test/gemm-microkernel-tester.h", |
| 2354 | "src/xnnpack/AlignedAllocator.h", |
| 2355 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2356 | deps = MICROKERNEL_TEST_DEPS, |
| 2357 | ) |
| 2358 | |
| 2359 | xnnpack_unit_test( |
| 2360 | name = "f32_conv_hwc_test", |
| 2361 | srcs = [ |
| 2362 | "test/f32-conv-hwc.cc", |
| 2363 | "test/conv-hwc-microkernel-tester.h", |
| 2364 | "src/xnnpack/AlignedAllocator.h", |
| 2365 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2366 | deps = MICROKERNEL_TEST_DEPS, |
| 2367 | ) |
| 2368 | |
| 2369 | xnnpack_unit_test( |
| 2370 | name = "f32_conv_hwc2spchw_test", |
| 2371 | srcs = [ |
| 2372 | "test/f32-conv-hwc2spchw.cc", |
| 2373 | "test/conv-hwc2spchw-microkernel-tester.h", |
| 2374 | "src/xnnpack/AlignedAllocator.h", |
| 2375 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2376 | deps = MICROKERNEL_TEST_DEPS, |
| 2377 | ) |
| 2378 | |
| 2379 | xnnpack_unit_test( |
| 2380 | name = "f32_dwconv_test", |
| 2381 | srcs = [ |
| 2382 | "test/f32-dwconv.cc", |
| 2383 | "test/dwconv-microkernel-tester.h", |
| 2384 | "src/xnnpack/AlignedAllocator.h", |
| 2385 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2386 | deps = MICROKERNEL_TEST_DEPS, |
| 2387 | ) |
| 2388 | |
| 2389 | xnnpack_unit_test( |
| 2390 | name = "f32_dwconv_spchw_test", |
| 2391 | srcs = [ |
| 2392 | "test/f32-dwconv-spchw.cc", |
| 2393 | "test/dwconv-spchw-microkernel-tester.h", |
| 2394 | "src/xnnpack/AlignedAllocator.h", |
| 2395 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2396 | deps = MICROKERNEL_TEST_DEPS, |
| 2397 | ) |
| 2398 | |
| 2399 | xnnpack_unit_test( |
| 2400 | name = "f32_gavgpool_test", |
| 2401 | srcs = [ |
| 2402 | "test/f32-gavgpool.cc", |
| 2403 | "test/gavgpool-microkernel-tester.h", |
| 2404 | "src/xnnpack/AlignedAllocator.h", |
| 2405 | ] + MICROKERNEL_TEST_HDRS, |
| 2406 | deps = MICROKERNEL_TEST_DEPS, |
| 2407 | ) |
| 2408 | |
| 2409 | xnnpack_unit_test( |
| 2410 | name = "f32_gavgpool_spchw_test", |
| 2411 | srcs = [ |
| 2412 | "test/f32-gavgpool-spchw.cc", |
| 2413 | "test/gavgpool-spchw-microkernel-tester.h", |
| 2414 | "src/xnnpack/AlignedAllocator.h", |
| 2415 | ] + MICROKERNEL_TEST_HDRS, |
| 2416 | deps = MICROKERNEL_TEST_DEPS, |
| 2417 | ) |
| 2418 | |
| 2419 | xnnpack_unit_test( |
| 2420 | name = "f32_gemm_test", |
| 2421 | srcs = [ |
| 2422 | "test/f32-gemm.cc", |
| 2423 | "test/gemm-microkernel-tester.h", |
| 2424 | "src/xnnpack/AlignedAllocator.h", |
| 2425 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2426 | deps = MICROKERNEL_TEST_DEPS, |
| 2427 | ) |
| 2428 | |
| 2429 | xnnpack_unit_test( |
| 2430 | name = "f32_gemminc_test", |
| 2431 | srcs = [ |
| 2432 | "test/f32-gemminc.cc", |
| 2433 | "test/gemm-microkernel-tester.h", |
| 2434 | "src/xnnpack/AlignedAllocator.h", |
| 2435 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2436 | deps = MICROKERNEL_TEST_DEPS, |
| 2437 | ) |
| 2438 | |
| 2439 | xnnpack_unit_test( |
| 2440 | name = "f32_hswish_test", |
| 2441 | srcs = [ |
| 2442 | "test/f32-hswish.cc", |
| 2443 | "test/hswish-microkernel-tester.h", |
| 2444 | ] + MICROKERNEL_TEST_HDRS, |
| 2445 | deps = MICROKERNEL_TEST_DEPS, |
| 2446 | ) |
| 2447 | |
| 2448 | xnnpack_unit_test( |
| 2449 | name = "f32_maxpool_test", |
| 2450 | srcs = [ |
| 2451 | "test/f32-maxpool.cc", |
| 2452 | "test/maxpool-microkernel-tester.h", |
| 2453 | ] + MICROKERNEL_TEST_HDRS, |
| 2454 | deps = MICROKERNEL_TEST_DEPS, |
| 2455 | ) |
| 2456 | |
| 2457 | xnnpack_unit_test( |
| 2458 | name = "f32_pavgpool_test", |
| 2459 | srcs = [ |
| 2460 | "test/f32-pavgpool.cc", |
| 2461 | "test/avgpool-microkernel-tester.h", |
| 2462 | "src/xnnpack/AlignedAllocator.h", |
| 2463 | ] + MICROKERNEL_TEST_HDRS, |
| 2464 | deps = MICROKERNEL_TEST_DEPS, |
| 2465 | ) |
| 2466 | |
| 2467 | xnnpack_unit_test( |
| 2468 | name = "f32_ppmm_test", |
| 2469 | srcs = [ |
| 2470 | "test/f32-ppmm.cc", |
| 2471 | "test/gemm-microkernel-tester.h", |
| 2472 | "src/xnnpack/AlignedAllocator.h", |
| 2473 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2474 | deps = MICROKERNEL_TEST_DEPS, |
| 2475 | ) |
| 2476 | |
| 2477 | xnnpack_unit_test( |
| 2478 | name = "f32_prelu_test", |
| 2479 | srcs = [ |
| 2480 | "test/f32-prelu.cc", |
| 2481 | "test/prelu-microkernel-tester.h", |
| 2482 | "src/xnnpack/AlignedAllocator.h", |
| 2483 | ] + MICROKERNEL_TEST_HDRS, |
| 2484 | deps = MICROKERNEL_TEST_DEPS, |
| 2485 | ) |
| 2486 | |
| 2487 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2488 | name = "f32_raddexpminusmax_test", |
| 2489 | srcs = [ |
| 2490 | "test/f32-raddexpminusmax.cc", |
| 2491 | "test/raddexpminusmax-microkernel-tester.h", |
| 2492 | ] + MICROKERNEL_TEST_HDRS, |
| 2493 | deps = MICROKERNEL_TEST_DEPS, |
| 2494 | ) |
| 2495 | |
| 2496 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2497 | name = "f32_raddextexp_test", |
| 2498 | srcs = [ |
| 2499 | "test/f32-raddextexp.cc", |
| 2500 | "test/raddextexp-microkernel-tester.h", |
| 2501 | ] + MICROKERNEL_TEST_HDRS, |
| 2502 | deps = MICROKERNEL_TEST_DEPS, |
| 2503 | ) |
| 2504 | |
| 2505 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2506 | name = "f32_raddstoreexpminusmax_test", |
| 2507 | srcs = [ |
| 2508 | "test/f32-raddstoreexpminusmax.cc", |
| 2509 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 2510 | ] + MICROKERNEL_TEST_HDRS, |
| 2511 | deps = MICROKERNEL_TEST_DEPS, |
| 2512 | ) |
| 2513 | |
| 2514 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2515 | name = "f32_rmax_test", |
| 2516 | srcs = [ |
| 2517 | "test/f32-rmax.cc", |
| 2518 | "test/rmax-microkernel-tester.h", |
| 2519 | ] + MICROKERNEL_TEST_HDRS, |
| 2520 | deps = MICROKERNEL_TEST_DEPS, |
| 2521 | ) |
| 2522 | |
| 2523 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2524 | name = "f32_sigmoid_test", |
| 2525 | srcs = [ |
| 2526 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2527 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2528 | ] + MICROKERNEL_TEST_HDRS, |
| 2529 | deps = MICROKERNEL_TEST_DEPS, |
| 2530 | ) |
| 2531 | |
| 2532 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2533 | name = "f32_spmm_test", |
| 2534 | srcs = [ |
| 2535 | "test/f32-spmm.cc", |
| 2536 | "test/spmm-microkernel-tester.h", |
| 2537 | "src/xnnpack/AlignedAllocator.h", |
| 2538 | ] + MICROKERNEL_TEST_HDRS, |
| 2539 | deps = MICROKERNEL_TEST_DEPS, |
| 2540 | ) |
| 2541 | |
| 2542 | xnnpack_unit_test( |
| 2543 | name = "f32_vadd_test", |
| 2544 | srcs = [ |
| 2545 | "test/f32-vadd.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2546 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2547 | ] + MICROKERNEL_TEST_HDRS, |
| 2548 | deps = MICROKERNEL_TEST_DEPS, |
| 2549 | ) |
| 2550 | |
| 2551 | xnnpack_unit_test( |
| 2552 | name = "f32_vaddc_test", |
| 2553 | srcs = [ |
| 2554 | "test/f32-vaddc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2555 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2556 | ] + MICROKERNEL_TEST_HDRS, |
| 2557 | deps = MICROKERNEL_TEST_DEPS, |
| 2558 | ) |
| 2559 | |
| 2560 | xnnpack_unit_test( |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 2561 | name = "f32_vdiv_test", |
| 2562 | srcs = [ |
| 2563 | "test/f32-vdiv.cc", |
| 2564 | "test/vbinary-microkernel-tester.h", |
| 2565 | ] + MICROKERNEL_TEST_HDRS, |
| 2566 | deps = MICROKERNEL_TEST_DEPS, |
| 2567 | ) |
| 2568 | |
| 2569 | xnnpack_unit_test( |
| 2570 | name = "f32_vdivc_test", |
| 2571 | srcs = [ |
| 2572 | "test/f32-vdivc.cc", |
| 2573 | "test/vbinaryc-microkernel-tester.h", |
| 2574 | ] + MICROKERNEL_TEST_HDRS, |
| 2575 | deps = MICROKERNEL_TEST_DEPS, |
| 2576 | ) |
| 2577 | |
| 2578 | xnnpack_unit_test( |
| 2579 | name = "f32_vrdivc_test", |
| 2580 | srcs = [ |
| 2581 | "test/f32-vrdivc.cc", |
| 2582 | "test/vbinaryc-microkernel-tester.h", |
| 2583 | ] + MICROKERNEL_TEST_HDRS, |
| 2584 | deps = MICROKERNEL_TEST_DEPS, |
| 2585 | ) |
| 2586 | |
| 2587 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 2588 | name = "f32_vmax_test", |
| 2589 | srcs = [ |
| 2590 | "test/f32-vmax.cc", |
| 2591 | "test/vbinary-microkernel-tester.h", |
| 2592 | ] + MICROKERNEL_TEST_HDRS, |
| 2593 | deps = MICROKERNEL_TEST_DEPS, |
| 2594 | ) |
| 2595 | |
| 2596 | xnnpack_unit_test( |
| 2597 | name = "f32_vmaxc_test", |
| 2598 | srcs = [ |
| 2599 | "test/f32-vmaxc.cc", |
| 2600 | "test/vbinaryc-microkernel-tester.h", |
| 2601 | ] + MICROKERNEL_TEST_HDRS, |
| 2602 | deps = MICROKERNEL_TEST_DEPS, |
| 2603 | ) |
| 2604 | |
| 2605 | xnnpack_unit_test( |
| 2606 | name = "f32_vmin_test", |
| 2607 | srcs = [ |
| 2608 | "test/f32-vmin.cc", |
| 2609 | "test/vbinary-microkernel-tester.h", |
| 2610 | ] + MICROKERNEL_TEST_HDRS, |
| 2611 | deps = MICROKERNEL_TEST_DEPS, |
| 2612 | ) |
| 2613 | |
| 2614 | xnnpack_unit_test( |
| 2615 | name = "f32_vminc_test", |
| 2616 | srcs = [ |
| 2617 | "test/f32-vminc.cc", |
| 2618 | "test/vbinaryc-microkernel-tester.h", |
| 2619 | ] + MICROKERNEL_TEST_HDRS, |
| 2620 | deps = MICROKERNEL_TEST_DEPS, |
| 2621 | ) |
| 2622 | |
| 2623 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2624 | name = "f32_vmul_test", |
| 2625 | srcs = [ |
| 2626 | "test/f32-vmul.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2627 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2628 | ] + MICROKERNEL_TEST_HDRS, |
| 2629 | deps = MICROKERNEL_TEST_DEPS, |
| 2630 | ) |
| 2631 | |
| 2632 | xnnpack_unit_test( |
| 2633 | name = "f32_vmulc_test", |
| 2634 | srcs = [ |
| 2635 | "test/f32-vmulc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2636 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2637 | ] + MICROKERNEL_TEST_HDRS, |
| 2638 | deps = MICROKERNEL_TEST_DEPS, |
| 2639 | ) |
| 2640 | |
| 2641 | xnnpack_unit_test( |
| 2642 | name = "f32_vmulcaddc_test", |
| 2643 | srcs = [ |
| 2644 | "test/f32-vmulcaddc.cc", |
| 2645 | "test/vmulcaddc-microkernel-tester.h", |
| 2646 | "src/xnnpack/AlignedAllocator.h", |
| 2647 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2648 | deps = MICROKERNEL_TEST_DEPS, |
| 2649 | ) |
| 2650 | |
| 2651 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2652 | name = "f32_vscale_test", |
| 2653 | srcs = [ |
| 2654 | "test/f32-vscale.cc", |
| 2655 | "test/vscale-microkernel-tester.h", |
| 2656 | ] + MICROKERNEL_TEST_HDRS, |
| 2657 | deps = MICROKERNEL_TEST_DEPS, |
| 2658 | ) |
| 2659 | |
| 2660 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2661 | name = "f32_vscaleexpminusmax_test", |
| 2662 | srcs = [ |
| 2663 | "test/f32-vscaleexpminusmax.cc", |
| 2664 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 2665 | ] + MICROKERNEL_TEST_HDRS, |
| 2666 | deps = MICROKERNEL_TEST_DEPS, |
| 2667 | ) |
| 2668 | |
| 2669 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2670 | name = "f32_vscaleextexp_test", |
| 2671 | srcs = [ |
| 2672 | "test/f32-vscaleextexp.cc", |
| 2673 | "test/vscaleextexp-microkernel-tester.h", |
| 2674 | ] + MICROKERNEL_TEST_HDRS, |
| 2675 | deps = MICROKERNEL_TEST_DEPS, |
| 2676 | ) |
| 2677 | |
| 2678 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2679 | name = "f32_vsub_test", |
| 2680 | srcs = [ |
| 2681 | "test/f32-vsub.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2682 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2683 | ] + MICROKERNEL_TEST_HDRS, |
| 2684 | deps = MICROKERNEL_TEST_DEPS, |
| 2685 | ) |
| 2686 | |
| 2687 | xnnpack_unit_test( |
| 2688 | name = "f32_vsubc_test", |
| 2689 | srcs = [ |
| 2690 | "test/f32-vsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2691 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2692 | ] + MICROKERNEL_TEST_HDRS, |
| 2693 | deps = MICROKERNEL_TEST_DEPS, |
| 2694 | ) |
| 2695 | |
| 2696 | xnnpack_unit_test( |
| 2697 | name = "f32_vrsubc_test", |
| 2698 | srcs = [ |
| 2699 | "test/f32-vrsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2700 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2701 | ] + MICROKERNEL_TEST_HDRS, |
| 2702 | deps = MICROKERNEL_TEST_DEPS, |
| 2703 | ) |
| 2704 | |
| 2705 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2706 | name = "q8_avgpool_test", |
| 2707 | srcs = [ |
| 2708 | "test/q8-avgpool.cc", |
| 2709 | "test/avgpool-microkernel-tester.h", |
| 2710 | "src/xnnpack/AlignedAllocator.h", |
| 2711 | ] + MICROKERNEL_TEST_HDRS, |
| 2712 | deps = MICROKERNEL_TEST_DEPS, |
| 2713 | ) |
| 2714 | |
| 2715 | xnnpack_unit_test( |
| 2716 | name = "q8_igemm_test", |
| 2717 | srcs = [ |
| 2718 | "test/q8-igemm.cc", |
| 2719 | "test/gemm-microkernel-tester.h", |
| 2720 | "src/xnnpack/AlignedAllocator.h", |
| 2721 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2722 | deps = MICROKERNEL_TEST_DEPS, |
| 2723 | ) |
| 2724 | |
| 2725 | xnnpack_unit_test( |
| 2726 | name = "q8_dwconv_test", |
| 2727 | srcs = [ |
| 2728 | "test/q8-dwconv.cc", |
| 2729 | "test/dwconv-microkernel-tester.h", |
| 2730 | "src/xnnpack/AlignedAllocator.h", |
| 2731 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2732 | deps = MICROKERNEL_TEST_DEPS, |
| 2733 | ) |
| 2734 | |
| 2735 | xnnpack_unit_test( |
| 2736 | name = "q8_gavgpool_test", |
| 2737 | srcs = [ |
| 2738 | "test/q8-gavgpool.cc", |
| 2739 | "test/gavgpool-microkernel-tester.h", |
| 2740 | "src/xnnpack/AlignedAllocator.h", |
| 2741 | ] + MICROKERNEL_TEST_HDRS, |
| 2742 | deps = MICROKERNEL_TEST_DEPS, |
| 2743 | ) |
| 2744 | |
| 2745 | xnnpack_unit_test( |
| 2746 | name = "q8_gemm_test", |
| 2747 | srcs = [ |
| 2748 | "test/q8-gemm.cc", |
| 2749 | "test/gemm-microkernel-tester.h", |
| 2750 | "src/xnnpack/AlignedAllocator.h", |
| 2751 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2752 | deps = MICROKERNEL_TEST_DEPS, |
| 2753 | ) |
| 2754 | |
| 2755 | xnnpack_unit_test( |
| 2756 | name = "q8_vadd_test", |
| 2757 | srcs = [ |
| 2758 | "test/q8-vadd.cc", |
| 2759 | "test/vadd-microkernel-tester.h", |
| 2760 | ] + MICROKERNEL_TEST_HDRS, |
| 2761 | deps = MICROKERNEL_TEST_DEPS, |
| 2762 | ) |
| 2763 | |
| 2764 | xnnpack_unit_test( |
| 2765 | name = "u8_clamp_test", |
| 2766 | srcs = [ |
| 2767 | "test/u8-clamp.cc", |
| 2768 | "test/clamp-microkernel-tester.h", |
| 2769 | ] + MICROKERNEL_TEST_HDRS, |
| 2770 | deps = MICROKERNEL_TEST_DEPS, |
| 2771 | ) |
| 2772 | |
| 2773 | xnnpack_unit_test( |
| 2774 | name = "u8_lut32norm_test", |
| 2775 | srcs = [ |
| 2776 | "test/u8-lut32norm.cc", |
| 2777 | "test/lut-norm-microkernel-tester.h", |
| 2778 | ] + MICROKERNEL_TEST_HDRS, |
| 2779 | deps = MICROKERNEL_TEST_DEPS, |
| 2780 | ) |
| 2781 | |
| 2782 | xnnpack_unit_test( |
| 2783 | name = "u8_maxpool_test", |
| 2784 | srcs = [ |
| 2785 | "test/u8-maxpool.cc", |
| 2786 | "test/maxpool-microkernel-tester.h", |
| 2787 | ] + MICROKERNEL_TEST_HDRS, |
| 2788 | deps = MICROKERNEL_TEST_DEPS, |
| 2789 | ) |
| 2790 | |
| 2791 | xnnpack_unit_test( |
| 2792 | name = "u8_rmax_test", |
| 2793 | srcs = [ |
| 2794 | "test/u8-rmax.cc", |
| 2795 | "test/rmax-microkernel-tester.h", |
| 2796 | ] + MICROKERNEL_TEST_HDRS, |
| 2797 | deps = MICROKERNEL_TEST_DEPS, |
| 2798 | ) |
| 2799 | |
| 2800 | xnnpack_unit_test( |
| 2801 | name = "x32_packx_test", |
| 2802 | srcs = [ |
| 2803 | "test/x32-packx.cc", |
| 2804 | "test/pack-microkernel-tester.h", |
| 2805 | "src/xnnpack/AlignedAllocator.h", |
| 2806 | ] + MICROKERNEL_TEST_HDRS, |
| 2807 | deps = MICROKERNEL_TEST_DEPS, |
| 2808 | ) |
| 2809 | |
| 2810 | xnnpack_unit_test( |
| 2811 | name = "x32_pad_test", |
| 2812 | srcs = [ |
| 2813 | "test/x32-pad.cc", |
| 2814 | "test/pad-microkernel-tester.h", |
| 2815 | ] + MICROKERNEL_TEST_HDRS, |
| 2816 | deps = MICROKERNEL_TEST_DEPS, |
| 2817 | ) |
| 2818 | |
| 2819 | xnnpack_unit_test( |
| 2820 | name = "x32_unpool_test", |
| 2821 | srcs = [ |
| 2822 | "test/x32-unpool.cc", |
| 2823 | "test/unpool-microkernel-tester.h", |
| 2824 | ] + MICROKERNEL_TEST_HDRS, |
| 2825 | deps = MICROKERNEL_TEST_DEPS, |
| 2826 | ) |
| 2827 | |
| 2828 | xnnpack_unit_test( |
| 2829 | name = "x32_zip_test", |
| 2830 | srcs = [ |
| 2831 | "test/x32-zip.cc", |
| 2832 | "test/zip-microkernel-tester.h", |
| 2833 | ] + MICROKERNEL_TEST_HDRS, |
| 2834 | deps = MICROKERNEL_TEST_DEPS, |
| 2835 | ) |
| 2836 | |
| 2837 | xnnpack_unit_test( |
| 2838 | name = "x8_lut_test", |
| 2839 | srcs = [ |
| 2840 | "test/x8-lut.cc", |
| 2841 | "test/lut-microkernel-tester.h", |
| 2842 | ] + MICROKERNEL_TEST_HDRS, |
| 2843 | deps = MICROKERNEL_TEST_DEPS, |
| 2844 | ) |
| 2845 | |
| 2846 | xnnpack_unit_test( |
| 2847 | name = "x8_zip_test", |
| 2848 | srcs = [ |
| 2849 | "test/x8-zip.cc", |
| 2850 | "test/zip-microkernel-tester.h", |
| 2851 | ] + MICROKERNEL_TEST_HDRS, |
| 2852 | deps = MICROKERNEL_TEST_DEPS, |
| 2853 | ) |
| 2854 | |
| 2855 | ########################### Size test for the library ########################## |
| 2856 | |
| 2857 | xnnpack_binary( |
| 2858 | name = "size_test", |
| 2859 | srcs = ["test/size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2860 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2861 | ) |
| 2862 | |
| 2863 | ########################### Unit tests for operators ########################### |
| 2864 | |
| 2865 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2866 | name = "add_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2867 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2868 | "test/add-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2869 | "test/add-operator-tester.h", |
| 2870 | ], |
| 2871 | deps = OPERATOR_TEST_DEPS, |
| 2872 | ) |
| 2873 | |
| 2874 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2875 | name = "add_nd_test", |
| 2876 | srcs = [ |
| 2877 | "test/add-nd.cc", |
| 2878 | "test/binary-elementwise-operator-tester.h", |
| 2879 | ], |
| 2880 | deps = OPERATOR_TEST_DEPS, |
| 2881 | ) |
| 2882 | |
| 2883 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2884 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2885 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2886 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2887 | "test/argmax-pooling-operator-tester.h", |
| 2888 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2889 | deps = OPERATOR_TEST_DEPS, |
| 2890 | ) |
| 2891 | |
| 2892 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2893 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2894 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2895 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2896 | "test/average-pooling-operator-tester.h", |
| 2897 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2898 | deps = OPERATOR_TEST_DEPS, |
| 2899 | ) |
| 2900 | |
| 2901 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2902 | name = "channel_pad_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2903 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2904 | "test/channel-pad-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2905 | "test/channel-pad-operator-tester.h", |
| 2906 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2907 | deps = OPERATOR_TEST_DEPS, |
| 2908 | ) |
| 2909 | |
| 2910 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2911 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2912 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2913 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2914 | "test/channel-shuffle-operator-tester.h", |
| 2915 | ], |
| 2916 | deps = OPERATOR_TEST_DEPS, |
| 2917 | ) |
| 2918 | |
| 2919 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2920 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2921 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2922 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2923 | "test/clamp-operator-tester.h", |
| 2924 | ], |
| 2925 | deps = OPERATOR_TEST_DEPS, |
| 2926 | ) |
| 2927 | |
| 2928 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2929 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2930 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2931 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2932 | "test/convolution-operator-tester.h", |
| 2933 | ], |
| 2934 | deps = OPERATOR_TEST_DEPS, |
| 2935 | ) |
| 2936 | |
| 2937 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2938 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2939 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2940 | "test/convolution-nchw.cc", |
| 2941 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2942 | ], |
| 2943 | deps = OPERATOR_TEST_DEPS, |
| 2944 | ) |
| 2945 | |
| 2946 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2947 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2948 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2949 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2950 | "test/deconvolution-operator-tester.h", |
| 2951 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2952 | deps = OPERATOR_TEST_DEPS, |
| 2953 | ) |
| 2954 | |
| 2955 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 2956 | name = "divide_nd_test", |
| 2957 | srcs = [ |
| 2958 | "test/binary-elementwise-operator-tester.h", |
| 2959 | "test/divide-nd.cc", |
| 2960 | ], |
| 2961 | deps = OPERATOR_TEST_DEPS, |
| 2962 | ) |
| 2963 | |
| 2964 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2965 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2966 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2967 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2968 | "test/fully-connected-operator-tester.h", |
| 2969 | ], |
| 2970 | deps = OPERATOR_TEST_DEPS, |
| 2971 | ) |
| 2972 | |
| 2973 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2974 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2975 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2976 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2977 | "test/global-average-pooling-operator-tester.h", |
| 2978 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 2979 | deps = OPERATOR_TEST_DEPS, |
| 2980 | ) |
| 2981 | |
| 2982 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2983 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2984 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2985 | "test/global-average-pooling-ncw.cc", |
| 2986 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2987 | ], |
| 2988 | deps = OPERATOR_TEST_DEPS, |
| 2989 | ) |
| 2990 | |
| 2991 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2992 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2993 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2994 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2995 | "test/hardswish-operator-tester.h", |
| 2996 | ], |
| 2997 | deps = OPERATOR_TEST_DEPS, |
| 2998 | ) |
| 2999 | |
| 3000 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3001 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3002 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3003 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3004 | "test/leaky-relu-operator-tester.h", |
| 3005 | ], |
| 3006 | deps = OPERATOR_TEST_DEPS, |
| 3007 | ) |
| 3008 | |
| 3009 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3010 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3011 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3012 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3013 | "test/max-pooling-operator-tester.h", |
| 3014 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3015 | deps = OPERATOR_TEST_DEPS, |
| 3016 | ) |
| 3017 | |
| 3018 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 3019 | name = "maximum_nd_test", |
| 3020 | srcs = [ |
| 3021 | "test/binary-elementwise-operator-tester.h", |
| 3022 | "test/maximum-nd.cc", |
| 3023 | ], |
| 3024 | deps = OPERATOR_TEST_DEPS, |
| 3025 | ) |
| 3026 | |
| 3027 | xnnpack_unit_test( |
| 3028 | name = "minimum_nd_test", |
| 3029 | srcs = [ |
| 3030 | "test/binary-elementwise-operator-tester.h", |
| 3031 | "test/minimum-nd.cc", |
| 3032 | ], |
| 3033 | deps = OPERATOR_TEST_DEPS, |
| 3034 | ) |
| 3035 | |
| 3036 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3037 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 3038 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 3039 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3040 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 3041 | ], |
| 3042 | deps = OPERATOR_TEST_DEPS, |
| 3043 | ) |
| 3044 | |
| 3045 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3046 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3047 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3048 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3049 | "test/prelu-operator-tester.h", |
| 3050 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3051 | deps = OPERATOR_TEST_DEPS, |
| 3052 | ) |
| 3053 | |
| 3054 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3055 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 3056 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3057 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 3058 | "test/resize-bilinear-operator-tester.h", |
| 3059 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3060 | deps = OPERATOR_TEST_DEPS, |
| 3061 | ) |
| 3062 | |
| 3063 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3064 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3065 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3066 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3067 | "test/sigmoid-operator-tester.h", |
| 3068 | ], |
| 3069 | deps = OPERATOR_TEST_DEPS, |
| 3070 | ) |
| 3071 | |
| 3072 | xnnpack_unit_test( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3073 | name = "softmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3074 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3075 | "test/softmax-nc.cc", |
| 3076 | "test/softmax-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3077 | ], |
| 3078 | deps = OPERATOR_TEST_DEPS, |
| 3079 | ) |
| 3080 | |
| 3081 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 3082 | name = "subtract_nd_test", |
| 3083 | srcs = [ |
| 3084 | "test/binary-elementwise-operator-tester.h", |
| 3085 | "test/subtract-nd.cc", |
| 3086 | ], |
| 3087 | deps = OPERATOR_TEST_DEPS, |
| 3088 | ) |
| 3089 | |
| 3090 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3091 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3092 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3093 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3094 | "test/unpooling-operator-tester.h", |
| 3095 | ], |
| 3096 | deps = OPERATOR_TEST_DEPS, |
| 3097 | ) |
| 3098 | |
| 3099 | ############################# Build configurations ############################# |
| 3100 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3101 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3102 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3103 | name = "xnn_enable_assembly_explicit_true", |
| 3104 | define_values = {"xnn_enable_assembly": "true"}, |
| 3105 | ) |
| 3106 | |
| 3107 | # Disables usage of assembly kernels. |
| 3108 | config_setting( |
| 3109 | name = "xnn_enable_assembly_explicit_false", |
| 3110 | define_values = {"xnn_enable_assembly": "false"}, |
| 3111 | ) |
| 3112 | |
| 3113 | # Builds with -c dbg |
| 3114 | config_setting( |
| 3115 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3116 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3117 | "compilation_mode": "dbg", |
| 3118 | }, |
| 3119 | ) |
| 3120 | |
| 3121 | # Builds with -c opt |
| 3122 | config_setting( |
| 3123 | name = "optimized_build", |
| 3124 | values = { |
| 3125 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3126 | }, |
| 3127 | ) |
| 3128 | |
| 3129 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3130 | name = "linux_k8", |
| 3131 | values = {"cpu": "k8"}, |
| 3132 | ) |
| 3133 | |
| 3134 | config_setting( |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 3135 | name = "linux_aarch64", |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3136 | values = {"cpu": "aarch64"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 3137 | ) |
| 3138 | |
| 3139 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3140 | name = "android", |
| 3141 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 3142 | ) |
| 3143 | |
| 3144 | config_setting( |
| 3145 | name = "android_armv7", |
| 3146 | values = { |
| 3147 | "crosstool_top": "//external:android/crosstool", |
| 3148 | "cpu": "armeabi-v7a", |
| 3149 | }, |
| 3150 | ) |
| 3151 | |
| 3152 | config_setting( |
| 3153 | name = "android_arm64", |
| 3154 | values = { |
| 3155 | "crosstool_top": "//external:android/crosstool", |
| 3156 | "cpu": "arm64-v8a", |
| 3157 | }, |
| 3158 | ) |
| 3159 | |
| 3160 | config_setting( |
| 3161 | name = "android_x86", |
| 3162 | values = { |
| 3163 | "crosstool_top": "//external:android/crosstool", |
| 3164 | "cpu": "x86", |
| 3165 | }, |
| 3166 | ) |
| 3167 | |
| 3168 | config_setting( |
| 3169 | name = "android_x86_64", |
| 3170 | values = { |
| 3171 | "crosstool_top": "//external:android/crosstool", |
| 3172 | "cpu": "x86_64", |
| 3173 | }, |
| 3174 | ) |
| 3175 | |
| 3176 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 3177 | name = "macos_x86_64", |
| 3178 | values = { |
| 3179 | "apple_platform_type": "macos", |
| 3180 | "cpu": "darwin", |
| 3181 | }, |
| 3182 | ) |
| 3183 | |
| 3184 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3185 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3186 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3187 | ) |
| 3188 | |
| 3189 | config_setting( |
| 3190 | name = "emscripten_wasm", |
| 3191 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3192 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3193 | "cpu": "wasm", |
| 3194 | }, |
| 3195 | ) |
| 3196 | |
| 3197 | config_setting( |
| 3198 | name = "emscripten_wasmsimd", |
| 3199 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3200 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3201 | "cpu": "wasm", |
Marat Dukhan | 8c19e3c | 2019-10-30 12:14:58 -0700 | [diff] [blame] | 3202 | "features": "wasm_simd", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3203 | }, |
| 3204 | ) |
| 3205 | |
| 3206 | config_setting( |
| 3207 | name = "emscripten_asmjs", |
| 3208 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3209 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3210 | "cpu": "asmjs", |
| 3211 | }, |
| 3212 | ) |