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 | 8d3c693 | 2020-03-06 20:27:27 -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_dnnl_copts", "xnnpack_optional_dnnl_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 | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 85 | "src/f32-avgpool/9p8x-scalar-c1.c", |
| 86 | "src/f32-avgpool/9x-scalar-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 87 | "src/f32-clamp/gen/scalar-x1.c", |
| 88 | "src/f32-clamp/gen/scalar-x2.c", |
| 89 | "src/f32-clamp/gen/scalar-x4.c", |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 90 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
Marat Dukhan | 6b7dfae | 2019-12-04 16:00:52 -0800 | [diff] [blame] | 91 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 92 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c", |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 93 | "src/f32-dwconv-spchw/3x3p1-scalar.c", |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 94 | "src/f32-dwconv-spchw/3x3s2p1-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 95 | "src/f32-dwconv-spchw/5x5p2-scalar.c", |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 96 | "src/f32-dwconv-spchw/5x5s2p2-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 97 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 98 | "src/f32-dwconv/gen/up1x25-scalar.c", |
| 99 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 100 | "src/f32-dwconv/gen/up1x4-scalar.c", |
| 101 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 102 | "src/f32-dwconv/gen/up1x9-scalar.c", |
| 103 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 104 | "src/f32-dwconv/gen/up2x25-scalar.c", |
| 105 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 106 | "src/f32-dwconv/gen/up2x4-scalar.c", |
| 107 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 108 | "src/f32-dwconv/gen/up2x9-scalar.c", |
Erich Elsen | 34dc2c0 | 2019-10-16 05:11:41 -0700 | [diff] [blame] | 109 | "src/f32-gavgpool-spchw/scalar-x1.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 110 | "src/f32-gavgpool/7p7x-scalar-c1.c", |
| 111 | "src/f32-gavgpool/7x-scalar-c1.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 112 | "src/f32-gemm/gen-inc/1x4-scalar.c", |
| 113 | "src/f32-gemm/gen-inc/2x4-scalar.c", |
| 114 | "src/f32-gemm/gen-inc/4x4-scalar.c", |
| 115 | "src/f32-gemm/gen/1x4-scalar.c", |
| 116 | "src/f32-gemm/gen/2x4-scalar.c", |
| 117 | "src/f32-gemm/gen/4x2-scalar.c", |
| 118 | "src/f32-gemm/gen/4x4-scalar.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 119 | "src/f32-hswish/gen/scalar-x1.c", |
| 120 | "src/f32-hswish/gen/scalar-x2.c", |
| 121 | "src/f32-hswish/gen/scalar-x4.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 122 | "src/f32-ibilinear/gen/scalar-c1.c", |
| 123 | "src/f32-ibilinear/gen/scalar-c2.c", |
| 124 | "src/f32-ibilinear/gen/scalar-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 125 | "src/f32-igemm/gen/1x4-scalar.c", |
| 126 | "src/f32-igemm/gen/2x4-scalar.c", |
| 127 | "src/f32-igemm/gen/4x2-scalar.c", |
| 128 | "src/f32-igemm/gen/4x4-scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 129 | "src/f32-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 130 | "src/f32-pavgpool/9p8x-scalar-c1.c", |
| 131 | "src/f32-pavgpool/9x-scalar-c1.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 132 | "src/f32-ppmm/gen/2x4-scalar.c", |
| 133 | "src/f32-ppmm/gen/3x3-scalar.c", |
| 134 | "src/f32-ppmm/gen/4x2-scalar.c", |
| 135 | "src/f32-ppmm/gen/4x4-scalar.c", |
| 136 | "src/f32-prelu/gen/scalar-2x1.c", |
| 137 | "src/f32-prelu/gen/scalar-2x4.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 138 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
| 139 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
| 140 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
| 141 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
| 142 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 143 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
| 144 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
| 145 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
| 146 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
| 147 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
| 148 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 149 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 150 | "src/f32-rmax/scalar.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 151 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 152 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 153 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
| 154 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 155 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 156 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
| 157 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 158 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 159 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 160 | "src/f32-spmm/gen/1x1-scalar-pipelined.c", |
| 161 | "src/f32-spmm/gen/1x1-scalar.c", |
| 162 | "src/f32-spmm/gen/2x1-scalar-pipelined.c", |
| 163 | "src/f32-spmm/gen/2x1-scalar.c", |
| 164 | "src/f32-spmm/gen/4x1-scalar-pipelined.c", |
| 165 | "src/f32-spmm/gen/4x1-scalar.c", |
| 166 | "src/f32-spmm/gen/8x1-scalar-pipelined.c", |
| 167 | "src/f32-spmm/gen/8x1-scalar.c", |
| 168 | "src/f32-spmm/gen/8x2-scalar.c", |
| 169 | "src/f32-spmm/gen/8x4-scalar.c", |
| 170 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 171 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 172 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
| 173 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 174 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 175 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 176 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 177 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 178 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
| 179 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 180 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 181 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 182 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 183 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 184 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| 185 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 186 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 187 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| 188 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 189 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 190 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| 191 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 192 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 193 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 194 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 195 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 196 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
| 197 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 198 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 199 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 200 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 201 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 202 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 203 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 204 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 205 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
| 206 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 207 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 208 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
| 209 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 210 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 211 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
| 212 | "src/f32-vmulcaddc/gen/c1-scalar-2x.c", |
| 213 | "src/f32-vmulcaddc/gen/c2-scalar-2x.c", |
| 214 | "src/f32-vmulcaddc/gen/c4-scalar-2x.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 215 | "src/math/expminus-scalar-lut2048-p1.c", |
| 216 | "src/math/expminus-scalar-lut64-p2.c", |
| 217 | "src/math/expminus-scalar-p5.c", |
| 218 | "src/math/sigmoid-scalar-lut2048-p1-div.c", |
| 219 | "src/math/sigmoid-scalar-lut64-p2-div.c", |
| 220 | "src/math/sigmoid-scalar-p5-div.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 221 | "src/q8-avgpool/9p8x-scalar-c1.c", |
| 222 | "src/q8-avgpool/9x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 223 | "src/q8-dwconv/up1x9-scalar.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 224 | "src/q8-gavgpool/7p7x-scalar-c1.c", |
| 225 | "src/q8-gavgpool/7x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 226 | "src/q8-gemm/2x2-scalar.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 227 | "src/q8-igemm/2x2-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 228 | "src/q8-vadd/scalar.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 229 | "src/u8-clamp/scalar-x4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 230 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 231 | "src/u8-maxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 232 | "src/u8-rmax/scalar.c", |
| 233 | "src/x32-packx/x2-scalar.c", |
| 234 | "src/x32-packx/x3-scalar.c", |
| 235 | "src/x32-packx/x4-scalar.c", |
| 236 | "src/x32-pad/x2-scalar.c", |
| 237 | "src/x32-unpool/scalar.c", |
| 238 | "src/x32-zip/x2-scalar.c", |
| 239 | "src/x32-zip/x3-scalar.c", |
| 240 | "src/x32-zip/x4-scalar.c", |
| 241 | "src/x32-zip/xm-scalar.c", |
| 242 | "src/x8-lut/scalar.c", |
| 243 | "src/x8-zip/x2-scalar.c", |
| 244 | "src/x8-zip/x3-scalar.c", |
| 245 | "src/x8-zip/x4-scalar.c", |
| 246 | "src/x8-zip/xm-scalar.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 247 | "src/requantization/precise-scalar.c", |
| 248 | "src/requantization/fp32-scalar.c", |
| 249 | "src/requantization/q31-scalar.c", |
| 250 | "src/requantization/gemmlowp-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 251 | ] |
| 252 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 253 | WASM_UKERNELS = [ |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 254 | "src/f32-avgpool/9p8x-wasm-c1.c", |
| 255 | "src/f32-avgpool/9x-wasm-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 256 | "src/f32-clamp/gen/wasm-x1.c", |
| 257 | "src/f32-clamp/gen/wasm-x2.c", |
| 258 | "src/f32-clamp/gen/wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 259 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 260 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 261 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 262 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 263 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 264 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 265 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 266 | "src/f32-dwconv/gen/up2x25-wasm.c", |
| 267 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 268 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 269 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 270 | "src/f32-dwconv/gen/up2x9-wasm.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 271 | "src/f32-gavgpool/7p7x-wasm-c1.c", |
| 272 | "src/f32-gavgpool/7x-wasm-c1.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 273 | "src/f32-gemm/gen-inc/1x4-wasm.c", |
| 274 | "src/f32-gemm/gen-inc/2x4-wasm.c", |
| 275 | "src/f32-gemm/gen-inc/4x4-wasm.c", |
| 276 | "src/f32-gemm/gen/1x4-wasm.c", |
| 277 | "src/f32-gemm/gen/2x4-wasm.c", |
| 278 | "src/f32-gemm/gen/4x2-wasm.c", |
| 279 | "src/f32-gemm/gen/4x4-wasm.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 280 | "src/f32-hswish/gen/wasm-x1.c", |
| 281 | "src/f32-hswish/gen/wasm-x2.c", |
| 282 | "src/f32-hswish/gen/wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 283 | "src/f32-igemm/gen/1x4-wasm.c", |
| 284 | "src/f32-igemm/gen/2x4-wasm.c", |
| 285 | "src/f32-igemm/gen/4x2-wasm.c", |
| 286 | "src/f32-igemm/gen/4x4-wasm.c", |
| 287 | "src/f32-maxpool/9p8x-wasm-c1.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 288 | "src/f32-pavgpool/9p8x-wasm-c1.c", |
| 289 | "src/f32-pavgpool/9x-wasm-c1.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 290 | "src/f32-vbinary/gen/vadd-wasm-x1.c", |
| 291 | "src/f32-vbinary/gen/vadd-wasm-x2.c", |
| 292 | "src/f32-vbinary/gen/vadd-wasm-x4.c", |
| 293 | "src/f32-vbinary/gen/vaddc-wasm-x1.c", |
| 294 | "src/f32-vbinary/gen/vaddc-wasm-x2.c", |
| 295 | "src/f32-vbinary/gen/vaddc-wasm-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 296 | "src/f32-vbinary/gen/vdiv-wasm-x1.c", |
| 297 | "src/f32-vbinary/gen/vdiv-wasm-x2.c", |
| 298 | "src/f32-vbinary/gen/vdiv-wasm-x4.c", |
| 299 | "src/f32-vbinary/gen/vdivc-wasm-x1.c", |
| 300 | "src/f32-vbinary/gen/vdivc-wasm-x2.c", |
| 301 | "src/f32-vbinary/gen/vdivc-wasm-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 302 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 303 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 304 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
| 305 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 306 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 307 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
| 308 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 309 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 310 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
| 311 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 312 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 313 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 314 | "src/f32-vbinary/gen/vmul-wasm-x1.c", |
| 315 | "src/f32-vbinary/gen/vmul-wasm-x2.c", |
| 316 | "src/f32-vbinary/gen/vmul-wasm-x4.c", |
| 317 | "src/f32-vbinary/gen/vmulc-wasm-x1.c", |
| 318 | "src/f32-vbinary/gen/vmulc-wasm-x2.c", |
| 319 | "src/f32-vbinary/gen/vmulc-wasm-x4.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 320 | "src/f32-vbinary/gen/vrdivc-wasm-x1.c", |
| 321 | "src/f32-vbinary/gen/vrdivc-wasm-x2.c", |
| 322 | "src/f32-vbinary/gen/vrdivc-wasm-x4.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 323 | "src/f32-vbinary/gen/vrsubc-wasm-x1.c", |
| 324 | "src/f32-vbinary/gen/vrsubc-wasm-x2.c", |
| 325 | "src/f32-vbinary/gen/vrsubc-wasm-x4.c", |
| 326 | "src/f32-vbinary/gen/vsub-wasm-x1.c", |
| 327 | "src/f32-vbinary/gen/vsub-wasm-x2.c", |
| 328 | "src/f32-vbinary/gen/vsub-wasm-x4.c", |
| 329 | "src/f32-vbinary/gen/vsubc-wasm-x1.c", |
| 330 | "src/f32-vbinary/gen/vsubc-wasm-x2.c", |
| 331 | "src/f32-vbinary/gen/vsubc-wasm-x4.c", |
| 332 | "src/f32-vmulcaddc/gen/c1-wasm-2x.c", |
| 333 | "src/f32-vmulcaddc/gen/c2-wasm-2x.c", |
| 334 | "src/f32-vmulcaddc/gen/c4-wasm-2x.c", |
| 335 | ] |
| 336 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 337 | PSIMD_FASTMATH_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 338 | "src/f32-argmaxpool/4x-psimd-c4.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 339 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 340 | "src/f32-argmaxpool/9x-psimd-c4.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 341 | "src/f32-avgpool/9p8x-psimd-c4.c", |
| 342 | "src/f32-avgpool/9x-psimd-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 343 | "src/f32-clamp/gen/psimd-x4.c", |
| 344 | "src/f32-clamp/gen/psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 345 | "src/f32-dwconv/gen/up4x25-psimd-acc2.c", |
| 346 | "src/f32-dwconv/gen/up4x25-psimd.c", |
| 347 | "src/f32-dwconv/gen/up4x4-psimd-acc2.c", |
| 348 | "src/f32-dwconv/gen/up4x4-psimd.c", |
| 349 | "src/f32-dwconv/gen/up4x9-psimd-acc2.c", |
| 350 | "src/f32-dwconv/gen/up4x9-psimd.c", |
| 351 | "src/f32-dwconv/gen/up8x25-psimd-acc2.c", |
| 352 | "src/f32-dwconv/gen/up8x25-psimd.c", |
| 353 | "src/f32-dwconv/gen/up8x4-psimd-acc2.c", |
| 354 | "src/f32-dwconv/gen/up8x4-psimd.c", |
| 355 | "src/f32-dwconv/gen/up8x9-psimd-acc2.c", |
| 356 | "src/f32-dwconv/gen/up8x9-psimd.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 357 | "src/f32-gavgpool/7p7x-psimd-c4.c", |
| 358 | "src/f32-gavgpool/7x-psimd-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 359 | "src/f32-gemm/gen/1x8-psimd-loadsplat.c", |
| 360 | "src/f32-gemm/gen/1x8-psimd-splat.c", |
| 361 | "src/f32-gemm/gen/1x8s4-psimd.c", |
Marat Dukhan | b00004d | 2020-02-13 09:28:35 -0800 | [diff] [blame] | 362 | "src/f32-gemm/gen/4x2c4-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 363 | "src/f32-gemm/gen/4x8-psimd-loadsplat.c", |
| 364 | "src/f32-gemm/gen/4x8-psimd-splat.c", |
| 365 | "src/f32-gemm/gen/4x8s4-psimd.c", |
| 366 | "src/f32-gemm/gen/6x8-psimd-loadsplat.c", |
| 367 | "src/f32-gemm/gen/6x8-psimd-splat.c", |
| 368 | "src/f32-gemm/gen/6x8s4-psimd.c", |
| 369 | "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c", |
| 370 | "src/f32-gemm/gen-inc/1x8-psimd-splat.c", |
| 371 | "src/f32-gemm/gen-inc/1x8s4-psimd.c", |
| 372 | "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c", |
| 373 | "src/f32-gemm/gen-inc/4x8-psimd-splat.c", |
| 374 | "src/f32-gemm/gen-inc/4x8s4-psimd.c", |
| 375 | "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c", |
| 376 | "src/f32-gemm/gen-inc/6x8-psimd-splat.c", |
| 377 | "src/f32-gemm/gen-inc/6x8s4-psimd.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 378 | "src/f32-hswish/gen/psimd-x4.c", |
| 379 | "src/f32-hswish/gen/psimd-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 380 | "src/f32-ibilinear/gen/psimd-c4.c", |
| 381 | "src/f32-ibilinear/gen/psimd-c8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 382 | "src/f32-igemm/gen/1x8-psimd-loadsplat.c", |
| 383 | "src/f32-igemm/gen/1x8-psimd-splat.c", |
| 384 | "src/f32-igemm/gen/1x8s4-psimd.c", |
| 385 | "src/f32-igemm/gen/4x2c4-psimd.c", |
| 386 | "src/f32-igemm/gen/4x8-psimd-loadsplat.c", |
| 387 | "src/f32-igemm/gen/4x8-psimd-splat.c", |
| 388 | "src/f32-igemm/gen/4x8s4-psimd.c", |
| 389 | "src/f32-igemm/gen/6x8-psimd-loadsplat.c", |
| 390 | "src/f32-igemm/gen/6x8-psimd-splat.c", |
| 391 | "src/f32-igemm/gen/6x8s4-psimd.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 392 | "src/f32-maxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 393 | "src/f32-pavgpool/9p8x-psimd-c4.c", |
| 394 | "src/f32-pavgpool/9x-psimd-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 395 | "src/f32-ppmm/gen/4x8-psimd.c", |
| 396 | "src/f32-prelu/gen/psimd-2x4.c", |
| 397 | "src/f32-prelu/gen/psimd-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 398 | "src/f32-rmax/psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 399 | "src/f32-vbinary/gen/vadd-psimd-x4.c", |
| 400 | "src/f32-vbinary/gen/vadd-psimd-x8.c", |
| 401 | "src/f32-vbinary/gen/vaddc-psimd-x4.c", |
| 402 | "src/f32-vbinary/gen/vaddc-psimd-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 403 | "src/f32-vbinary/gen/vdiv-psimd-x4.c", |
| 404 | "src/f32-vbinary/gen/vdiv-psimd-x8.c", |
| 405 | "src/f32-vbinary/gen/vdivc-psimd-x4.c", |
| 406 | "src/f32-vbinary/gen/vdivc-psimd-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 407 | "src/f32-vbinary/gen/vmax-psimd-x4.c", |
| 408 | "src/f32-vbinary/gen/vmax-psimd-x8.c", |
| 409 | "src/f32-vbinary/gen/vmaxc-psimd-x4.c", |
| 410 | "src/f32-vbinary/gen/vmaxc-psimd-x8.c", |
| 411 | "src/f32-vbinary/gen/vmin-psimd-x4.c", |
| 412 | "src/f32-vbinary/gen/vmin-psimd-x8.c", |
| 413 | "src/f32-vbinary/gen/vminc-psimd-x4.c", |
| 414 | "src/f32-vbinary/gen/vminc-psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 415 | "src/f32-vbinary/gen/vmul-psimd-x4.c", |
| 416 | "src/f32-vbinary/gen/vmul-psimd-x8.c", |
| 417 | "src/f32-vbinary/gen/vmulc-psimd-x4.c", |
| 418 | "src/f32-vbinary/gen/vmulc-psimd-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 419 | "src/f32-vbinary/gen/vrdivc-psimd-x4.c", |
| 420 | "src/f32-vbinary/gen/vrdivc-psimd-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 421 | "src/f32-vbinary/gen/vrsubc-psimd-x4.c", |
| 422 | "src/f32-vbinary/gen/vrsubc-psimd-x8.c", |
| 423 | "src/f32-vbinary/gen/vsub-psimd-x4.c", |
| 424 | "src/f32-vbinary/gen/vsub-psimd-x8.c", |
| 425 | "src/f32-vbinary/gen/vsubc-psimd-x4.c", |
| 426 | "src/f32-vbinary/gen/vsubc-psimd-x8.c", |
| 427 | "src/f32-vmulcaddc/gen/c4-psimd-2x.c", |
| 428 | "src/f32-vmulcaddc/gen/c8-psimd-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 429 | "src/x32-packx/x4-psimd.c", |
| 430 | "src/x32-pad/x2-psimd.c", |
| 431 | "src/x32-unpool/psimd.c", |
| 432 | "src/x32-zip/x2-psimd.c", |
| 433 | "src/x32-zip/x3-psimd.c", |
| 434 | "src/x32-zip/x4-psimd.c", |
| 435 | "src/x32-zip/xm-psimd.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 436 | "src/requantization/precise-psimd.c", |
| 437 | "src/requantization/fp32-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 438 | ] |
| 439 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 440 | PSIMD_ACCMATH_UKERNELS = [ |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 441 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c", |
| 442 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c", |
| 443 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c", |
| 444 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c", |
| 445 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c", |
| 446 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c", |
| 447 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c", |
| 448 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c", |
| 449 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c", |
| 450 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c", |
| 451 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c", |
| 452 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 453 | "src/f32-sigmoid/gen/psimd-p5-div-x4.c", |
| 454 | "src/f32-sigmoid/gen/psimd-p5-div-x8.c", |
| 455 | "src/f32-sigmoid/gen/psimd-p5-div-x12.c", |
| 456 | "src/f32-sigmoid/gen/psimd-p5-div-x16.c", |
| 457 | "src/f32-sigmoid/gen/psimd-p5-div-x20.c", |
| 458 | "src/f32-sigmoid/gen/psimd-p5-div-x24.c", |
| 459 | "src/math/sigmoid-psimd-p5-div.c", |
| 460 | ] |
| 461 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 462 | # ISA-specific micro-kernels |
| 463 | NEON_UKERNELS = [ |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 464 | "src/f32-avgpool/9p8x-neon-c4.c", |
| 465 | "src/f32-avgpool/9x-neon-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 466 | "src/f32-clamp/gen/neon-x4.c", |
| 467 | "src/f32-clamp/gen/neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 468 | "src/f32-dwconv/gen/up4x9-neon.c", |
| 469 | "src/f32-dwconv/gen/up4x9-neon-acc2.c", |
| 470 | "src/f32-dwconv/gen/up8x9-neon.c", |
| 471 | "src/f32-dwconv/gen/up8x9-neon-acc2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 472 | "src/f32-gavgpool-spchw/neon-x4.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 473 | "src/f32-gavgpool/7p7x-neon-c4.c", |
| 474 | "src/f32-gavgpool/7x-neon-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 475 | "src/f32-gemm/gen/1x8-neon-lane-ld64.c", |
| 476 | "src/f32-gemm/gen/4x2-neon-lane-ld64.c", |
| 477 | "src/f32-gemm/gen/4x8-neon-lane-ld128.c", |
| 478 | "src/f32-gemm/gen/4x8-neon-lane-ld64.c", |
| 479 | "src/f32-gemm/gen/5x8-neon-lane-ld64.c", |
| 480 | "src/f32-gemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 481 | "src/f32-gemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 482 | "src/f32-gemm/gen/1x8-neon-dup-ld64.c", |
| 483 | "src/f32-gemm/gen/4x8-neon-dup-ld128.c", |
| 484 | "src/f32-gemm/gen/4x8-neon-dup-ld64.c", |
| 485 | "src/f32-gemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 486 | "src/f32-gemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 487 | "src/f32-gemm/gen/1x8s4-neon.c", |
| 488 | "src/f32-gemm/gen/4x8s4-neon.c", |
| 489 | "src/f32-gemm/gen/6x8s4-neon.c", |
| 490 | "src/f32-gemm/gen/8x8s4-neon.c", |
| 491 | "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c", |
| 492 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c", |
| 493 | "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c", |
| 494 | "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c", |
| 495 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 496 | "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 497 | "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c", |
| 498 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c", |
| 499 | "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c", |
| 500 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 501 | "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 502 | "src/f32-gemm/gen-inc/1x8s4-neon.c", |
| 503 | "src/f32-gemm/gen-inc/4x8s4-neon.c", |
| 504 | "src/f32-gemm/gen-inc/6x8s4-neon.c", |
| 505 | "src/f32-gemm/gen-inc/8x8s4-neon.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 506 | "src/f32-hswish/gen/neon-x4.c", |
| 507 | "src/f32-hswish/gen/neon-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 508 | "src/f32-ibilinear/gen/neon-c4.c", |
| 509 | "src/f32-ibilinear/gen/neon-c8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 510 | "src/f32-igemm/gen/1x8-neon-lane-ld64.c", |
| 511 | "src/f32-igemm/gen/4x2-neon-lane-ld64.c", |
| 512 | "src/f32-igemm/gen/4x4-neon-lane-ld64.c", |
| 513 | "src/f32-igemm/gen/4x8-neon-lane-ld128.c", |
| 514 | "src/f32-igemm/gen/4x8-neon-lane-ld64.c", |
| 515 | "src/f32-igemm/gen/6x8-neon-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 516 | "src/f32-igemm/gen/6x8-neon-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 517 | "src/f32-igemm/gen/1x8-neon-dup-ld64.c", |
| 518 | "src/f32-igemm/gen/4x8-neon-dup-ld128.c", |
| 519 | "src/f32-igemm/gen/4x8-neon-dup-ld64.c", |
| 520 | "src/f32-igemm/gen/6x8-neon-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 521 | "src/f32-igemm/gen/6x8-neon-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 522 | "src/f32-igemm/gen/1x8s4-neon.c", |
| 523 | "src/f32-igemm/gen/4x8s4-neon.c", |
| 524 | "src/f32-igemm/gen/6x8s4-neon.c", |
| 525 | "src/f32-igemm/gen/8x8s4-neon.c", |
Frank Barchard | f092a4a | 2020-03-03 14:22:46 -0800 | [diff] [blame] | 526 | "src/f32-maxpool/9p8x-neon-c4.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 527 | "src/f32-pavgpool/9p8x-neon-c4.c", |
| 528 | "src/f32-pavgpool/9x-neon-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 529 | "src/f32-ppmm/gen/4x8-neon.c", |
| 530 | "src/f32-ppmm/gen/8x8-neon.c", |
| 531 | "src/f32-prelu/gen/neon-2x4.c", |
| 532 | "src/f32-prelu/gen/neon-2x8.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 533 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 534 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 535 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 536 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 537 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 538 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 539 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 540 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 541 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 542 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| 543 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 544 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 545 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| 546 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| 547 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| 548 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| 549 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 550 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| 551 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| 552 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 553 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| 554 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 555 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 556 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 557 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 558 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 559 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 560 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 561 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 562 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 563 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 564 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| 565 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 566 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 567 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 568 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 569 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 570 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 571 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 572 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 573 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 574 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 575 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 576 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 577 | "src/f32-vbinary/gen/vadd-neon-x4.c", |
| 578 | "src/f32-vbinary/gen/vadd-neon-x8.c", |
| 579 | "src/f32-vbinary/gen/vaddc-neon-x4.c", |
| 580 | "src/f32-vbinary/gen/vaddc-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 581 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 582 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 583 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 584 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 585 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 586 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 587 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 588 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 589 | "src/f32-vbinary/gen/vmul-neon-x4.c", |
| 590 | "src/f32-vbinary/gen/vmul-neon-x8.c", |
| 591 | "src/f32-vbinary/gen/vmulc-neon-x4.c", |
| 592 | "src/f32-vbinary/gen/vmulc-neon-x8.c", |
| 593 | "src/f32-vbinary/gen/vrsubc-neon-x4.c", |
| 594 | "src/f32-vbinary/gen/vrsubc-neon-x8.c", |
| 595 | "src/f32-vbinary/gen/vsub-neon-x4.c", |
| 596 | "src/f32-vbinary/gen/vsub-neon-x8.c", |
| 597 | "src/f32-vbinary/gen/vsubc-neon-x4.c", |
| 598 | "src/f32-vbinary/gen/vsubc-neon-x8.c", |
| 599 | "src/f32-vmulcaddc/gen/c4-neon-2x.c", |
| 600 | "src/f32-vmulcaddc/gen/c8-neon-2x.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 601 | "src/q8-avgpool/9p8x-neon-c8.c", |
| 602 | "src/q8-avgpool/9x-neon-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 603 | "src/q8-dwconv/up8x9-neon.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 604 | "src/q8-gavgpool/7p7x-neon-c8.c", |
| 605 | "src/q8-gavgpool/7x-neon-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 606 | "src/q8-gemm/4x8-neon.c", |
| 607 | "src/q8-gemm/8x8-neon.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 608 | "src/q8-igemm/4x8-neon.c", |
| 609 | "src/q8-igemm/8x8-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 610 | "src/q8-vadd/neon.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 611 | "src/u8-clamp/neon-x64.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 612 | "src/u8-maxpool/9p8x-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 613 | "src/u8-rmax/neon.c", |
| 614 | "src/x32-packx/x4-neon-st4.c", |
| 615 | "src/x32-pad/x2-neon.c", |
| 616 | "src/x32-zip/x2-neon.c", |
| 617 | "src/x32-zip/x3-neon.c", |
| 618 | "src/x32-zip/x4-neon.c", |
| 619 | "src/x32-zip/xm-neon.c", |
| 620 | "src/x8-zip/x2-neon.c", |
| 621 | "src/x8-zip/x3-neon.c", |
| 622 | "src/x8-zip/x4-neon.c", |
| 623 | "src/x8-zip/xm-neon.c", |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 624 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 625 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 626 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 627 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 628 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 629 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 630 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 631 | "src/requantization/precise-neon.c", |
| 632 | "src/requantization/fp32-neon.c", |
| 633 | "src/requantization/q31-neon.c", |
| 634 | "src/requantization/gemmlowp-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 635 | ] |
| 636 | |
| 637 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 638 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 639 | "src/f32-ibilinear/gen/neonfma-c8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 640 | "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c", |
| 641 | "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c", |
| 642 | "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c", |
| 643 | "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 644 | "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 645 | "src/f32-igemm/gen/1x8s4-neonfma.c", |
| 646 | "src/f32-igemm/gen/4x8s4-neonfma.c", |
| 647 | "src/f32-igemm/gen/6x8s4-neonfma.c", |
| 648 | "src/f32-igemm/gen/8x8s4-neonfma.c", |
| 649 | "src/f32-dwconv/gen/up4x9-neonfma.c", |
| 650 | "src/f32-dwconv/gen/up4x9-neonfma-acc2.c", |
| 651 | "src/f32-dwconv/gen/up8x9-neonfma.c", |
| 652 | "src/f32-dwconv/gen/up8x9-neonfma-acc2.c", |
| 653 | "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c", |
| 654 | "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c", |
| 655 | "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c", |
| 656 | "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 657 | "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 658 | "src/f32-gemm/gen/1x8s4-neonfma.c", |
| 659 | "src/f32-gemm/gen/4x8s4-neonfma.c", |
| 660 | "src/f32-gemm/gen/6x8s4-neonfma.c", |
| 661 | "src/f32-gemm/gen/8x8s4-neonfma.c", |
| 662 | "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c", |
| 663 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c", |
| 664 | "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c", |
| 665 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 666 | "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 667 | "src/f32-gemm/gen-inc/1x8s4-neonfma.c", |
| 668 | "src/f32-gemm/gen-inc/4x8s4-neonfma.c", |
| 669 | "src/f32-gemm/gen-inc/6x8s4-neonfma.c", |
| 670 | "src/f32-gemm/gen-inc/8x8s4-neonfma.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 671 | "src/f32-hswish/gen/neonfma-x4.c", |
| 672 | "src/f32-hswish/gen/neonfma-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 673 | "src/f32-ppmm/gen/4x8-neonfma.c", |
| 674 | "src/f32-ppmm/gen/8x8-neonfma.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 675 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 676 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 677 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 678 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 679 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 680 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 681 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 682 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 683 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 684 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| 685 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 686 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 687 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| 688 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| 689 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| 690 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| 691 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 692 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| 693 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| 694 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 695 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| 696 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 697 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 698 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 699 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 700 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 701 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 702 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 703 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 704 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 705 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 706 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 707 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 708 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 709 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 710 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 711 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 712 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 713 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 714 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 715 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 716 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| 717 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 718 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 719 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 720 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 721 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 722 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| 723 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 724 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 725 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 726 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 727 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 728 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| 729 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 730 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 731 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 732 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 733 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 734 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| 735 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 736 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 737 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 738 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 739 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 740 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| 741 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 742 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 743 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 744 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 745 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 746 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| 747 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 748 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 749 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 750 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 751 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 752 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 753 | "src/f32-vmulcaddc/gen/c4-neonfma-2x.c", |
| 754 | "src/f32-vmulcaddc/gen/c8-neonfma-2x.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 755 | "src/math/exp-neonfma-lut64-p2.c", |
| 756 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 757 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 758 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 759 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 760 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 761 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 762 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| 763 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 764 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 765 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| 766 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 767 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 768 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| 769 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 770 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 771 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| 772 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 773 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 774 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| 775 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 776 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 777 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 778 | ] |
| 779 | |
| 780 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 781 | "src/f32-vbinary/gen/vdiv-neon-x4.c", |
| 782 | "src/f32-vbinary/gen/vdiv-neon-x8.c", |
| 783 | "src/f32-vbinary/gen/vdivc-neon-x4.c", |
| 784 | "src/f32-vbinary/gen/vdivc-neon-x8.c", |
| 785 | "src/f32-vbinary/gen/vrdivc-neon-x4.c", |
| 786 | "src/f32-vbinary/gen/vrdivc-neon-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 787 | "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c", |
| 788 | "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c", |
| 789 | "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c", |
| 790 | "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c", |
| 791 | "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c", |
| 792 | "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 793 | "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 794 | "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c", |
| 795 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c", |
| 796 | "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c", |
| 797 | "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c", |
| 798 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 799 | "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 800 | "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c", |
| 801 | "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c", |
| 802 | "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c", |
| 803 | "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c", |
| 804 | "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c", |
| 805 | "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c", |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 806 | "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 807 | "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c", |
| 808 | "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c", |
| 809 | "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c", |
| 810 | "src/f32-dwconv-spchw/3x3p1-neonfma.c", |
| 811 | "src/f32-dwconv-spchw/5x5p2-neonfma.c", |
| 812 | "src/f32-dwconv-spchw/3x3s2p1-neonfma.c", |
| 813 | "src/f32-dwconv-spchw/5x5s2p2-neonfma.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 814 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 815 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 816 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 817 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 818 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 819 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| 820 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 821 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 822 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 823 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 824 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 825 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 826 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 827 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 828 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 829 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 830 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 831 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 832 | "src/f32-spmm/gen/12x1-neonfma.c", |
| 833 | "src/f32-spmm/gen/12x2-neonfma.c", |
| 834 | "src/f32-spmm/gen/12x4-neonfma.c", |
| 835 | "src/f32-spmm/gen/16x1-neonfma-pipelined.c", |
| 836 | "src/f32-spmm/gen/16x1-neonfma-unroll2.c", |
| 837 | "src/f32-spmm/gen/16x1-neonfma.c", |
| 838 | "src/f32-spmm/gen/16x2-neonfma.c", |
| 839 | "src/f32-spmm/gen/16x4-neonfma.c", |
| 840 | "src/f32-spmm/gen/4x1-neonfma-pipelined.c", |
| 841 | "src/f32-spmm/gen/4x1-neonfma-unroll2.c", |
| 842 | "src/f32-spmm/gen/4x1-neonfma.c", |
| 843 | "src/f32-spmm/gen/4x2-neonfma.c", |
| 844 | "src/f32-spmm/gen/4x4-neonfma.c", |
| 845 | "src/f32-spmm/gen/8x1-neonfma-pipelined.c", |
| 846 | "src/f32-spmm/gen/8x1-neonfma-unroll2.c", |
| 847 | "src/f32-spmm/gen/8x1-neonfma.c", |
| 848 | "src/f32-spmm/gen/8x2-neonfma.c", |
| 849 | "src/f32-spmm/gen/8x4-neonfma.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 850 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| 851 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| 852 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| 853 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| 854 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| 855 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 856 | ] |
| 857 | |
| 858 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 859 | "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c", |
| 860 | "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c", |
| 861 | "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 862 | "src/f16-spmm/gen/8x1-neonfp16arith.c", |
| 863 | "src/f16-spmm/gen/8x1-neonfp16arith-unroll2.c", |
| 864 | "src/f16-spmm/gen/16x1-neonfp16arith.c", |
| 865 | "src/f16-spmm/gen/16x1-neonfp16arith-unroll2.c", |
| 866 | "src/f16-spmm/gen/24x1-neonfp16arith.c", |
| 867 | "src/f16-spmm/gen/24x1-neonfp16arith-unroll2.c", |
| 868 | "src/f16-spmm/gen/32x1-neonfp16arith.c", |
| 869 | "src/f16-spmm/gen/32x1-neonfp16arith-unroll2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 870 | ] |
| 871 | |
| 872 | SSE_UKERNELS = [ |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 873 | "src/f32-avgpool/9p8x-sse-c4.c", |
| 874 | "src/f32-avgpool/9x-sse-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 875 | "src/f32-clamp/gen/sse-x4.c", |
| 876 | "src/f32-clamp/gen/sse-x8.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 877 | "src/f32-dwconv-spchw/3x3p1-sse.c", |
| 878 | "src/f32-dwconv-spchw/3x3s2p1-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 879 | "src/f32-dwconv/gen/up4x25-sse-acc2.c", |
| 880 | "src/f32-dwconv/gen/up4x25-sse.c", |
| 881 | "src/f32-dwconv/gen/up4x4-sse-acc2.c", |
| 882 | "src/f32-dwconv/gen/up4x4-sse.c", |
| 883 | "src/f32-dwconv/gen/up4x9-sse-acc2.c", |
| 884 | "src/f32-dwconv/gen/up4x9-sse.c", |
| 885 | "src/f32-dwconv/gen/up8x25-sse-acc2.c", |
| 886 | "src/f32-dwconv/gen/up8x25-sse.c", |
| 887 | "src/f32-dwconv/gen/up8x4-sse-acc2.c", |
| 888 | "src/f32-dwconv/gen/up8x4-sse.c", |
| 889 | "src/f32-dwconv/gen/up8x9-sse-acc2.c", |
| 890 | "src/f32-dwconv/gen/up8x9-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 891 | "src/f32-gavgpool-spchw/sse-x4.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 892 | "src/f32-gavgpool/7p7x-sse-c4.c", |
| 893 | "src/f32-gavgpool/7x-sse-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 894 | "src/f32-gemm/gen/1x8-sse-dup.c", |
| 895 | "src/f32-gemm/gen/1x8-sse-load1.c", |
| 896 | "src/f32-gemm/gen/1x8s4-sse.c", |
Marat Dukhan | b00004d | 2020-02-13 09:28:35 -0800 | [diff] [blame] | 897 | "src/f32-gemm/gen/4x2c4-sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 898 | "src/f32-gemm/gen/4x8-sse-dup.c", |
| 899 | "src/f32-gemm/gen/4x8-sse-load1.c", |
| 900 | "src/f32-gemm/gen/4x8s4-sse.c", |
| 901 | "src/f32-gemm/gen-inc/1x8-sse-dup.c", |
| 902 | "src/f32-gemm/gen-inc/1x8-sse-load1.c", |
| 903 | "src/f32-gemm/gen-inc/1x8s4-sse.c", |
| 904 | "src/f32-gemm/gen-inc/4x8-sse-dup.c", |
| 905 | "src/f32-gemm/gen-inc/4x8-sse-load1.c", |
| 906 | "src/f32-gemm/gen-inc/4x8s4-sse.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 907 | "src/f32-hswish/gen/sse-x4.c", |
| 908 | "src/f32-hswish/gen/sse-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 909 | "src/f32-ibilinear/gen/sse-c4.c", |
| 910 | "src/f32-ibilinear/gen/sse-c8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 911 | "src/f32-igemm/gen/1x8-sse-dup.c", |
| 912 | "src/f32-igemm/gen/1x8-sse-load1.c", |
| 913 | "src/f32-igemm/gen/1x8s4-sse.c", |
| 914 | "src/f32-igemm/gen/4x2c4-sse.c", |
| 915 | "src/f32-igemm/gen/4x8-sse-dup.c", |
| 916 | "src/f32-igemm/gen/4x8-sse-load1.c", |
| 917 | "src/f32-igemm/gen/4x8s4-sse.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 918 | "src/f32-maxpool/9p8x-sse-c4.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 919 | "src/f32-pavgpool/9p8x-sse-c4.c", |
| 920 | "src/f32-pavgpool/9x-sse-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 921 | "src/f32-ppmm/gen/4x8-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 922 | "src/f32-rmax/sse.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 923 | "src/f32-spmm/gen/4x1-sse.c", |
| 924 | "src/f32-spmm/gen/8x1-sse.c", |
| 925 | "src/f32-vbinary/gen/vadd-sse-x4.c", |
| 926 | "src/f32-vbinary/gen/vadd-sse-x8.c", |
| 927 | "src/f32-vbinary/gen/vaddc-sse-x4.c", |
| 928 | "src/f32-vbinary/gen/vaddc-sse-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 929 | "src/f32-vbinary/gen/vdiv-sse-x4.c", |
| 930 | "src/f32-vbinary/gen/vdiv-sse-x8.c", |
| 931 | "src/f32-vbinary/gen/vdivc-sse-x4.c", |
| 932 | "src/f32-vbinary/gen/vdivc-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 933 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 934 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 935 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 936 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 937 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 938 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 939 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 940 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 941 | "src/f32-vbinary/gen/vmul-sse-x4.c", |
| 942 | "src/f32-vbinary/gen/vmul-sse-x8.c", |
| 943 | "src/f32-vbinary/gen/vmulc-sse-x4.c", |
| 944 | "src/f32-vbinary/gen/vmulc-sse-x8.c", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 945 | "src/f32-vbinary/gen/vrdivc-sse-x4.c", |
| 946 | "src/f32-vbinary/gen/vrdivc-sse-x8.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 947 | "src/f32-vbinary/gen/vrsubc-sse-x4.c", |
| 948 | "src/f32-vbinary/gen/vrsubc-sse-x8.c", |
| 949 | "src/f32-vbinary/gen/vsub-sse-x4.c", |
| 950 | "src/f32-vbinary/gen/vsub-sse-x8.c", |
| 951 | "src/f32-vbinary/gen/vsubc-sse-x4.c", |
| 952 | "src/f32-vbinary/gen/vsubc-sse-x8.c", |
| 953 | "src/f32-vmulcaddc/gen/c4-sse-2x.c", |
| 954 | "src/f32-vmulcaddc/gen/c8-sse-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 955 | "src/x32-packx/x4-sse.c", |
| 956 | ] |
| 957 | |
| 958 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 959 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 960 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 961 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 962 | "src/f32-prelu/gen/sse2-2x4.c", |
| 963 | "src/f32-prelu/gen/sse2-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 964 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| 965 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| 966 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| 967 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| 968 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 969 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| 970 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| 971 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 972 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| 973 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 974 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 975 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 976 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 977 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 978 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 979 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 980 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 981 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Marat Dukhan | 6ee435a | 2020-02-26 22:33:38 -0800 | [diff] [blame] | 982 | "src/q8-avgpool/9p8x-sse2-c8.c", |
| 983 | "src/q8-avgpool/9x-sse2-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 984 | "src/q8-igemm/4x4c2-sse2.c", |
| 985 | "src/q8-dwconv/up8x9-sse2.c", |
Marat Dukhan | a63a6fc | 2020-03-10 06:12:48 -0700 | [diff] [blame] | 986 | "src/q8-gavgpool/7p7x-sse2-c8.c", |
| 987 | "src/q8-gavgpool/7x-sse2-c8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 988 | "src/q8-gemm/2x4c8-sse2.c", |
| 989 | "src/q8-gemm/4x4c2-sse2.c", |
| 990 | "src/q8-vadd/sse2.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 991 | "src/u8-clamp/sse2-x64.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 992 | "src/u8-maxpool/9p8x-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 993 | "src/u8-rmax/sse2.c", |
| 994 | "src/x32-pad/x2-sse2.c", |
| 995 | "src/x32-zip/x2-sse2.c", |
| 996 | "src/x32-zip/x3-sse2.c", |
| 997 | "src/x32-zip/x4-sse2.c", |
| 998 | "src/x32-zip/xm-sse2.c", |
| 999 | "src/x8-zip/x2-sse2.c", |
| 1000 | "src/x8-zip/x3-sse2.c", |
| 1001 | "src/x8-zip/x4-sse2.c", |
| 1002 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 1003 | "src/math/exp-sse2-p5.c", |
| 1004 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 1005 | "src/math/sigmoid-sse2-p5-div.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1006 | "src/requantization/precise-sse2.c", |
| 1007 | "src/requantization/fp32-sse2.c", |
| 1008 | "src/requantization/q31-sse2.c", |
| 1009 | "src/requantization/gemmlowp-sse2.c", |
| 1010 | ] |
| 1011 | |
| 1012 | SSSE3_UKERNELS = [ |
| 1013 | "src/requantization/precise-ssse3.c", |
| 1014 | "src/requantization/q31-ssse3.c", |
| 1015 | "src/requantization/gemmlowp-ssse3.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1016 | ] |
| 1017 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1018 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1019 | "src/f32-prelu/gen/sse41-2x4.c", |
| 1020 | "src/f32-prelu/gen/sse41-2x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1021 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1022 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1023 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1024 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1025 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 1026 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1027 | "src/requantization/precise-sse4.c", |
| 1028 | "src/requantization/q31-sse4.c", |
| 1029 | "src/requantization/gemmlowp-sse4.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1030 | ] |
| 1031 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1032 | AVX_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1033 | "src/f32-clamp/gen/avx-x8.c", |
| 1034 | "src/f32-clamp/gen/avx-x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1035 | "src/f32-dwconv/gen/up16x4-avx-acc2.c", |
| 1036 | "src/f32-dwconv/gen/up16x4-avx.c", |
| 1037 | "src/f32-dwconv/gen/up8x4-avx-acc2.c", |
| 1038 | "src/f32-dwconv/gen/up8x4-avx.c", |
| 1039 | "src/f32-dwconv/gen/up16x9-avx-acc2.c", |
| 1040 | "src/f32-dwconv/gen/up16x9-avx.c", |
| 1041 | "src/f32-dwconv/gen/up8x9-avx-acc2.c", |
| 1042 | "src/f32-dwconv/gen/up8x9-avx.c", |
| 1043 | "src/f32-dwconv/gen/up16x25-avx-acc2.c", |
| 1044 | "src/f32-dwconv/gen/up16x25-avx.c", |
| 1045 | "src/f32-dwconv/gen/up8x25-avx-acc2.c", |
| 1046 | "src/f32-dwconv/gen/up8x25-avx.c", |
| 1047 | "src/f32-gemm/gen/1x8-avx-broadcast.c", |
| 1048 | "src/f32-gemm/gen/4x8-avx-broadcast.c", |
| 1049 | "src/f32-gemm/gen/5x8-avx-broadcast.c", |
| 1050 | "src/f32-gemm/gen/6x8-avx-broadcast.c", |
| 1051 | "src/f32-gemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1052 | "src/f32-gemm/gen/1x16-avx-broadcast.c", |
| 1053 | "src/f32-gemm/gen/3x16-avx-broadcast.c", |
| 1054 | "src/f32-gemm/gen/4x16-avx-broadcast.c", |
| 1055 | "src/f32-gemm/gen/5x16-avx-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1056 | "src/f32-gemm/gen-inc/1x8-avx-broadcast.c", |
| 1057 | "src/f32-gemm/gen-inc/4x8-avx-broadcast.c", |
| 1058 | "src/f32-gemm/gen-inc/5x8-avx-broadcast.c", |
| 1059 | "src/f32-gemm/gen-inc/6x8-avx-broadcast.c", |
| 1060 | "src/f32-gemm/gen-inc/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1061 | "src/f32-gemm/gen-inc/1x16-avx-broadcast.c", |
| 1062 | "src/f32-gemm/gen-inc/3x16-avx-broadcast.c", |
| 1063 | "src/f32-gemm/gen-inc/4x16-avx-broadcast.c", |
| 1064 | "src/f32-gemm/gen-inc/5x16-avx-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1065 | "src/f32-hswish/gen/avx-x8.c", |
| 1066 | "src/f32-hswish/gen/avx-x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1067 | "src/f32-igemm/gen/1x8-avx-broadcast.c", |
| 1068 | "src/f32-igemm/gen/4x8-avx-broadcast.c", |
| 1069 | "src/f32-igemm/gen/5x8-avx-broadcast.c", |
| 1070 | "src/f32-igemm/gen/6x8-avx-broadcast.c", |
| 1071 | "src/f32-igemm/gen/7x8-avx-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1072 | "src/f32-igemm/gen/1x16-avx-broadcast.c", |
| 1073 | "src/f32-igemm/gen/3x16-avx-broadcast.c", |
| 1074 | "src/f32-igemm/gen/4x16-avx-broadcast.c", |
| 1075 | "src/f32-igemm/gen/5x16-avx-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame^] | 1076 | "src/f32-prelu/gen/avx-2x8.c", |
| 1077 | "src/f32-prelu/gen/avx-2x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1078 | "src/f32-rmax/avx.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1079 | "src/f32-vbinary/gen/vadd-avx-x8.c", |
| 1080 | "src/f32-vbinary/gen/vadd-avx-x16.c", |
| 1081 | "src/f32-vbinary/gen/vaddc-avx-x8.c", |
| 1082 | "src/f32-vbinary/gen/vaddc-avx-x16.c", |
| 1083 | "src/f32-vbinary/gen/vdiv-avx-x8.c", |
| 1084 | "src/f32-vbinary/gen/vdiv-avx-x16.c", |
| 1085 | "src/f32-vbinary/gen/vdivc-avx-x8.c", |
| 1086 | "src/f32-vbinary/gen/vdivc-avx-x16.c", |
| 1087 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 1088 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 1089 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 1090 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 1091 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 1092 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 1093 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 1094 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
| 1095 | "src/f32-vbinary/gen/vmul-avx-x8.c", |
| 1096 | "src/f32-vbinary/gen/vmul-avx-x16.c", |
| 1097 | "src/f32-vbinary/gen/vmulc-avx-x8.c", |
| 1098 | "src/f32-vbinary/gen/vmulc-avx-x16.c", |
| 1099 | "src/f32-vbinary/gen/vrdivc-avx-x8.c", |
| 1100 | "src/f32-vbinary/gen/vrdivc-avx-x16.c", |
| 1101 | "src/f32-vbinary/gen/vrsubc-avx-x8.c", |
| 1102 | "src/f32-vbinary/gen/vrsubc-avx-x16.c", |
| 1103 | "src/f32-vbinary/gen/vsub-avx-x8.c", |
| 1104 | "src/f32-vbinary/gen/vsub-avx-x16.c", |
| 1105 | "src/f32-vbinary/gen/vsubc-avx-x8.c", |
| 1106 | "src/f32-vbinary/gen/vsubc-avx-x16.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1107 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1108 | ] |
| 1109 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1110 | FMA3_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1111 | "src/f32-dwconv/gen/up16x4-fma3-acc2.c", |
| 1112 | "src/f32-dwconv/gen/up16x4-fma3.c", |
| 1113 | "src/f32-dwconv/gen/up8x4-fma3-acc2.c", |
| 1114 | "src/f32-dwconv/gen/up8x4-fma3.c", |
| 1115 | "src/f32-dwconv/gen/up16x9-fma3-acc2.c", |
| 1116 | "src/f32-dwconv/gen/up16x9-fma3.c", |
| 1117 | "src/f32-dwconv/gen/up8x9-fma3-acc2.c", |
| 1118 | "src/f32-dwconv/gen/up8x9-fma3.c", |
| 1119 | "src/f32-dwconv/gen/up16x25-fma3-acc2.c", |
| 1120 | "src/f32-dwconv/gen/up16x25-fma3.c", |
| 1121 | "src/f32-dwconv/gen/up8x25-fma3-acc2.c", |
| 1122 | "src/f32-dwconv/gen/up8x25-fma3.c", |
| 1123 | "src/f32-gemm/gen/1x8-fma3-broadcast.c", |
| 1124 | "src/f32-gemm/gen/4x8-fma3-broadcast.c", |
| 1125 | "src/f32-gemm/gen/5x8-fma3-broadcast.c", |
| 1126 | "src/f32-gemm/gen/6x8-fma3-broadcast.c", |
| 1127 | "src/f32-gemm/gen/7x8-fma3-broadcast.c", |
| 1128 | "src/f32-gemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1129 | "src/f32-gemm/gen/1x16-fma3-broadcast.c", |
| 1130 | "src/f32-gemm/gen/3x16-fma3-broadcast.c", |
| 1131 | "src/f32-gemm/gen/4x16-fma3-broadcast.c", |
| 1132 | "src/f32-gemm/gen/5x16-fma3-broadcast.c", |
Ashkan Aliabadi | d94b856 | 2019-12-10 11:33:51 -0800 | [diff] [blame] | 1133 | "src/f32-gemm/gen/1x16s4-fma3-broadcast.c", |
| 1134 | "src/f32-gemm/gen/3x16s4-fma3-broadcast.c", |
| 1135 | "src/f32-gemm/gen/4x16s4-fma3-broadcast.c", |
| 1136 | "src/f32-gemm/gen/5x16s4-fma3-broadcast.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1137 | "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c", |
| 1138 | "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c", |
| 1139 | "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c", |
| 1140 | "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c", |
| 1141 | "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c", |
| 1142 | "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1143 | "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c", |
| 1144 | "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c", |
| 1145 | "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c", |
| 1146 | "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c", |
Ashkan Aliabadi | d94b856 | 2019-12-10 11:33:51 -0800 | [diff] [blame] | 1147 | "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c", |
| 1148 | "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c", |
| 1149 | "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c", |
| 1150 | "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1151 | "src/f32-hswish/gen/fma3-x8.c", |
| 1152 | "src/f32-hswish/gen/fma3-x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1153 | "src/f32-igemm/gen/1x8-fma3-broadcast.c", |
| 1154 | "src/f32-igemm/gen/4x8-fma3-broadcast.c", |
| 1155 | "src/f32-igemm/gen/5x8-fma3-broadcast.c", |
| 1156 | "src/f32-igemm/gen/6x8-fma3-broadcast.c", |
| 1157 | "src/f32-igemm/gen/7x8-fma3-broadcast.c", |
| 1158 | "src/f32-igemm/gen/8x8-fma3-broadcast.c", |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 1159 | "src/f32-igemm/gen/1x16-fma3-broadcast.c", |
| 1160 | "src/f32-igemm/gen/3x16-fma3-broadcast.c", |
| 1161 | "src/f32-igemm/gen/4x16-fma3-broadcast.c", |
| 1162 | "src/f32-igemm/gen/5x16-fma3-broadcast.c", |
Ashkan Aliabadi | d94b856 | 2019-12-10 11:33:51 -0800 | [diff] [blame] | 1163 | "src/f32-igemm/gen/1x16s4-fma3-broadcast.c", |
| 1164 | "src/f32-igemm/gen/3x16s4-fma3-broadcast.c", |
| 1165 | "src/f32-igemm/gen/4x16s4-fma3-broadcast.c", |
| 1166 | "src/f32-igemm/gen/5x16s4-fma3-broadcast.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1167 | ] |
| 1168 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1169 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1170 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 1171 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1172 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1173 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 1174 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1175 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 1176 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1177 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1178 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 1179 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1180 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1181 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 1182 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 1183 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 1184 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 1185 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 1186 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 1187 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 1188 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 1189 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 1190 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 1191 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 1192 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 1193 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 1194 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 1195 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1196 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1197 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 1198 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1199 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 1200 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1201 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1202 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 1203 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1204 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1205 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1206 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 1207 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 1208 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 1209 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 1210 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 1211 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 1212 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 1213 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 1214 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 1215 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 1216 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 1217 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 1218 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 1219 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 1220 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 1221 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 1222 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 1223 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 1224 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 1225 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 1226 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 1227 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 1228 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 1229 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 1230 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 1231 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 1232 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 1233 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 1234 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 1235 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1236 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 1237 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 1238 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 1239 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 1240 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 1241 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 1242 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 1243 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 1244 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 1245 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 1246 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 1247 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 1248 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 1249 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 1250 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 1251 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 1252 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 1253 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 1254 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 1255 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 1256 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 1257 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 1258 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 1259 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1260 | "src/math/exp-avx2-p5.c", |
| 1261 | "src/math/exp-avx2-perm-p3.c", |
| 1262 | "src/math/exp-avx2-perm-p4.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 1263 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1264 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 1265 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 1266 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 1267 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| 1268 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 1269 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 1270 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1271 | ] |
| 1272 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1273 | AVX512F_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1274 | "src/f32-clamp/gen/avx512f-x16.c", |
| 1275 | "src/f32-clamp/gen/avx512f-x32.c", |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 1276 | "src/f32-dwconv/gen/up32x4-avx512f-acc2.c", |
| 1277 | "src/f32-dwconv/gen/up32x4-avx512f.c", |
| 1278 | "src/f32-dwconv/gen/up16x4-avx512f-acc2.c", |
| 1279 | "src/f32-dwconv/gen/up16x4-avx512f.c", |
| 1280 | "src/f32-dwconv/gen/up32x9-avx512f-acc2.c", |
| 1281 | "src/f32-dwconv/gen/up32x9-avx512f.c", |
| 1282 | "src/f32-dwconv/gen/up16x9-avx512f-acc2.c", |
| 1283 | "src/f32-dwconv/gen/up16x9-avx512f.c", |
| 1284 | "src/f32-dwconv/gen/up32x25-avx512f-acc2.c", |
| 1285 | "src/f32-dwconv/gen/up32x25-avx512f.c", |
| 1286 | "src/f32-dwconv/gen/up16x25-avx512f-acc2.c", |
| 1287 | "src/f32-dwconv/gen/up16x25-avx512f.c", |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1288 | "src/f32-gemm/gen/1x16-avx512f-broadcast.c", |
| 1289 | "src/f32-gemm/gen/4x16-avx512f-broadcast.c", |
| 1290 | "src/f32-gemm/gen/5x16-avx512f-broadcast.c", |
| 1291 | "src/f32-gemm/gen/6x16-avx512f-broadcast.c", |
| 1292 | "src/f32-gemm/gen/7x16-avx512f-broadcast.c", |
| 1293 | "src/f32-gemm/gen/8x16-avx512f-broadcast.c", |
| 1294 | "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c", |
| 1295 | "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c", |
| 1296 | "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c", |
| 1297 | "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c", |
| 1298 | "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c", |
| 1299 | "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1300 | "src/f32-hswish/gen/avx512f-x16.c", |
| 1301 | "src/f32-hswish/gen/avx512f-x32.c", |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1302 | "src/f32-igemm/gen/1x16-avx512f-broadcast.c", |
| 1303 | "src/f32-igemm/gen/4x16-avx512f-broadcast.c", |
| 1304 | "src/f32-igemm/gen/5x16-avx512f-broadcast.c", |
| 1305 | "src/f32-igemm/gen/6x16-avx512f-broadcast.c", |
| 1306 | "src/f32-igemm/gen/7x16-avx512f-broadcast.c", |
| 1307 | "src/f32-igemm/gen/8x16-avx512f-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame^] | 1308 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 1309 | "src/f32-prelu/gen/avx512f-2x32.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1310 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1311 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1312 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1313 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1314 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1315 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1316 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1317 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1318 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1319 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1320 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1321 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1322 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 1323 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1324 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1325 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 1326 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1327 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 1328 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1329 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1330 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 1331 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1332 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1333 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1334 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1335 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1336 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1337 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1338 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1339 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1340 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1341 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1342 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1343 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1344 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1345 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1346 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1347 | "src/f32-vbinary/gen/vadd-avx512f-x16.c", |
| 1348 | "src/f32-vbinary/gen/vadd-avx512f-x32.c", |
| 1349 | "src/f32-vbinary/gen/vaddc-avx512f-x16.c", |
| 1350 | "src/f32-vbinary/gen/vaddc-avx512f-x32.c", |
| 1351 | "src/f32-vbinary/gen/vdiv-avx512f-x16.c", |
| 1352 | "src/f32-vbinary/gen/vdiv-avx512f-x32.c", |
| 1353 | "src/f32-vbinary/gen/vdivc-avx512f-x16.c", |
| 1354 | "src/f32-vbinary/gen/vdivc-avx512f-x32.c", |
| 1355 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 1356 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 1357 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 1358 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 1359 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 1360 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 1361 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 1362 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
| 1363 | "src/f32-vbinary/gen/vmul-avx512f-x16.c", |
| 1364 | "src/f32-vbinary/gen/vmul-avx512f-x32.c", |
| 1365 | "src/f32-vbinary/gen/vmulc-avx512f-x16.c", |
| 1366 | "src/f32-vbinary/gen/vmulc-avx512f-x32.c", |
| 1367 | "src/f32-vbinary/gen/vrdivc-avx512f-x16.c", |
| 1368 | "src/f32-vbinary/gen/vrdivc-avx512f-x32.c", |
| 1369 | "src/f32-vbinary/gen/vrsubc-avx512f-x16.c", |
| 1370 | "src/f32-vbinary/gen/vrsubc-avx512f-x32.c", |
| 1371 | "src/f32-vbinary/gen/vsub-avx512f-x16.c", |
| 1372 | "src/f32-vbinary/gen/vsub-avx512f-x32.c", |
| 1373 | "src/f32-vbinary/gen/vsubc-avx512f-x16.c", |
| 1374 | "src/f32-vbinary/gen/vsubc-avx512f-x32.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1375 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1376 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 1377 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 1378 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 1379 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 1380 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 1381 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 1382 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 1383 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1384 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1385 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1386 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 1387 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1388 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 1389 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 1390 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 1391 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 1392 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 1393 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 1394 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 1395 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 1396 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 1397 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 1398 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 1399 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1400 | "src/math/exp-avx512f-p5-scalef.c", |
| 1401 | "src/math/exp-avx512f-p5.c", |
| 1402 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | feb4923 | 2019-10-28 11:03:31 -0700 | [diff] [blame] | 1403 | "src/math/exp-avx512f-perm2-p2.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1404 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1405 | ] |
| 1406 | |
| 1407 | AARCH32_ASM_UKERNELS = [ |
| 1408 | "src/q8-dwconv/up8x9-aarch32-neon.S", |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 1409 | "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S", |
Frank Barchard | abf8154 | 2019-12-13 16:18:30 -0800 | [diff] [blame] | 1410 | "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S", |
| 1411 | "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S", |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 1412 | "src/f32-gemm/4x8-aarch32-neon-ld64.S", |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 1413 | "src/f32-igemm/4x8-aarch32-neon-ld64.S", |
Frank Barchard | 90ce789 | 2020-02-10 23:35:45 -0800 | [diff] [blame] | 1414 | "src/f32-igemm/gen/4x8-aarch32-neon-cortex-a75.S", |
| 1415 | "src/f32-igemm/gen/4x8-aarch32-neon-pld-cortex-a75.S", |
Frank Barchard | c87a8fd | 2020-02-12 13:02:52 -0800 | [diff] [blame] | 1416 | "src/f32-igemm/4x8-aarch32-neon-cortex-a53.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1417 | ] |
| 1418 | |
| 1419 | AARCH64_ASM_UKERNELS = [ |
| 1420 | "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S", |
| 1421 | "src/f32-dwconv/up4x9-aarch64-neonfma.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1422 | "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S", |
| 1423 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S", |
| 1424 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1425 | "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
| 1426 | "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S", |
| 1427 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S", |
| 1428 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1429 | "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1430 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S", |
| 1431 | "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1432 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1433 | "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
| 1434 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 91e1999 | 2020-03-09 18:46:14 -0700 | [diff] [blame] | 1435 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a55.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1436 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1437 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1438 | "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
| 1439 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S", |
| 1440 | "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S", |
| 1441 | "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S", |
| 1442 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S", |
| 1443 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S", |
| 1444 | "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S", |
| 1445 | "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S", |
| 1446 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S", |
| 1447 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S", |
| 1448 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S", |
| 1449 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S", |
| 1450 | "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1451 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1452 | "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S", |
| 1453 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 91e1999 | 2020-03-09 18:46:14 -0700 | [diff] [blame] | 1454 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a55.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1455 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1456 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1457 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S", |
| 1458 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S", |
| 1459 | "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1460 | "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 1461 | "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1462 | "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S", |
| 1463 | "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1464 | "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 1465 | "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1466 | "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S", |
| 1467 | "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S", |
| 1468 | "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S", |
| 1469 | "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 1470 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 91e1999 | 2020-03-09 18:46:14 -0700 | [diff] [blame] | 1471 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a55.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1472 | "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S", |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 1473 | "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S", |
| 1474 | "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1475 | ] |
| 1476 | |
| 1477 | INTERNAL_MICROKERNEL_HDRS = [ |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1478 | "src/requantization/gemmlowp-requantization.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1479 | "src/xnnpack/argmaxpool.h", |
| 1480 | "src/xnnpack/avgpool.h", |
| 1481 | "src/xnnpack/clamp.h", |
| 1482 | "src/xnnpack/common.h", |
| 1483 | "src/xnnpack/conv.h", |
| 1484 | "src/xnnpack/dwconv.h", |
| 1485 | "src/xnnpack/gavgpool.h", |
| 1486 | "src/xnnpack/gemm.h", |
| 1487 | "src/xnnpack/hswish.h", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1488 | "src/xnnpack/ibilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1489 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 1490 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1491 | "src/xnnpack/lut.h", |
| 1492 | "src/xnnpack/math.h", |
| 1493 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1494 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1495 | "src/xnnpack/packx.h", |
| 1496 | "src/xnnpack/pad.h", |
| 1497 | "src/xnnpack/params.h", |
| 1498 | "src/xnnpack/pavgpool.h", |
| 1499 | "src/xnnpack/ppmm.h", |
| 1500 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1501 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1502 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1503 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1504 | "src/xnnpack/rmax.h", |
| 1505 | "src/xnnpack/scalar-utils.h", |
| 1506 | "src/xnnpack/spmm.h", |
| 1507 | "src/xnnpack/unpool.h", |
| 1508 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1509 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1510 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1511 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 1512 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 1513 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1514 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1515 | "src/xnnpack/zip.h", |
| 1516 | ] |
| 1517 | |
| 1518 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1519 | "include/xnnpack.h", |
| 1520 | "src/xnnpack/allocator.h", |
| 1521 | "src/xnnpack/compute.h", |
| 1522 | "src/xnnpack/im2col.h", |
| 1523 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1524 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1525 | "src/xnnpack/operator.h", |
| 1526 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1527 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1528 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1529 | "src/xnnpack/requantization.h", |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 1530 | "src/xnnpack/subgraph.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1531 | ] |
| 1532 | |
| 1533 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1534 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1535 | ] |
| 1536 | |
| 1537 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1538 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1539 | "include/xnnpack.h", |
| 1540 | ] |
| 1541 | |
| 1542 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| 1543 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 1544 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1545 | "src/xnnpack/requantization.h", |
| 1546 | "include/xnnpack.h", |
| 1547 | ] |
| 1548 | |
| 1549 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 1550 | "src/xnnpack/params.h", |
| 1551 | "src/xnnpack/common.h", |
| 1552 | ] |
| 1553 | |
| 1554 | WEIGHTS_PACK_HDRS = [ |
| 1555 | "src/xnnpack/pack.h", |
| 1556 | "src/xnnpack/operator.h", |
| 1557 | "src/xnnpack/compute.h", |
| 1558 | ] |
| 1559 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1560 | LOGGING_COPTS = select({ |
| 1561 | # No logging in optimized mode |
| 1562 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 1563 | # Full logging in debug mode |
| 1564 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 1565 | # Error-only logging in default (fastbuild) mode |
| 1566 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 1567 | }) |
| 1568 | |
| 1569 | LOGGING_HDRS = [ |
| 1570 | "src/xnnpack/log.h", |
| 1571 | ] |
| 1572 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1573 | xnnpack_cc_library( |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1574 | name = "tables", |
| 1575 | srcs = TABLE_SRCS, |
| 1576 | hdrs = INTERNAL_HDRS, |
| 1577 | copts = xnnpack_std_copts(), |
| 1578 | ) |
| 1579 | |
| 1580 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1581 | name = "scalar_ukernels", |
| 1582 | srcs = SCALAR_UKERNELS, |
| 1583 | hdrs = INTERNAL_HDRS, |
| 1584 | aarch32_copts = ["-marm"], |
| 1585 | copts = xnnpack_std_copts(), |
| 1586 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1587 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1588 | "@FP16", |
| 1589 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1590 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1591 | ], |
| 1592 | ) |
| 1593 | |
| 1594 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1595 | name = "wasm_ukernels", |
| 1596 | hdrs = INTERNAL_HDRS, |
| 1597 | copts = xnnpack_std_copts(), |
| 1598 | wasm_srcs = WASM_UKERNELS, |
| 1599 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1600 | ":tables", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1601 | "@FP16", |
| 1602 | "@FXdiv", |
| 1603 | "@pthreadpool", |
| 1604 | ], |
| 1605 | ) |
| 1606 | |
| 1607 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1608 | name = "psimd_fastmath_ukernels", |
| 1609 | srcs = PSIMD_FASTMATH_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1610 | hdrs = INTERNAL_HDRS, |
| 1611 | aarch32_copts = [ |
| 1612 | "-marm", |
| 1613 | "-mfpu=neon", |
| 1614 | ], |
| 1615 | copts = xnnpack_std_copts(), |
| 1616 | optimized_copts = [ |
| 1617 | "-O3", |
| 1618 | "-ffast-math", |
| 1619 | ], |
| 1620 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1621 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1622 | "@FP16", |
| 1623 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1624 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1625 | ], |
| 1626 | ) |
| 1627 | |
| 1628 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1629 | name = "psimd_accmath_ukernels", |
| 1630 | srcs = PSIMD_ACCMATH_UKERNELS, |
| 1631 | hdrs = INTERNAL_HDRS, |
| 1632 | aarch32_copts = [ |
| 1633 | "-marm", |
| 1634 | "-mfpu=neon", |
| 1635 | ], |
| 1636 | copts = xnnpack_std_copts(), |
| 1637 | optimized_copts = [ |
| 1638 | "-O3", |
| 1639 | ], |
| 1640 | deps = [ |
| 1641 | ":tables", |
| 1642 | "@FP16", |
| 1643 | "@psimd", |
| 1644 | "@pthreadpool", |
| 1645 | ], |
| 1646 | ) |
| 1647 | |
| 1648 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1649 | name = "neon_ukernels", |
| 1650 | hdrs = INTERNAL_HDRS, |
| 1651 | aarch32_copts = [ |
| 1652 | "-marm", |
| 1653 | "-mfpu=neon", |
| 1654 | ], |
| 1655 | aarch32_srcs = NEON_UKERNELS, |
| 1656 | aarch64_srcs = NEON_UKERNELS, |
| 1657 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1658 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1659 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1660 | "@FP16", |
| 1661 | "@pthreadpool", |
| 1662 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1663 | ) |
| 1664 | |
| 1665 | xnnpack_cc_library( |
| 1666 | name = "neonfma_ukernels", |
| 1667 | hdrs = INTERNAL_HDRS, |
| 1668 | aarch32_copts = [ |
| 1669 | "-marm", |
| 1670 | "-mfpu=neon-vfpv4", |
| 1671 | ], |
| 1672 | aarch32_srcs = NEONFMA_UKERNELS, |
| 1673 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 1674 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1675 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1676 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1677 | "@FP16", |
| 1678 | "@pthreadpool", |
| 1679 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1680 | ) |
| 1681 | |
| 1682 | xnnpack_cc_library( |
| 1683 | name = "neonfp16arith_ukernels", |
| 1684 | hdrs = INTERNAL_HDRS, |
| 1685 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 1686 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 1687 | copts = xnnpack_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1688 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1689 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1690 | "@FP16", |
| 1691 | "@pthreadpool", |
| 1692 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1693 | ) |
| 1694 | |
| 1695 | xnnpack_cc_library( |
| 1696 | name = "sse2_ukernels", |
| 1697 | hdrs = INTERNAL_HDRS, |
| 1698 | copts = xnnpack_std_copts(), |
| 1699 | x86_copts = ["-msse2"], |
| 1700 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1701 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1702 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1703 | "@FP16", |
| 1704 | "@pthreadpool", |
| 1705 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1706 | ) |
| 1707 | |
| 1708 | xnnpack_cc_library( |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1709 | name = "ssse3_ukernels", |
| 1710 | hdrs = INTERNAL_HDRS, |
| 1711 | copts = xnnpack_std_copts(), |
| 1712 | x86_copts = ["-mssse3"], |
| 1713 | x86_srcs = SSSE3_UKERNELS, |
| 1714 | deps = [ |
| 1715 | ":tables", |
| 1716 | "@FP16", |
| 1717 | "@pthreadpool", |
| 1718 | ], |
| 1719 | ) |
| 1720 | |
| 1721 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1722 | name = "sse41_ukernels", |
| 1723 | hdrs = INTERNAL_HDRS, |
| 1724 | copts = xnnpack_std_copts(), |
| 1725 | x86_copts = ["-msse4.1"], |
| 1726 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1727 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1728 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1729 | "@FP16", |
| 1730 | "@pthreadpool", |
| 1731 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1732 | ) |
| 1733 | |
| 1734 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1735 | name = "avx_ukernels", |
| 1736 | hdrs = INTERNAL_HDRS, |
| 1737 | copts = xnnpack_std_copts(), |
| 1738 | x86_copts = ["-mavx"], |
| 1739 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1740 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1741 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1742 | "@FP16", |
| 1743 | "@pthreadpool", |
| 1744 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1745 | ) |
| 1746 | |
| 1747 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1748 | name = "fma3_ukernels", |
| 1749 | hdrs = INTERNAL_HDRS, |
| 1750 | copts = xnnpack_std_copts(), |
| 1751 | x86_copts = [ |
| 1752 | "-mfma", |
| 1753 | ], |
| 1754 | x86_srcs = FMA3_UKERNELS, |
| 1755 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1756 | ":tables", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1757 | "@FP16", |
| 1758 | "@pthreadpool", |
| 1759 | ], |
| 1760 | ) |
| 1761 | |
| 1762 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1763 | name = "avx2_ukernels", |
| 1764 | hdrs = INTERNAL_HDRS, |
| 1765 | copts = xnnpack_std_copts(), |
| 1766 | x86_copts = [ |
| 1767 | "-mfma", |
| 1768 | "-mavx2", |
| 1769 | ], |
| 1770 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1771 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1772 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1773 | "@FP16", |
| 1774 | "@pthreadpool", |
| 1775 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1776 | ) |
| 1777 | |
| 1778 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1779 | name = "avx512f_ukernels", |
| 1780 | hdrs = INTERNAL_HDRS, |
| 1781 | copts = xnnpack_std_copts(), |
| 1782 | x86_copts = ["-mavx512f"], |
| 1783 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1784 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 1785 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1786 | "@FP16", |
| 1787 | "@pthreadpool", |
| 1788 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1789 | ) |
| 1790 | |
| 1791 | xnnpack_cc_library( |
| 1792 | name = "asm_ukernels", |
| 1793 | hdrs = ["src/xnnpack/assembly.h"], |
| 1794 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
| 1795 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 1796 | ) |
| 1797 | |
| 1798 | xnnpack_aggregate_library( |
| 1799 | name = "ukernels", |
| 1800 | aarch32_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1801 | ":psimd_fastmath_ukernels", |
| 1802 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1803 | ":neon_ukernels", |
| 1804 | ":neonfma_ukernels", |
| 1805 | ":asm_ukernels", |
| 1806 | ], |
| 1807 | aarch64_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1808 | ":psimd_fastmath_ukernels", |
| 1809 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1810 | ":neon_ukernels", |
| 1811 | ":neonfma_ukernels", |
| 1812 | ":neonfp16arith_ukernels", |
| 1813 | ":asm_ukernels", |
| 1814 | ], |
| 1815 | generic_deps = [":scalar_ukernels"], |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1816 | wasm_deps = [ |
| 1817 | ":wasm_ukernels", |
| 1818 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1819 | wasmsimd_deps = [ |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1820 | ":wasm_ukernels", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1821 | ":psimd_fastmath_ukernels", |
| 1822 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1823 | ], |
| 1824 | x86_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1825 | ":psimd_fastmath_ukernels", |
| 1826 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1827 | ":sse2_ukernels", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1828 | ":ssse3_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1829 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1830 | ":avx_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1831 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1832 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1833 | ":avx512f_ukernels", |
| 1834 | ], |
| 1835 | ) |
| 1836 | |
| 1837 | xnnpack_cc_library( |
| 1838 | name = "im2col", |
| 1839 | srcs = ["src/im2col.c"], |
| 1840 | hdrs = [ |
| 1841 | "src/xnnpack/common.h", |
| 1842 | "src/xnnpack/im2col.h", |
| 1843 | ], |
| 1844 | copts = xnnpack_std_copts(), |
| 1845 | ) |
| 1846 | |
| 1847 | xnnpack_cc_library( |
| 1848 | name = "indirection", |
| 1849 | srcs = ["src/indirection.c"], |
| 1850 | hdrs = INTERNAL_HDRS, |
| 1851 | copts = xnnpack_std_copts(), |
| 1852 | deps = [ |
| 1853 | "@FP16", |
| 1854 | "@FXdiv", |
| 1855 | "@pthreadpool", |
| 1856 | ], |
| 1857 | ) |
| 1858 | |
| 1859 | xnnpack_cc_library( |
| 1860 | name = "operator_run", |
| 1861 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1862 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 1863 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1864 | # Wrappers for multi-pass microkernels use VLAs for temporary buffers. |
| 1865 | "-Wno-vla", |
| 1866 | ], |
| 1867 | deps = [ |
| 1868 | "@FP16", |
| 1869 | "@FXdiv", |
| 1870 | "@clog", |
| 1871 | "@pthreadpool", |
| 1872 | ], |
| 1873 | ) |
| 1874 | |
| 1875 | cc_library( |
| 1876 | name = "enable_assembly", |
| 1877 | defines = select({ |
| 1878 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 1879 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 1880 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1881 | }), |
| 1882 | ) |
| 1883 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1884 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1885 | name = "operators", |
| 1886 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1887 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1888 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1889 | ], |
| 1890 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 1891 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1892 | "-Isrc", |
| 1893 | "-Iinclude", |
| 1894 | ] + select({ |
| 1895 | ":debug_build": [], |
| 1896 | "//conditions:default": xnnpack_min_size_copts(), |
| 1897 | }), |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1898 | wasm_srcs = ["src/wasm-stubs.c"], |
| 1899 | wasmsimd_srcs = ["src/wasm-stubs.c"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1900 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1901 | ":indirection", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1902 | "@FP16", |
| 1903 | "@FXdiv", |
| 1904 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1905 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1906 | ], |
| 1907 | ) |
| 1908 | |
| 1909 | cc_library( |
| 1910 | name = "XNNPACK", |
| 1911 | srcs = [ |
| 1912 | "src/init.c", |
Marat Dukhan | ccfdbd1 | 2020-02-03 14:27:45 -0800 | [diff] [blame] | 1913 | "src/runtime.c", |
| 1914 | "src/subgraph.c", |
| 1915 | "src/tensor.c", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1916 | ], |
| 1917 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1918 | "-Isrc", |
| 1919 | "-Iinclude", |
| 1920 | ] + select({ |
| 1921 | ":debug_build": [], |
| 1922 | "//conditions:default": xnnpack_min_size_copts(), |
| 1923 | }), |
| 1924 | includes = ["include"], |
| 1925 | linkstatic = True, |
| 1926 | textual_hdrs = ["include/xnnpack.h"], |
| 1927 | visibility = xnnpack_visibility(), |
| 1928 | deps = [ |
| 1929 | ":enable_assembly", |
| 1930 | ":ukernels", |
| 1931 | ":operator_run", |
| 1932 | ":operators", |
| 1933 | "@clog", |
| 1934 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 1935 | ] + select({ |
| 1936 | ":emscripten": [], |
| 1937 | "//conditions:default": ["@cpuinfo"], |
| 1938 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1939 | ) |
| 1940 | |
| 1941 | cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1942 | name = "xnnpack_operators_nhwc_f32", |
| 1943 | srcs = [ |
| 1944 | "src/init.c", |
| 1945 | ], |
| 1946 | copts = xnnpack_std_copts() + LOGGING_COPTS + [ |
| 1947 | "-Isrc", |
| 1948 | "-Iinclude", |
| 1949 | ] + select({ |
| 1950 | ":debug_build": [], |
| 1951 | "//conditions:default": xnnpack_min_size_copts(), |
| 1952 | }), |
| 1953 | defines = [ |
| 1954 | "XNN_NO_Q8_OPERATORS", |
| 1955 | "XNN_NO_U8_OPERATORS", |
| 1956 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1957 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1958 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1959 | includes = ["include"], |
| 1960 | linkstatic = True, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1961 | textual_hdrs = ["include/xnnpack.h"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1962 | visibility = xnnpack_visibility(), |
| 1963 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1964 | ":enable_assembly", |
| 1965 | ":ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1966 | ":operator_run", |
| 1967 | ":operators", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1968 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1969 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1970 | ] + select({ |
| 1971 | ":emscripten": [], |
| 1972 | "//conditions:default": ["@cpuinfo"], |
| 1973 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1974 | ) |
| 1975 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 1976 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1977 | name = "bench_utils", |
| 1978 | srcs = ["bench/utils.cc"], |
| 1979 | hdrs = ["bench/utils.h"], |
| 1980 | copts = ["-Wno-unused-result"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 1981 | deps = [ |
| 1982 | "@com_google_benchmark//:benchmark", |
| 1983 | "@cpuinfo", |
| 1984 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1985 | ) |
| 1986 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 1987 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1988 | |
| 1989 | xnnpack_benchmark( |
| 1990 | name = "q8_gemm_bench", |
| 1991 | srcs = [ |
| 1992 | "bench/gemm.h", |
| 1993 | "bench/q8-gemm.cc", |
| 1994 | "src/xnnpack/AlignedAllocator.h", |
| 1995 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 1996 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 1997 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
| 1998 | ) |
| 1999 | |
| 2000 | xnnpack_benchmark( |
| 2001 | name = "f16_gemm_bench", |
| 2002 | srcs = [ |
| 2003 | "bench/f16-gemm.cc", |
| 2004 | "bench/gemm.h", |
| 2005 | "src/xnnpack/AlignedAllocator.h", |
| 2006 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2007 | copts = ["-Wno-unused-function"], |
| 2008 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2009 | ) |
| 2010 | |
| 2011 | xnnpack_benchmark( |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 2012 | name = "f16_spmm_bench", |
| 2013 | srcs = [ |
| 2014 | "bench/f16-spmm.cc", |
| 2015 | "bench/gemm.h", |
| 2016 | "bench/google/gemm.h", |
| 2017 | "src/xnnpack/AlignedAllocator.h", |
| 2018 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2019 | copts = ["-Wno-unused-function"], |
| 2020 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2021 | ) |
| 2022 | |
| 2023 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2024 | name = "f32_igemm_bench", |
| 2025 | srcs = [ |
| 2026 | "bench/f32-igemm.cc", |
| 2027 | "bench/conv.h", |
| 2028 | "src/xnnpack/AlignedAllocator.h", |
| 2029 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 2030 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2031 | ) |
| 2032 | |
| 2033 | xnnpack_benchmark( |
| 2034 | name = "f32_conv_hwc_bench", |
| 2035 | srcs = [ |
| 2036 | "bench/f32-conv-hwc.cc", |
| 2037 | "bench/dconv.h", |
| 2038 | "src/xnnpack/AlignedAllocator.h", |
| 2039 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2040 | copts = ["-Wno-unused-function"], |
| 2041 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2042 | ) |
| 2043 | |
| 2044 | xnnpack_benchmark( |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 2045 | name = "f32_conv_hwc2spchw_bench", |
| 2046 | srcs = [ |
| 2047 | "bench/f32-conv-hwc2spchw.cc", |
| 2048 | "bench/dconv.h", |
| 2049 | "src/xnnpack/AlignedAllocator.h", |
| 2050 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2051 | copts = ["-Wno-unused-function"], |
| 2052 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2053 | ) |
| 2054 | |
| 2055 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2056 | name = "f32_dwconv_bench", |
| 2057 | srcs = [ |
| 2058 | "bench/f32-dwconv.cc", |
| 2059 | "bench/dwconv.h", |
| 2060 | "src/xnnpack/AlignedAllocator.h", |
| 2061 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2062 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 2063 | ) |
| 2064 | |
| 2065 | xnnpack_benchmark( |
| 2066 | name = "f32_dwconv_spchw_bench", |
| 2067 | srcs = [ |
| 2068 | "bench/f32-dwconv-spchw.cc", |
| 2069 | "bench/dwconv.h", |
| 2070 | "src/xnnpack/AlignedAllocator.h", |
| 2071 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2072 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 2073 | ) |
| 2074 | |
| 2075 | xnnpack_benchmark( |
| 2076 | name = "f32_gemm_bench", |
| 2077 | srcs = [ |
| 2078 | "bench/f32-gemm.cc", |
| 2079 | "bench/gemm.h", |
| 2080 | "src/xnnpack/AlignedAllocator.h", |
| 2081 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2082 | copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(), |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 2083 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2084 | ) |
| 2085 | |
| 2086 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2087 | name = "f32_raddexpminusmax_bench", |
| 2088 | srcs = [ |
| 2089 | "bench/f32-raddexpminusmax.cc", |
| 2090 | "src/xnnpack/AlignedAllocator.h", |
| 2091 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2092 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2093 | ) |
| 2094 | |
| 2095 | xnnpack_benchmark( |
| 2096 | name = "f32_raddextexp_bench", |
| 2097 | srcs = [ |
| 2098 | "bench/f32-raddextexp.cc", |
| 2099 | "src/xnnpack/AlignedAllocator.h", |
| 2100 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2101 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2102 | ) |
| 2103 | |
| 2104 | xnnpack_benchmark( |
| 2105 | name = "f32_raddstoreexpminusmax_bench", |
| 2106 | srcs = [ |
| 2107 | "bench/f32-raddstoreexpminusmax.cc", |
| 2108 | "src/xnnpack/AlignedAllocator.h", |
| 2109 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2110 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2111 | ) |
| 2112 | |
| 2113 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2114 | name = "f32_rmax_bench", |
| 2115 | srcs = [ |
| 2116 | "bench/f32-rmax.cc", |
| 2117 | "src/xnnpack/AlignedAllocator.h", |
| 2118 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2119 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2120 | ) |
| 2121 | |
| 2122 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 2123 | name = "f32_sigmoid_bench", |
| 2124 | srcs = [ |
| 2125 | "bench/f32-sigmoid.cc", |
| 2126 | "src/xnnpack/AlignedAllocator.h", |
| 2127 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2128 | copts = ["-Wno-unused-function"], |
| 2129 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2130 | ) |
| 2131 | |
| 2132 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2133 | name = "f32_spmm_bench", |
| 2134 | srcs = [ |
| 2135 | "bench/f32-spmm.cc", |
| 2136 | "bench/gemm.h", |
| 2137 | "src/xnnpack/AlignedAllocator.h", |
| 2138 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2139 | copts = ["-Wno-unused-function"], |
| 2140 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2141 | ) |
| 2142 | |
| 2143 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 2144 | name = "f32_softmax_bench", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 2145 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 2146 | "bench/f32-softmax.cc", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 2147 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 8d3c693 | 2020-03-06 20:27:27 -0800 | [diff] [blame] | 2148 | copts = ["-Wno-unused-function"] + xnnpack_optional_dnnl_copts(), |
| 2149 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(), |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 2150 | ) |
| 2151 | |
| 2152 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2153 | name = "f32_vscaleexpminusmax_bench", |
| 2154 | srcs = [ |
| 2155 | "bench/f32-vscaleexpminusmax.cc", |
| 2156 | "src/xnnpack/AlignedAllocator.h", |
| 2157 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2158 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2159 | ) |
| 2160 | |
| 2161 | xnnpack_benchmark( |
| 2162 | name = "f32_vscaleextexp_bench", |
| 2163 | srcs = [ |
| 2164 | "bench/f32-vscaleextexp.cc", |
| 2165 | "src/xnnpack/AlignedAllocator.h", |
| 2166 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2167 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2168 | ) |
| 2169 | |
| 2170 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2171 | name = "f32_im2col_gemm_bench", |
| 2172 | srcs = [ |
| 2173 | "bench/f32-im2col-gemm.cc", |
| 2174 | "bench/conv.h", |
| 2175 | "src/xnnpack/AlignedAllocator.h", |
| 2176 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 2177 | deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"], |
| 2178 | ) |
| 2179 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2180 | xnnpack_benchmark( |
| 2181 | name = "requantization_bench", |
| 2182 | srcs = [ |
| 2183 | "bench/requantization.cc", |
| 2184 | "src/xnnpack/requantization-stubs.h", |
| 2185 | "src/xnnpack/AlignedAllocator.h", |
| 2186 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 2187 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 2188 | ) |
| 2189 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2190 | ########################### Benchmarks for operators ########################### |
| 2191 | |
| 2192 | xnnpack_benchmark( |
| 2193 | name = "add_bench", |
| 2194 | srcs = ["bench/add.cc"], |
| 2195 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2196 | ) |
| 2197 | |
| 2198 | xnnpack_benchmark( |
| 2199 | name = "average_pooling_bench", |
| 2200 | srcs = ["bench/average-pooling.cc"], |
| 2201 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2202 | ) |
| 2203 | |
| 2204 | xnnpack_benchmark( |
| 2205 | name = "channel_shuffle_bench", |
| 2206 | srcs = ["bench/channel-shuffle.cc"], |
| 2207 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2208 | ) |
| 2209 | |
| 2210 | xnnpack_benchmark( |
| 2211 | name = "convolution_bench", |
| 2212 | srcs = ["bench/convolution.cc"], |
| 2213 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
| 2214 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
| 2215 | ) |
| 2216 | |
| 2217 | xnnpack_benchmark( |
| 2218 | name = "deconvolution_bench", |
| 2219 | srcs = ["bench/deconvolution.cc"], |
| 2220 | copts = xnnpack_optional_tflite_copts(), |
| 2221 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 2222 | ) |
| 2223 | |
| 2224 | xnnpack_benchmark( |
| 2225 | name = "global_average_pooling_bench", |
| 2226 | srcs = ["bench/global-average-pooling.cc"], |
| 2227 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2228 | ) |
| 2229 | |
| 2230 | xnnpack_benchmark( |
| 2231 | name = "max_pooling_bench", |
| 2232 | srcs = ["bench/max-pooling.cc"], |
| 2233 | deps = OPERATOR_BENCHMARK_DEPS, |
| 2234 | ) |
| 2235 | |
| 2236 | xnnpack_benchmark( |
| 2237 | name = "sigmoid_bench", |
| 2238 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 2239 | copts = xnnpack_optional_tflite_copts(), |
| 2240 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2241 | ) |
| 2242 | |
| 2243 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 2244 | name = "prelu_bench", |
| 2245 | srcs = ["bench/prelu.cc"], |
| 2246 | copts = xnnpack_optional_tflite_copts(), |
| 2247 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 2248 | ) |
| 2249 | |
| 2250 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 2251 | name = "softmax_bench", |
| 2252 | srcs = ["bench/softmax.cc"], |
Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame] | 2253 | copts = xnnpack_optional_tflite_copts(), |
| 2254 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2255 | ) |
| 2256 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2257 | ############################# End-to-end benchmarks ############################ |
| 2258 | |
| 2259 | cc_library( |
| 2260 | name = "mobilenet_v1", |
| 2261 | srcs = ["models/mobilenet-v1.cc"], |
| 2262 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2263 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2264 | linkstatic = True, |
| 2265 | deps = [ |
| 2266 | ":XNNPACK", |
| 2267 | "@pthreadpool", |
| 2268 | ], |
| 2269 | ) |
| 2270 | |
| 2271 | cc_library( |
| 2272 | name = "mobilenet_v2", |
| 2273 | srcs = ["models/mobilenet-v2.cc"], |
| 2274 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2275 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2276 | linkstatic = True, |
| 2277 | deps = [ |
| 2278 | ":XNNPACK", |
| 2279 | "@pthreadpool", |
| 2280 | ], |
| 2281 | ) |
| 2282 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2283 | cc_library( |
| 2284 | name = "mobilenet_v3_large", |
| 2285 | srcs = ["models/mobilenet-v3-large.cc"], |
| 2286 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2287 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2288 | linkstatic = True, |
| 2289 | deps = [ |
| 2290 | ":XNNPACK", |
| 2291 | "@pthreadpool", |
| 2292 | ], |
| 2293 | ) |
| 2294 | |
| 2295 | cc_library( |
| 2296 | name = "mobilenet_v3_small", |
| 2297 | srcs = ["models/mobilenet-v3-small.cc"], |
| 2298 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 2299 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2300 | linkstatic = True, |
| 2301 | deps = [ |
| 2302 | ":XNNPACK", |
| 2303 | "@pthreadpool", |
| 2304 | ], |
| 2305 | ) |
| 2306 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2307 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 2308 | name = "f32_dwconv_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2309 | srcs = [ |
| 2310 | "bench/f32-dwconv-e2e.cc", |
| 2311 | "bench/end2end.h", |
| 2312 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 2313 | copts = ["-Wno-unused-function"], |
| 2314 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 2315 | ":XNNPACK", |
| 2316 | ":mobilenet_v1", |
| 2317 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2318 | ":mobilenet_v3_large", |
| 2319 | ":mobilenet_v3_small", |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 2320 | ], |
| 2321 | ) |
| 2322 | |
| 2323 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 2324 | name = "f32_gemm_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2325 | srcs = [ |
| 2326 | "bench/f32-gemm-e2e.cc", |
| 2327 | "bench/end2end.h", |
| 2328 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 2329 | copts = ["-Wno-unused-function"], |
| 2330 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 2331 | ":XNNPACK", |
| 2332 | ":mobilenet_v1", |
| 2333 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2334 | ":mobilenet_v3_large", |
| 2335 | ":mobilenet_v3_small", |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 2336 | ], |
| 2337 | ) |
| 2338 | |
| 2339 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2340 | name = "end2end_bench", |
| 2341 | srcs = ["bench/end2end.cc"], |
| 2342 | deps = [ |
| 2343 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 2344 | ":bench_utils", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2345 | ":mobilenet_v1", |
| 2346 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2347 | ":mobilenet_v3_large", |
| 2348 | ":mobilenet_v3_small", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 2349 | "@pthreadpool", |
| 2350 | ], |
| 2351 | ) |
| 2352 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2353 | #################### Accuracy evaluation for math functions #################### |
| 2354 | |
| 2355 | xnnpack_benchmark( |
| 2356 | name = "f32_exp_eval", |
| 2357 | srcs = [ |
| 2358 | "eval/f32-exp.cc", |
| 2359 | "src/xnnpack/AlignedAllocator.h", |
| 2360 | ] + ACCURACY_EVAL_HDRS, |
| 2361 | deps = ACCURACY_EVAL_DEPS, |
| 2362 | ) |
| 2363 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 2364 | xnnpack_benchmark( |
| 2365 | name = "f32_expminus_eval", |
| 2366 | srcs = [ |
| 2367 | "eval/f32-expminus.cc", |
| 2368 | "src/xnnpack/AlignedAllocator.h", |
| 2369 | ] + ACCURACY_EVAL_HDRS, |
| 2370 | deps = ACCURACY_EVAL_DEPS, |
| 2371 | ) |
| 2372 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2373 | xnnpack_benchmark( |
| 2374 | name = "f32_extexp_eval", |
| 2375 | srcs = [ |
| 2376 | "eval/f32-extexp.cc", |
| 2377 | "src/xnnpack/AlignedAllocator.h", |
| 2378 | ] + ACCURACY_EVAL_HDRS, |
| 2379 | deps = ACCURACY_EVAL_DEPS, |
| 2380 | ) |
| 2381 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2382 | xnnpack_benchmark( |
| 2383 | name = "f32_sigmoid_eval", |
| 2384 | srcs = [ |
| 2385 | "eval/f32-sigmoid.cc", |
| 2386 | "src/xnnpack/AlignedAllocator.h", |
| 2387 | ] + ACCURACY_EVAL_HDRS, |
| 2388 | deps = ACCURACY_EVAL_DEPS, |
| 2389 | ) |
| 2390 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2391 | ######################### Unit tests for micro-kernels ######################### |
| 2392 | |
| 2393 | xnnpack_unit_test( |
| 2394 | name = "f16_gemm_test", |
| 2395 | srcs = [ |
| 2396 | "test/f16-gemm.cc", |
| 2397 | "test/gemm-microkernel-tester.h", |
| 2398 | "src/xnnpack/AlignedAllocator.h", |
| 2399 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2400 | deps = MICROKERNEL_TEST_DEPS, |
| 2401 | ) |
| 2402 | |
| 2403 | xnnpack_unit_test( |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 2404 | name = "f16_spmm_test", |
| 2405 | srcs = [ |
| 2406 | "test/f16-spmm.cc", |
| 2407 | "test/spmm-microkernel-tester.h", |
| 2408 | "src/xnnpack/AlignedAllocator.h", |
| 2409 | ] + MICROKERNEL_TEST_HDRS, |
| 2410 | deps = MICROKERNEL_TEST_DEPS, |
| 2411 | ) |
| 2412 | |
| 2413 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2414 | name = "f32_argmaxpool_test", |
| 2415 | srcs = [ |
| 2416 | "test/f32-argmaxpool.cc", |
| 2417 | "test/argmaxpool-microkernel-tester.h", |
| 2418 | "src/xnnpack/AlignedAllocator.h", |
| 2419 | ] + MICROKERNEL_TEST_HDRS, |
| 2420 | deps = MICROKERNEL_TEST_DEPS, |
| 2421 | ) |
| 2422 | |
| 2423 | xnnpack_unit_test( |
| 2424 | name = "f32_avgpool_test", |
| 2425 | srcs = [ |
| 2426 | "test/f32-avgpool.cc", |
| 2427 | "test/avgpool-microkernel-tester.h", |
| 2428 | "src/xnnpack/AlignedAllocator.h", |
| 2429 | ] + MICROKERNEL_TEST_HDRS, |
| 2430 | deps = MICROKERNEL_TEST_DEPS, |
| 2431 | ) |
| 2432 | |
| 2433 | xnnpack_unit_test( |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2434 | name = "f32_ibilinear_test", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 2435 | srcs = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2436 | "test/f32-ibilinear.cc", |
| 2437 | "test/ibilinear-microkernel-tester.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 2438 | "src/xnnpack/AlignedAllocator.h", |
| 2439 | ] + MICROKERNEL_TEST_HDRS, |
| 2440 | deps = MICROKERNEL_TEST_DEPS, |
| 2441 | ) |
| 2442 | |
| 2443 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2444 | name = "f32_clamp_test", |
| 2445 | srcs = [ |
| 2446 | "test/f32-clamp.cc", |
| 2447 | "test/clamp-microkernel-tester.h", |
| 2448 | ] + MICROKERNEL_TEST_HDRS, |
| 2449 | deps = MICROKERNEL_TEST_DEPS, |
| 2450 | ) |
| 2451 | |
| 2452 | xnnpack_unit_test( |
| 2453 | name = "f32_igemm_test", |
| 2454 | srcs = [ |
| 2455 | "test/f32-igemm.cc", |
| 2456 | "test/gemm-microkernel-tester.h", |
| 2457 | "src/xnnpack/AlignedAllocator.h", |
| 2458 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2459 | deps = MICROKERNEL_TEST_DEPS, |
| 2460 | ) |
| 2461 | |
| 2462 | xnnpack_unit_test( |
| 2463 | name = "f32_conv_hwc_test", |
| 2464 | srcs = [ |
| 2465 | "test/f32-conv-hwc.cc", |
| 2466 | "test/conv-hwc-microkernel-tester.h", |
| 2467 | "src/xnnpack/AlignedAllocator.h", |
| 2468 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2469 | deps = MICROKERNEL_TEST_DEPS, |
| 2470 | ) |
| 2471 | |
| 2472 | xnnpack_unit_test( |
| 2473 | name = "f32_conv_hwc2spchw_test", |
| 2474 | srcs = [ |
| 2475 | "test/f32-conv-hwc2spchw.cc", |
| 2476 | "test/conv-hwc2spchw-microkernel-tester.h", |
| 2477 | "src/xnnpack/AlignedAllocator.h", |
| 2478 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2479 | deps = MICROKERNEL_TEST_DEPS, |
| 2480 | ) |
| 2481 | |
| 2482 | xnnpack_unit_test( |
| 2483 | name = "f32_dwconv_test", |
| 2484 | srcs = [ |
| 2485 | "test/f32-dwconv.cc", |
| 2486 | "test/dwconv-microkernel-tester.h", |
| 2487 | "src/xnnpack/AlignedAllocator.h", |
| 2488 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2489 | deps = MICROKERNEL_TEST_DEPS, |
| 2490 | ) |
| 2491 | |
| 2492 | xnnpack_unit_test( |
| 2493 | name = "f32_dwconv_spchw_test", |
| 2494 | srcs = [ |
| 2495 | "test/f32-dwconv-spchw.cc", |
| 2496 | "test/dwconv-spchw-microkernel-tester.h", |
| 2497 | "src/xnnpack/AlignedAllocator.h", |
| 2498 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2499 | deps = MICROKERNEL_TEST_DEPS, |
| 2500 | ) |
| 2501 | |
| 2502 | xnnpack_unit_test( |
| 2503 | name = "f32_gavgpool_test", |
| 2504 | srcs = [ |
| 2505 | "test/f32-gavgpool.cc", |
| 2506 | "test/gavgpool-microkernel-tester.h", |
| 2507 | "src/xnnpack/AlignedAllocator.h", |
| 2508 | ] + MICROKERNEL_TEST_HDRS, |
| 2509 | deps = MICROKERNEL_TEST_DEPS, |
| 2510 | ) |
| 2511 | |
| 2512 | xnnpack_unit_test( |
| 2513 | name = "f32_gavgpool_spchw_test", |
| 2514 | srcs = [ |
| 2515 | "test/f32-gavgpool-spchw.cc", |
| 2516 | "test/gavgpool-spchw-microkernel-tester.h", |
| 2517 | "src/xnnpack/AlignedAllocator.h", |
| 2518 | ] + MICROKERNEL_TEST_HDRS, |
| 2519 | deps = MICROKERNEL_TEST_DEPS, |
| 2520 | ) |
| 2521 | |
| 2522 | xnnpack_unit_test( |
| 2523 | name = "f32_gemm_test", |
| 2524 | srcs = [ |
| 2525 | "test/f32-gemm.cc", |
| 2526 | "test/gemm-microkernel-tester.h", |
| 2527 | "src/xnnpack/AlignedAllocator.h", |
| 2528 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2529 | deps = MICROKERNEL_TEST_DEPS, |
| 2530 | ) |
| 2531 | |
| 2532 | xnnpack_unit_test( |
| 2533 | name = "f32_gemminc_test", |
| 2534 | srcs = [ |
| 2535 | "test/f32-gemminc.cc", |
| 2536 | "test/gemm-microkernel-tester.h", |
| 2537 | "src/xnnpack/AlignedAllocator.h", |
| 2538 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2539 | deps = MICROKERNEL_TEST_DEPS, |
| 2540 | ) |
| 2541 | |
| 2542 | xnnpack_unit_test( |
| 2543 | name = "f32_hswish_test", |
| 2544 | srcs = [ |
| 2545 | "test/f32-hswish.cc", |
| 2546 | "test/hswish-microkernel-tester.h", |
| 2547 | ] + MICROKERNEL_TEST_HDRS, |
| 2548 | deps = MICROKERNEL_TEST_DEPS, |
| 2549 | ) |
| 2550 | |
| 2551 | xnnpack_unit_test( |
| 2552 | name = "f32_maxpool_test", |
| 2553 | srcs = [ |
| 2554 | "test/f32-maxpool.cc", |
| 2555 | "test/maxpool-microkernel-tester.h", |
| 2556 | ] + MICROKERNEL_TEST_HDRS, |
| 2557 | deps = MICROKERNEL_TEST_DEPS, |
| 2558 | ) |
| 2559 | |
| 2560 | xnnpack_unit_test( |
| 2561 | name = "f32_pavgpool_test", |
| 2562 | srcs = [ |
| 2563 | "test/f32-pavgpool.cc", |
| 2564 | "test/avgpool-microkernel-tester.h", |
| 2565 | "src/xnnpack/AlignedAllocator.h", |
| 2566 | ] + MICROKERNEL_TEST_HDRS, |
| 2567 | deps = MICROKERNEL_TEST_DEPS, |
| 2568 | ) |
| 2569 | |
| 2570 | xnnpack_unit_test( |
| 2571 | name = "f32_ppmm_test", |
| 2572 | srcs = [ |
| 2573 | "test/f32-ppmm.cc", |
| 2574 | "test/gemm-microkernel-tester.h", |
| 2575 | "src/xnnpack/AlignedAllocator.h", |
| 2576 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2577 | deps = MICROKERNEL_TEST_DEPS, |
| 2578 | ) |
| 2579 | |
| 2580 | xnnpack_unit_test( |
| 2581 | name = "f32_prelu_test", |
| 2582 | srcs = [ |
| 2583 | "test/f32-prelu.cc", |
| 2584 | "test/prelu-microkernel-tester.h", |
| 2585 | "src/xnnpack/AlignedAllocator.h", |
| 2586 | ] + MICROKERNEL_TEST_HDRS, |
| 2587 | deps = MICROKERNEL_TEST_DEPS, |
| 2588 | ) |
| 2589 | |
| 2590 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2591 | name = "f32_raddexpminusmax_test", |
| 2592 | srcs = [ |
| 2593 | "test/f32-raddexpminusmax.cc", |
| 2594 | "test/raddexpminusmax-microkernel-tester.h", |
| 2595 | ] + MICROKERNEL_TEST_HDRS, |
| 2596 | deps = MICROKERNEL_TEST_DEPS, |
| 2597 | ) |
| 2598 | |
| 2599 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2600 | name = "f32_raddextexp_test", |
| 2601 | srcs = [ |
| 2602 | "test/f32-raddextexp.cc", |
| 2603 | "test/raddextexp-microkernel-tester.h", |
| 2604 | ] + MICROKERNEL_TEST_HDRS, |
| 2605 | deps = MICROKERNEL_TEST_DEPS, |
| 2606 | ) |
| 2607 | |
| 2608 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2609 | name = "f32_raddstoreexpminusmax_test", |
| 2610 | srcs = [ |
| 2611 | "test/f32-raddstoreexpminusmax.cc", |
| 2612 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 2613 | ] + MICROKERNEL_TEST_HDRS, |
| 2614 | deps = MICROKERNEL_TEST_DEPS, |
| 2615 | ) |
| 2616 | |
| 2617 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2618 | name = "f32_rmax_test", |
| 2619 | srcs = [ |
| 2620 | "test/f32-rmax.cc", |
| 2621 | "test/rmax-microkernel-tester.h", |
| 2622 | ] + MICROKERNEL_TEST_HDRS, |
| 2623 | deps = MICROKERNEL_TEST_DEPS, |
| 2624 | ) |
| 2625 | |
| 2626 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2627 | name = "f32_sigmoid_test", |
| 2628 | srcs = [ |
| 2629 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2630 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 2631 | ] + MICROKERNEL_TEST_HDRS, |
| 2632 | deps = MICROKERNEL_TEST_DEPS, |
| 2633 | ) |
| 2634 | |
| 2635 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2636 | name = "f32_spmm_test", |
| 2637 | srcs = [ |
| 2638 | "test/f32-spmm.cc", |
| 2639 | "test/spmm-microkernel-tester.h", |
| 2640 | "src/xnnpack/AlignedAllocator.h", |
| 2641 | ] + MICROKERNEL_TEST_HDRS, |
| 2642 | deps = MICROKERNEL_TEST_DEPS, |
| 2643 | ) |
| 2644 | |
| 2645 | xnnpack_unit_test( |
| 2646 | name = "f32_vadd_test", |
| 2647 | srcs = [ |
| 2648 | "test/f32-vadd.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2649 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2650 | ] + MICROKERNEL_TEST_HDRS, |
| 2651 | deps = MICROKERNEL_TEST_DEPS, |
| 2652 | ) |
| 2653 | |
| 2654 | xnnpack_unit_test( |
| 2655 | name = "f32_vaddc_test", |
| 2656 | srcs = [ |
| 2657 | "test/f32-vaddc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2658 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2659 | ] + MICROKERNEL_TEST_HDRS, |
| 2660 | deps = MICROKERNEL_TEST_DEPS, |
| 2661 | ) |
| 2662 | |
| 2663 | xnnpack_unit_test( |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 2664 | name = "f32_vdiv_test", |
| 2665 | srcs = [ |
| 2666 | "test/f32-vdiv.cc", |
| 2667 | "test/vbinary-microkernel-tester.h", |
| 2668 | ] + MICROKERNEL_TEST_HDRS, |
| 2669 | deps = MICROKERNEL_TEST_DEPS, |
| 2670 | ) |
| 2671 | |
| 2672 | xnnpack_unit_test( |
| 2673 | name = "f32_vdivc_test", |
| 2674 | srcs = [ |
| 2675 | "test/f32-vdivc.cc", |
| 2676 | "test/vbinaryc-microkernel-tester.h", |
| 2677 | ] + MICROKERNEL_TEST_HDRS, |
| 2678 | deps = MICROKERNEL_TEST_DEPS, |
| 2679 | ) |
| 2680 | |
| 2681 | xnnpack_unit_test( |
| 2682 | name = "f32_vrdivc_test", |
| 2683 | srcs = [ |
| 2684 | "test/f32-vrdivc.cc", |
| 2685 | "test/vbinaryc-microkernel-tester.h", |
| 2686 | ] + MICROKERNEL_TEST_HDRS, |
| 2687 | deps = MICROKERNEL_TEST_DEPS, |
| 2688 | ) |
| 2689 | |
| 2690 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 2691 | name = "f32_vmax_test", |
| 2692 | srcs = [ |
| 2693 | "test/f32-vmax.cc", |
| 2694 | "test/vbinary-microkernel-tester.h", |
| 2695 | ] + MICROKERNEL_TEST_HDRS, |
| 2696 | deps = MICROKERNEL_TEST_DEPS, |
| 2697 | ) |
| 2698 | |
| 2699 | xnnpack_unit_test( |
| 2700 | name = "f32_vmaxc_test", |
| 2701 | srcs = [ |
| 2702 | "test/f32-vmaxc.cc", |
| 2703 | "test/vbinaryc-microkernel-tester.h", |
| 2704 | ] + MICROKERNEL_TEST_HDRS, |
| 2705 | deps = MICROKERNEL_TEST_DEPS, |
| 2706 | ) |
| 2707 | |
| 2708 | xnnpack_unit_test( |
| 2709 | name = "f32_vmin_test", |
| 2710 | srcs = [ |
| 2711 | "test/f32-vmin.cc", |
| 2712 | "test/vbinary-microkernel-tester.h", |
| 2713 | ] + MICROKERNEL_TEST_HDRS, |
| 2714 | deps = MICROKERNEL_TEST_DEPS, |
| 2715 | ) |
| 2716 | |
| 2717 | xnnpack_unit_test( |
| 2718 | name = "f32_vminc_test", |
| 2719 | srcs = [ |
| 2720 | "test/f32-vminc.cc", |
| 2721 | "test/vbinaryc-microkernel-tester.h", |
| 2722 | ] + MICROKERNEL_TEST_HDRS, |
| 2723 | deps = MICROKERNEL_TEST_DEPS, |
| 2724 | ) |
| 2725 | |
| 2726 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2727 | name = "f32_vmul_test", |
| 2728 | srcs = [ |
| 2729 | "test/f32-vmul.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2730 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2731 | ] + MICROKERNEL_TEST_HDRS, |
| 2732 | deps = MICROKERNEL_TEST_DEPS, |
| 2733 | ) |
| 2734 | |
| 2735 | xnnpack_unit_test( |
| 2736 | name = "f32_vmulc_test", |
| 2737 | srcs = [ |
| 2738 | "test/f32-vmulc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2739 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2740 | ] + MICROKERNEL_TEST_HDRS, |
| 2741 | deps = MICROKERNEL_TEST_DEPS, |
| 2742 | ) |
| 2743 | |
| 2744 | xnnpack_unit_test( |
| 2745 | name = "f32_vmulcaddc_test", |
| 2746 | srcs = [ |
| 2747 | "test/f32-vmulcaddc.cc", |
| 2748 | "test/vmulcaddc-microkernel-tester.h", |
| 2749 | "src/xnnpack/AlignedAllocator.h", |
| 2750 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2751 | deps = MICROKERNEL_TEST_DEPS, |
| 2752 | ) |
| 2753 | |
| 2754 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2755 | name = "f32_vscale_test", |
| 2756 | srcs = [ |
| 2757 | "test/f32-vscale.cc", |
| 2758 | "test/vscale-microkernel-tester.h", |
| 2759 | ] + MICROKERNEL_TEST_HDRS, |
| 2760 | deps = MICROKERNEL_TEST_DEPS, |
| 2761 | ) |
| 2762 | |
| 2763 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2764 | name = "f32_vscaleexpminusmax_test", |
| 2765 | srcs = [ |
| 2766 | "test/f32-vscaleexpminusmax.cc", |
| 2767 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 2768 | ] + MICROKERNEL_TEST_HDRS, |
| 2769 | deps = MICROKERNEL_TEST_DEPS, |
| 2770 | ) |
| 2771 | |
| 2772 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2773 | name = "f32_vscaleextexp_test", |
| 2774 | srcs = [ |
| 2775 | "test/f32-vscaleextexp.cc", |
| 2776 | "test/vscaleextexp-microkernel-tester.h", |
| 2777 | ] + MICROKERNEL_TEST_HDRS, |
| 2778 | deps = MICROKERNEL_TEST_DEPS, |
| 2779 | ) |
| 2780 | |
| 2781 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2782 | name = "f32_vsub_test", |
| 2783 | srcs = [ |
| 2784 | "test/f32-vsub.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2785 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2786 | ] + MICROKERNEL_TEST_HDRS, |
| 2787 | deps = MICROKERNEL_TEST_DEPS, |
| 2788 | ) |
| 2789 | |
| 2790 | xnnpack_unit_test( |
| 2791 | name = "f32_vsubc_test", |
| 2792 | srcs = [ |
| 2793 | "test/f32-vsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2794 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 2795 | ] + MICROKERNEL_TEST_HDRS, |
| 2796 | deps = MICROKERNEL_TEST_DEPS, |
| 2797 | ) |
| 2798 | |
| 2799 | xnnpack_unit_test( |
| 2800 | name = "f32_vrsubc_test", |
| 2801 | srcs = [ |
| 2802 | "test/f32-vrsubc.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2803 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2804 | ] + MICROKERNEL_TEST_HDRS, |
| 2805 | deps = MICROKERNEL_TEST_DEPS, |
| 2806 | ) |
| 2807 | |
| 2808 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2809 | name = "q8_avgpool_test", |
| 2810 | srcs = [ |
| 2811 | "test/q8-avgpool.cc", |
| 2812 | "test/avgpool-microkernel-tester.h", |
| 2813 | "src/xnnpack/AlignedAllocator.h", |
| 2814 | ] + MICROKERNEL_TEST_HDRS, |
| 2815 | deps = MICROKERNEL_TEST_DEPS, |
| 2816 | ) |
| 2817 | |
| 2818 | xnnpack_unit_test( |
| 2819 | name = "q8_igemm_test", |
| 2820 | srcs = [ |
| 2821 | "test/q8-igemm.cc", |
| 2822 | "test/gemm-microkernel-tester.h", |
| 2823 | "src/xnnpack/AlignedAllocator.h", |
| 2824 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2825 | deps = MICROKERNEL_TEST_DEPS, |
| 2826 | ) |
| 2827 | |
| 2828 | xnnpack_unit_test( |
| 2829 | name = "q8_dwconv_test", |
| 2830 | srcs = [ |
| 2831 | "test/q8-dwconv.cc", |
| 2832 | "test/dwconv-microkernel-tester.h", |
| 2833 | "src/xnnpack/AlignedAllocator.h", |
| 2834 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2835 | deps = MICROKERNEL_TEST_DEPS, |
| 2836 | ) |
| 2837 | |
| 2838 | xnnpack_unit_test( |
| 2839 | name = "q8_gavgpool_test", |
| 2840 | srcs = [ |
| 2841 | "test/q8-gavgpool.cc", |
| 2842 | "test/gavgpool-microkernel-tester.h", |
| 2843 | "src/xnnpack/AlignedAllocator.h", |
| 2844 | ] + MICROKERNEL_TEST_HDRS, |
| 2845 | deps = MICROKERNEL_TEST_DEPS, |
| 2846 | ) |
| 2847 | |
| 2848 | xnnpack_unit_test( |
| 2849 | name = "q8_gemm_test", |
| 2850 | srcs = [ |
| 2851 | "test/q8-gemm.cc", |
| 2852 | "test/gemm-microkernel-tester.h", |
| 2853 | "src/xnnpack/AlignedAllocator.h", |
| 2854 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 2855 | deps = MICROKERNEL_TEST_DEPS, |
| 2856 | ) |
| 2857 | |
| 2858 | xnnpack_unit_test( |
| 2859 | name = "q8_vadd_test", |
| 2860 | srcs = [ |
| 2861 | "test/q8-vadd.cc", |
| 2862 | "test/vadd-microkernel-tester.h", |
| 2863 | ] + MICROKERNEL_TEST_HDRS, |
| 2864 | deps = MICROKERNEL_TEST_DEPS, |
| 2865 | ) |
| 2866 | |
| 2867 | xnnpack_unit_test( |
| 2868 | name = "u8_clamp_test", |
| 2869 | srcs = [ |
| 2870 | "test/u8-clamp.cc", |
| 2871 | "test/clamp-microkernel-tester.h", |
| 2872 | ] + MICROKERNEL_TEST_HDRS, |
| 2873 | deps = MICROKERNEL_TEST_DEPS, |
| 2874 | ) |
| 2875 | |
| 2876 | xnnpack_unit_test( |
| 2877 | name = "u8_lut32norm_test", |
| 2878 | srcs = [ |
| 2879 | "test/u8-lut32norm.cc", |
| 2880 | "test/lut-norm-microkernel-tester.h", |
| 2881 | ] + MICROKERNEL_TEST_HDRS, |
| 2882 | deps = MICROKERNEL_TEST_DEPS, |
| 2883 | ) |
| 2884 | |
| 2885 | xnnpack_unit_test( |
| 2886 | name = "u8_maxpool_test", |
| 2887 | srcs = [ |
| 2888 | "test/u8-maxpool.cc", |
| 2889 | "test/maxpool-microkernel-tester.h", |
| 2890 | ] + MICROKERNEL_TEST_HDRS, |
| 2891 | deps = MICROKERNEL_TEST_DEPS, |
| 2892 | ) |
| 2893 | |
| 2894 | xnnpack_unit_test( |
| 2895 | name = "u8_rmax_test", |
| 2896 | srcs = [ |
| 2897 | "test/u8-rmax.cc", |
| 2898 | "test/rmax-microkernel-tester.h", |
| 2899 | ] + MICROKERNEL_TEST_HDRS, |
| 2900 | deps = MICROKERNEL_TEST_DEPS, |
| 2901 | ) |
| 2902 | |
| 2903 | xnnpack_unit_test( |
| 2904 | name = "x32_packx_test", |
| 2905 | srcs = [ |
| 2906 | "test/x32-packx.cc", |
| 2907 | "test/pack-microkernel-tester.h", |
| 2908 | "src/xnnpack/AlignedAllocator.h", |
| 2909 | ] + MICROKERNEL_TEST_HDRS, |
| 2910 | deps = MICROKERNEL_TEST_DEPS, |
| 2911 | ) |
| 2912 | |
| 2913 | xnnpack_unit_test( |
| 2914 | name = "x32_pad_test", |
| 2915 | srcs = [ |
| 2916 | "test/x32-pad.cc", |
| 2917 | "test/pad-microkernel-tester.h", |
| 2918 | ] + MICROKERNEL_TEST_HDRS, |
| 2919 | deps = MICROKERNEL_TEST_DEPS, |
| 2920 | ) |
| 2921 | |
| 2922 | xnnpack_unit_test( |
| 2923 | name = "x32_unpool_test", |
| 2924 | srcs = [ |
| 2925 | "test/x32-unpool.cc", |
| 2926 | "test/unpool-microkernel-tester.h", |
| 2927 | ] + MICROKERNEL_TEST_HDRS, |
| 2928 | deps = MICROKERNEL_TEST_DEPS, |
| 2929 | ) |
| 2930 | |
| 2931 | xnnpack_unit_test( |
| 2932 | name = "x32_zip_test", |
| 2933 | srcs = [ |
| 2934 | "test/x32-zip.cc", |
| 2935 | "test/zip-microkernel-tester.h", |
| 2936 | ] + MICROKERNEL_TEST_HDRS, |
| 2937 | deps = MICROKERNEL_TEST_DEPS, |
| 2938 | ) |
| 2939 | |
| 2940 | xnnpack_unit_test( |
| 2941 | name = "x8_lut_test", |
| 2942 | srcs = [ |
| 2943 | "test/x8-lut.cc", |
| 2944 | "test/lut-microkernel-tester.h", |
| 2945 | ] + MICROKERNEL_TEST_HDRS, |
| 2946 | deps = MICROKERNEL_TEST_DEPS, |
| 2947 | ) |
| 2948 | |
| 2949 | xnnpack_unit_test( |
| 2950 | name = "x8_zip_test", |
| 2951 | srcs = [ |
| 2952 | "test/x8-zip.cc", |
| 2953 | "test/zip-microkernel-tester.h", |
| 2954 | ] + MICROKERNEL_TEST_HDRS, |
| 2955 | deps = MICROKERNEL_TEST_DEPS, |
| 2956 | ) |
| 2957 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2958 | xnnpack_unit_test( |
| 2959 | name = "requantization_test", |
| 2960 | srcs = [ |
| 2961 | "src/xnnpack/requantization-stubs.h", |
| 2962 | "test/requantization.cc", |
| 2963 | "test/requantization-tester.h", |
| 2964 | ] + MICROKERNEL_TEST_HDRS, |
| 2965 | deps = MICROKERNEL_TEST_DEPS, |
| 2966 | ) |
| 2967 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 2968 | ########################## Size tests for the library ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2969 | |
| 2970 | xnnpack_binary( |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 2971 | name = "operator_size_test", |
| 2972 | srcs = ["test/operator-size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2973 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2974 | ) |
| 2975 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 2976 | xnnpack_binary( |
| 2977 | name = "subgraph_size_test", |
| 2978 | srcs = ["test/subgraph-size.c"], |
| 2979 | deps = [":XNNPACK"], |
| 2980 | ) |
| 2981 | |
| 2982 | ########################### Unit tests for operators ########################## |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2983 | |
| 2984 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2985 | name = "add_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2986 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2987 | "test/add-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2988 | "test/add-operator-tester.h", |
| 2989 | ], |
| 2990 | deps = OPERATOR_TEST_DEPS, |
| 2991 | ) |
| 2992 | |
| 2993 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2994 | name = "add_nd_test", |
| 2995 | srcs = [ |
| 2996 | "test/add-nd.cc", |
| 2997 | "test/binary-elementwise-operator-tester.h", |
| 2998 | ], |
| 2999 | deps = OPERATOR_TEST_DEPS, |
| 3000 | ) |
| 3001 | |
| 3002 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3003 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3004 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3005 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3006 | "test/argmax-pooling-operator-tester.h", |
| 3007 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3008 | deps = OPERATOR_TEST_DEPS, |
| 3009 | ) |
| 3010 | |
| 3011 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3012 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3013 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3014 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3015 | "test/average-pooling-operator-tester.h", |
| 3016 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3017 | deps = OPERATOR_TEST_DEPS, |
| 3018 | ) |
| 3019 | |
| 3020 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3021 | name = "channel_pad_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3022 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3023 | "test/channel-pad-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3024 | "test/channel-pad-operator-tester.h", |
| 3025 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3026 | deps = OPERATOR_TEST_DEPS, |
| 3027 | ) |
| 3028 | |
| 3029 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3030 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3031 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3032 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3033 | "test/channel-shuffle-operator-tester.h", |
| 3034 | ], |
| 3035 | deps = OPERATOR_TEST_DEPS, |
| 3036 | ) |
| 3037 | |
| 3038 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3039 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3040 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3041 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3042 | "test/clamp-operator-tester.h", |
| 3043 | ], |
| 3044 | deps = OPERATOR_TEST_DEPS, |
| 3045 | ) |
| 3046 | |
| 3047 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3048 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3049 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3050 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3051 | "test/convolution-operator-tester.h", |
| 3052 | ], |
| 3053 | deps = OPERATOR_TEST_DEPS, |
| 3054 | ) |
| 3055 | |
| 3056 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3057 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3058 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3059 | "test/convolution-nchw.cc", |
| 3060 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3061 | ], |
| 3062 | deps = OPERATOR_TEST_DEPS, |
| 3063 | ) |
| 3064 | |
| 3065 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3066 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3067 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3068 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3069 | "test/deconvolution-operator-tester.h", |
| 3070 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3071 | deps = OPERATOR_TEST_DEPS, |
| 3072 | ) |
| 3073 | |
| 3074 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 3075 | name = "divide_nd_test", |
| 3076 | srcs = [ |
| 3077 | "test/binary-elementwise-operator-tester.h", |
| 3078 | "test/divide-nd.cc", |
| 3079 | ], |
| 3080 | deps = OPERATOR_TEST_DEPS, |
| 3081 | ) |
| 3082 | |
| 3083 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3084 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3085 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3086 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3087 | "test/fully-connected-operator-tester.h", |
| 3088 | ], |
| 3089 | deps = OPERATOR_TEST_DEPS, |
| 3090 | ) |
| 3091 | |
| 3092 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3093 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3094 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3095 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3096 | "test/global-average-pooling-operator-tester.h", |
| 3097 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3098 | deps = OPERATOR_TEST_DEPS, |
| 3099 | ) |
| 3100 | |
| 3101 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3102 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3103 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3104 | "test/global-average-pooling-ncw.cc", |
| 3105 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3106 | ], |
| 3107 | deps = OPERATOR_TEST_DEPS, |
| 3108 | ) |
| 3109 | |
| 3110 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3111 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3112 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3113 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3114 | "test/hardswish-operator-tester.h", |
| 3115 | ], |
| 3116 | deps = OPERATOR_TEST_DEPS, |
| 3117 | ) |
| 3118 | |
| 3119 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3120 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3121 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3122 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3123 | "test/leaky-relu-operator-tester.h", |
| 3124 | ], |
| 3125 | deps = OPERATOR_TEST_DEPS, |
| 3126 | ) |
| 3127 | |
| 3128 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3129 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3130 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3131 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3132 | "test/max-pooling-operator-tester.h", |
| 3133 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3134 | deps = OPERATOR_TEST_DEPS, |
| 3135 | ) |
| 3136 | |
| 3137 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 3138 | name = "maximum_nd_test", |
| 3139 | srcs = [ |
| 3140 | "test/binary-elementwise-operator-tester.h", |
| 3141 | "test/maximum-nd.cc", |
| 3142 | ], |
| 3143 | deps = OPERATOR_TEST_DEPS, |
| 3144 | ) |
| 3145 | |
| 3146 | xnnpack_unit_test( |
| 3147 | name = "minimum_nd_test", |
| 3148 | srcs = [ |
| 3149 | "test/binary-elementwise-operator-tester.h", |
| 3150 | "test/minimum-nd.cc", |
| 3151 | ], |
| 3152 | deps = OPERATOR_TEST_DEPS, |
| 3153 | ) |
| 3154 | |
| 3155 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3156 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 3157 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 3158 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3159 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 3160 | ], |
| 3161 | deps = OPERATOR_TEST_DEPS, |
| 3162 | ) |
| 3163 | |
| 3164 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3165 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3166 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3167 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3168 | "test/prelu-operator-tester.h", |
| 3169 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3170 | deps = OPERATOR_TEST_DEPS, |
| 3171 | ) |
| 3172 | |
| 3173 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3174 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 3175 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3176 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 3177 | "test/resize-bilinear-operator-tester.h", |
| 3178 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 3179 | deps = OPERATOR_TEST_DEPS, |
| 3180 | ) |
| 3181 | |
| 3182 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3183 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3184 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3185 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3186 | "test/sigmoid-operator-tester.h", |
| 3187 | ], |
| 3188 | deps = OPERATOR_TEST_DEPS, |
| 3189 | ) |
| 3190 | |
| 3191 | xnnpack_unit_test( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3192 | name = "softmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3193 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3194 | "test/softmax-nc.cc", |
| 3195 | "test/softmax-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3196 | ], |
| 3197 | deps = OPERATOR_TEST_DEPS, |
| 3198 | ) |
| 3199 | |
| 3200 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 3201 | name = "subtract_nd_test", |
| 3202 | srcs = [ |
| 3203 | "test/binary-elementwise-operator-tester.h", |
| 3204 | "test/subtract-nd.cc", |
| 3205 | ], |
| 3206 | deps = OPERATOR_TEST_DEPS, |
| 3207 | ) |
| 3208 | |
| 3209 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3210 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3211 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3212 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3213 | "test/unpooling-operator-tester.h", |
| 3214 | ], |
| 3215 | deps = OPERATOR_TEST_DEPS, |
| 3216 | ) |
| 3217 | |
| 3218 | ############################# Build configurations ############################# |
| 3219 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3220 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3221 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3222 | name = "xnn_enable_assembly_explicit_true", |
| 3223 | define_values = {"xnn_enable_assembly": "true"}, |
| 3224 | ) |
| 3225 | |
| 3226 | # Disables usage of assembly kernels. |
| 3227 | config_setting( |
| 3228 | name = "xnn_enable_assembly_explicit_false", |
| 3229 | define_values = {"xnn_enable_assembly": "false"}, |
| 3230 | ) |
| 3231 | |
| 3232 | # Builds with -c dbg |
| 3233 | config_setting( |
| 3234 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3235 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3236 | "compilation_mode": "dbg", |
| 3237 | }, |
| 3238 | ) |
| 3239 | |
| 3240 | # Builds with -c opt |
| 3241 | config_setting( |
| 3242 | name = "optimized_build", |
| 3243 | values = { |
| 3244 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3245 | }, |
| 3246 | ) |
| 3247 | |
| 3248 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3249 | name = "linux_k8", |
| 3250 | values = {"cpu": "k8"}, |
| 3251 | ) |
| 3252 | |
| 3253 | config_setting( |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 3254 | name = "linux_aarch64", |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 3255 | values = {"cpu": "aarch64"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 3256 | ) |
| 3257 | |
| 3258 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3259 | name = "android", |
| 3260 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 3261 | ) |
| 3262 | |
| 3263 | config_setting( |
| 3264 | name = "android_armv7", |
| 3265 | values = { |
| 3266 | "crosstool_top": "//external:android/crosstool", |
| 3267 | "cpu": "armeabi-v7a", |
| 3268 | }, |
| 3269 | ) |
| 3270 | |
| 3271 | config_setting( |
| 3272 | name = "android_arm64", |
| 3273 | values = { |
| 3274 | "crosstool_top": "//external:android/crosstool", |
| 3275 | "cpu": "arm64-v8a", |
| 3276 | }, |
| 3277 | ) |
| 3278 | |
| 3279 | config_setting( |
| 3280 | name = "android_x86", |
| 3281 | values = { |
| 3282 | "crosstool_top": "//external:android/crosstool", |
| 3283 | "cpu": "x86", |
| 3284 | }, |
| 3285 | ) |
| 3286 | |
| 3287 | config_setting( |
| 3288 | name = "android_x86_64", |
| 3289 | values = { |
| 3290 | "crosstool_top": "//external:android/crosstool", |
| 3291 | "cpu": "x86_64", |
| 3292 | }, |
| 3293 | ) |
| 3294 | |
| 3295 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 3296 | name = "macos_x86_64", |
| 3297 | values = { |
| 3298 | "apple_platform_type": "macos", |
| 3299 | "cpu": "darwin", |
| 3300 | }, |
| 3301 | ) |
| 3302 | |
| 3303 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3304 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3305 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3306 | ) |
| 3307 | |
| 3308 | config_setting( |
| 3309 | name = "emscripten_wasm", |
| 3310 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3311 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3312 | "cpu": "wasm", |
| 3313 | }, |
| 3314 | ) |
| 3315 | |
| 3316 | config_setting( |
| 3317 | name = "emscripten_wasmsimd", |
| 3318 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3319 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3320 | "cpu": "wasm", |
Marat Dukhan | 8c19e3c | 2019-10-30 12:14:58 -0700 | [diff] [blame] | 3321 | "features": "wasm_simd", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3322 | }, |
| 3323 | ) |
| 3324 | |
| 3325 | config_setting( |
| 3326 | name = "emscripten_asmjs", |
| 3327 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 3328 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3329 | "cpu": "asmjs", |
| 3330 | }, |
| 3331 | ) |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3332 | |
| 3333 | config_setting( |
| 3334 | name = "ios_armv7", |
| 3335 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3336 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3337 | "cpu": "ios_armv7", |
| 3338 | }, |
| 3339 | ) |
| 3340 | |
| 3341 | config_setting( |
| 3342 | name = "ios_arm64", |
| 3343 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3344 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3345 | "cpu": "ios_arm64", |
| 3346 | }, |
| 3347 | ) |
| 3348 | |
| 3349 | config_setting( |
| 3350 | name = "ios_arm64e", |
| 3351 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3352 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3353 | "cpu": "ios_arm64e", |
| 3354 | }, |
| 3355 | ) |
| 3356 | |
| 3357 | config_setting( |
| 3358 | name = "ios_x86", |
| 3359 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3360 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3361 | "cpu": "ios_i386", |
| 3362 | }, |
| 3363 | ) |
| 3364 | |
| 3365 | config_setting( |
| 3366 | name = "ios_x86_64", |
| 3367 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3368 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3369 | "cpu": "ios_x86_64", |
| 3370 | }, |
| 3371 | ) |
| 3372 | |
| 3373 | config_setting( |
| 3374 | name = "watchos_armv7k", |
| 3375 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3376 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3377 | "cpu": "watchos_armv7k", |
| 3378 | }, |
| 3379 | ) |
| 3380 | |
| 3381 | config_setting( |
| 3382 | name = "watchos_arm64_32", |
| 3383 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3384 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3385 | "cpu": "watchos_arm64_32", |
| 3386 | }, |
| 3387 | ) |
| 3388 | |
| 3389 | config_setting( |
| 3390 | name = "watchos_x86", |
| 3391 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3392 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3393 | "cpu": "watchos_i386", |
| 3394 | }, |
| 3395 | ) |
| 3396 | |
| 3397 | config_setting( |
| 3398 | name = "watchos_x86_64", |
| 3399 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3400 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3401 | "cpu": "watchos_x86_64", |
| 3402 | }, |
| 3403 | ) |
| 3404 | |
| 3405 | config_setting( |
| 3406 | name = "tvos_arm64", |
| 3407 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3408 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3409 | "cpu": "tvos_arm64", |
| 3410 | }, |
| 3411 | ) |
| 3412 | |
| 3413 | config_setting( |
| 3414 | name = "tvos_x86_64", |
| 3415 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 3416 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 3417 | "cpu": "tvos_x86_64", |
| 3418 | }, |
| 3419 | ) |