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 | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 9 | load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_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_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 | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 15 | OPERATOR_BENCHMARK_DEPS = [ |
| 16 | ":XNNPACK", |
| 17 | ":bench_utils", |
| 18 | "@cpuinfo", |
Frank Barchard | 0c84973 | 2020-06-12 13:31:32 -0700 | [diff] [blame] | 19 | "@FP16", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 20 | "@pthreadpool", |
| 21 | ] |
| 22 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 23 | MICROKERNEL_BENCHMARK_DEPS = [ |
| 24 | ":ukernels", |
| 25 | ":bench_utils", |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 26 | ":enable_assembly", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 27 | "@cpuinfo", |
| 28 | "@FP16", |
| 29 | "@pthreadpool", |
| 30 | ] |
| 31 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 32 | ACCURACY_EVAL_DEPS = [ |
| 33 | ":XNNPACK", |
| 34 | ":ukernels", |
| 35 | "@FP16", |
| 36 | "@pthreadpool", |
| 37 | ] |
| 38 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 39 | MICROKERNEL_TEST_DEPS = [ |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 40 | ":ukernels_test_mode", |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 41 | ":enable_assembly", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 42 | "@cpuinfo", |
| 43 | "@FP16", |
| 44 | "@pthreadpool", |
| 45 | ] |
| 46 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 47 | OPERATOR_TEST_DEPS = [ |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 48 | ":XNNPACK_test_mode", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 49 | "@pthreadpool", |
| 50 | "@FP16", |
| 51 | ] |
| 52 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 53 | OPERATOR_SRCS = [ |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 54 | "src/operators/argmax-pooling-nhwc.c", |
| 55 | "src/operators/average-pooling-nhwc.c", |
| 56 | "src/operators/binary-elementwise-nd.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 57 | "src/operators/channel-shuffle-nc.c", |
Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 58 | "src/operators/constant-pad-nd.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 59 | "src/operators/convolution-nchw.c", |
| 60 | "src/operators/convolution-nhwc.c", |
| 61 | "src/operators/deconvolution-nhwc.c", |
| 62 | "src/operators/fully-connected-nc.c", |
| 63 | "src/operators/global-average-pooling-ncw.c", |
| 64 | "src/operators/global-average-pooling-nwc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 65 | "src/operators/leaky-relu-nc.c", |
| 66 | "src/operators/max-pooling-nhwc.c", |
| 67 | "src/operators/prelu-nc.c", |
| 68 | "src/operators/resize-bilinear-nhwc.c", |
| 69 | "src/operators/sigmoid-nc.c", |
| 70 | "src/operators/softmax-nc.c", |
Marat Dukhan | c3065f5 | 2020-06-04 13:33:32 -0700 | [diff] [blame] | 71 | "src/operators/unary-elementwise-nc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 72 | "src/operators/unpooling-nhwc.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 73 | ] |
| 74 | |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 75 | SUBGRAPH_SRCS = [ |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 76 | "src/subgraph/abs.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 77 | "src/subgraph/add2.c", |
| 78 | "src/subgraph/argmax-pooling-2d.c", |
| 79 | "src/subgraph/average-pooling-2d.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 80 | "src/subgraph/bankers-rounding.c", |
| 81 | "src/subgraph/ceiling.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 82 | "src/subgraph/clamp.c", |
| 83 | "src/subgraph/convolution-2d.c", |
| 84 | "src/subgraph/deconvolution-2d.c", |
| 85 | "src/subgraph/depthwise-convolution-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 86 | "src/subgraph/divide.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 87 | "src/subgraph/fully-connected.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 88 | "src/subgraph/floor.c", |
Marat Dukhan | a059b7d | 2020-06-11 11:41:27 -0700 | [diff] [blame] | 89 | "src/subgraph/global-average-pooling-2d.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 90 | "src/subgraph/hardswish.c", |
Marat Dukhan | 5bbebac | 2020-06-10 19:42:15 -0700 | [diff] [blame] | 91 | "src/subgraph/leaky-relu.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 92 | "src/subgraph/max-pooling-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 93 | "src/subgraph/maximum2.c", |
| 94 | "src/subgraph/minimum2.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 95 | "src/subgraph/multiply2.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 96 | "src/subgraph/negate.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 97 | "src/subgraph/prelu.c", |
| 98 | "src/subgraph/sigmoid.c", |
| 99 | "src/subgraph/softmax.c", |
| 100 | "src/subgraph/static-constant-pad.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 101 | "src/subgraph/square.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 102 | "src/subgraph/squared-difference.c", |
| 103 | "src/subgraph/subtract.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 104 | "src/subgraph/unpooling-2d.c", |
| 105 | ] |
| 106 | |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 107 | TABLE_SRCS = [ |
| 108 | "src/tables/exp2-k-over-64.c", |
| 109 | "src/tables/exp2-k-over-2048.c", |
| 110 | ] |
| 111 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 112 | SCALAR_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 113 | "src/f32-argmaxpool/4x-scalar-c1.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 114 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 115 | "src/f32-argmaxpool/9x-scalar-c1.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 116 | "src/f32-avgpool/9p8x-minmax-scalar-c1.c", |
| 117 | "src/f32-avgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 118 | "src/f32-clamp/gen/scalar-x1.c", |
| 119 | "src/f32-clamp/gen/scalar-x2.c", |
| 120 | "src/f32-clamp/gen/scalar-x4.c", |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 121 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
Marat Dukhan | 6b7dfae | 2019-12-04 16:00:52 -0800 | [diff] [blame] | 122 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 123 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c", |
| 124 | "src/f32-dwconv-chw/3x3p1-scalar.c", |
| 125 | "src/f32-dwconv-chw/3x3s2p1-scalar.c", |
| 126 | "src/f32-dwconv-chw/5x5p2-scalar.c", |
| 127 | "src/f32-dwconv-chw/5x5s2p2-scalar.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 128 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 129 | "src/f32-dwconv/gen/up1x4-scalar.c", |
| 130 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 131 | "src/f32-dwconv/gen/up1x9-scalar.c", |
| 132 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 133 | "src/f32-dwconv/gen/up1x25-scalar.c", |
| 134 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 135 | "src/f32-dwconv/gen/up2x4-scalar.c", |
| 136 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 137 | "src/f32-dwconv/gen/up2x9-scalar.c", |
| 138 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 139 | "src/f32-dwconv/gen/up2x25-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 140 | "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c", |
| 141 | "src/f32-dwconv/gen/up1x4-minmax-scalar.c", |
| 142 | "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c", |
| 143 | "src/f32-dwconv/gen/up1x9-minmax-scalar.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 144 | "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c", |
| 145 | "src/f32-dwconv/gen/up1x25-minmax-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 146 | "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c", |
| 147 | "src/f32-dwconv/gen/up2x4-minmax-scalar.c", |
| 148 | "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c", |
| 149 | "src/f32-dwconv/gen/up2x9-minmax-scalar.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 150 | "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c", |
| 151 | "src/f32-dwconv/gen/up2x25-minmax-scalar.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 152 | "src/f32-gavgpool-cw/scalar-x1.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 153 | "src/f32-gavgpool/7p7x-minmax-scalar-c1.c", |
| 154 | "src/f32-gavgpool/7x-minmax-scalar-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 155 | "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c", |
| 156 | "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c", |
| 157 | "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 158 | "src/f32-gemm/gen/1x4-scalar.c", |
| 159 | "src/f32-gemm/gen/2x4-scalar.c", |
| 160 | "src/f32-gemm/gen/4x2-scalar.c", |
| 161 | "src/f32-gemm/gen/4x4-scalar.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 162 | "src/f32-gemm/gen/1x4-relu-scalar.c", |
| 163 | "src/f32-gemm/gen/2x4-relu-scalar.c", |
| 164 | "src/f32-gemm/gen/4x2-relu-scalar.c", |
| 165 | "src/f32-gemm/gen/4x4-relu-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 166 | "src/f32-gemm/gen/1x4-minmax-scalar.c", |
| 167 | "src/f32-gemm/gen/2x4-minmax-scalar.c", |
| 168 | "src/f32-gemm/gen/4x2-minmax-scalar.c", |
| 169 | "src/f32-gemm/gen/4x4-minmax-scalar.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 170 | "src/f32-hswish/gen/scalar-x1.c", |
| 171 | "src/f32-hswish/gen/scalar-x2.c", |
| 172 | "src/f32-hswish/gen/scalar-x4.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 173 | "src/f32-ibilinear/gen/scalar-c1.c", |
| 174 | "src/f32-ibilinear/gen/scalar-c2.c", |
| 175 | "src/f32-ibilinear/gen/scalar-c4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 176 | "src/f32-igemm/gen/1x4-scalar.c", |
| 177 | "src/f32-igemm/gen/2x4-scalar.c", |
| 178 | "src/f32-igemm/gen/4x2-scalar.c", |
| 179 | "src/f32-igemm/gen/4x4-scalar.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 180 | "src/f32-igemm/gen/1x4-relu-scalar.c", |
| 181 | "src/f32-igemm/gen/2x4-relu-scalar.c", |
| 182 | "src/f32-igemm/gen/4x2-relu-scalar.c", |
| 183 | "src/f32-igemm/gen/4x4-relu-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 184 | "src/f32-igemm/gen/1x4-minmax-scalar.c", |
| 185 | "src/f32-igemm/gen/2x4-minmax-scalar.c", |
| 186 | "src/f32-igemm/gen/4x2-minmax-scalar.c", |
| 187 | "src/f32-igemm/gen/4x4-minmax-scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 188 | "src/f32-maxpool/9p8x-minmax-scalar-c1.c", |
| 189 | "src/f32-pavgpool/9p8x-minmax-scalar-c1.c", |
| 190 | "src/f32-pavgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 191 | "src/f32-ppmm/gen/2x4-minmax-scalar.c", |
| 192 | "src/f32-ppmm/gen/3x3-minmax-scalar.c", |
| 193 | "src/f32-ppmm/gen/4x2-minmax-scalar.c", |
| 194 | "src/f32-ppmm/gen/4x4-minmax-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 195 | "src/f32-prelu/gen/scalar-2x1.c", |
| 196 | "src/f32-prelu/gen/scalar-2x4.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 197 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
| 198 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
| 199 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
| 200 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
| 201 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 202 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
| 203 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
| 204 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
| 205 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
| 206 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
| 207 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 208 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 209 | "src/f32-rmax/scalar.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 210 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 211 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 212 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
| 213 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 214 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 215 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
| 216 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 217 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 218 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 219 | "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c", |
| 220 | "src/f32-spmm/gen/1x1-minmax-scalar.c", |
| 221 | "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c", |
| 222 | "src/f32-spmm/gen/2x1-minmax-scalar.c", |
| 223 | "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c", |
| 224 | "src/f32-spmm/gen/4x1-minmax-scalar.c", |
| 225 | "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c", |
| 226 | "src/f32-spmm/gen/8x1-minmax-scalar.c", |
| 227 | "src/f32-spmm/gen/8x2-minmax-scalar.c", |
| 228 | "src/f32-spmm/gen/8x4-minmax-scalar.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 229 | "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c", |
| 230 | "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c", |
| 231 | "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c", |
| 232 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c", |
| 233 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c", |
| 234 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c", |
| 235 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c", |
| 236 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c", |
| 237 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c", |
| 238 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c", |
| 239 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c", |
| 240 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 241 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 242 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 243 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| 244 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 245 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 246 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| 247 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 248 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 249 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| 250 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 251 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 252 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 253 | "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c", |
| 254 | "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c", |
| 255 | "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c", |
| 256 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c", |
| 257 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c", |
| 258 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c", |
| 259 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c", |
| 260 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c", |
| 261 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 262 | "src/f32-vbinary/gen/vrsqrdiffc-scalar-x1.c", |
| 263 | "src/f32-vbinary/gen/vrsqrdiffc-scalar-x2.c", |
| 264 | "src/f32-vbinary/gen/vrsqrdiffc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 265 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c", |
| 266 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c", |
| 267 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 268 | "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c", |
| 269 | "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c", |
| 270 | "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c", |
| 271 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c", |
| 272 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c", |
| 273 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 274 | "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c", |
| 275 | "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c", |
| 276 | "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c", |
| 277 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c", |
| 278 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c", |
| 279 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 280 | "src/f32-vlrelu/gen/scalar-x1.c", |
| 281 | "src/f32-vlrelu/gen/scalar-x2.c", |
| 282 | "src/f32-vlrelu/gen/scalar-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 283 | "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c", |
| 284 | "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c", |
| 285 | "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 286 | "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c", |
| 287 | "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c", |
| 288 | "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c", |
| 289 | "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c", |
| 290 | "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c", |
| 291 | "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c", |
| 292 | "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c", |
| 293 | "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c", |
| 294 | "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c", |
| 295 | "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c", |
| 296 | "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c", |
| 297 | "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 298 | "src/f32-vunary/gen/vabs-scalar-x1.c", |
| 299 | "src/f32-vunary/gen/vabs-scalar-x2.c", |
| 300 | "src/f32-vunary/gen/vabs-scalar-x4.c", |
| 301 | "src/f32-vunary/gen/vneg-scalar-x1.c", |
| 302 | "src/f32-vunary/gen/vneg-scalar-x2.c", |
| 303 | "src/f32-vunary/gen/vneg-scalar-x4.c", |
| 304 | "src/f32-vunary/gen/vsqr-scalar-x1.c", |
| 305 | "src/f32-vunary/gen/vsqr-scalar-x2.c", |
| 306 | "src/f32-vunary/gen/vsqr-scalar-x4.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 307 | "src/math/expminus-scalar-lut2048-p1.c", |
| 308 | "src/math/expminus-scalar-lut64-p2.c", |
| 309 | "src/math/expminus-scalar-p5.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 310 | "src/math/roundne-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 311 | "src/math/roundne-scalar-nearbyint.c", |
| 312 | "src/math/roundne-scalar-rint.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 313 | "src/math/roundd-scalar-addsub.c", |
| 314 | "src/math/roundd-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 315 | "src/math/roundd-scalar-floor.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 316 | "src/math/roundu-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 317 | "src/math/roundu-scalar-ceil.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 318 | "src/math/roundu-scalar-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 319 | "src/math/roundz-scalar-addsub.c", |
| 320 | "src/math/roundz-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 321 | "src/math/roundz-scalar-trunc.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 322 | "src/math/sigmoid-scalar-lut2048-p1-div.c", |
| 323 | "src/math/sigmoid-scalar-lut64-p2-div.c", |
| 324 | "src/math/sigmoid-scalar-p5-div.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 325 | "src/q8-avgpool/9p8x-minmax-scalar-c1.c", |
| 326 | "src/q8-avgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 327 | "src/q8-dwconv/up1x9-minmax-scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 328 | "src/q8-gavgpool/7p7x-minmax-scalar-c1.c", |
| 329 | "src/q8-gavgpool/7x-minmax-scalar-c1.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 330 | "src/q8-gemm/2x2-minmax-scalar.c", |
| 331 | "src/q8-igemm/2x2-minmax-scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 332 | "src/q8-vadd/minmax-scalar.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 333 | "src/u8-clamp/scalar-x4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 334 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 335 | "src/u8-maxpool/9p8x-minmax-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 336 | "src/u8-rmax/scalar.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 337 | "src/x32-fill/scalar-float.c", |
| 338 | "src/x32-fill/scalar-int.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 339 | "src/x32-packx/x2-scalar.c", |
| 340 | "src/x32-packx/x3-scalar.c", |
| 341 | "src/x32-packx/x4-scalar.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 342 | "src/x32-pad/scalar-int.c", |
| 343 | "src/x32-pad/scalar-float.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 344 | "src/x32-unpool/scalar.c", |
| 345 | "src/x32-zip/x2-scalar.c", |
| 346 | "src/x32-zip/x3-scalar.c", |
| 347 | "src/x32-zip/x4-scalar.c", |
| 348 | "src/x32-zip/xm-scalar.c", |
| 349 | "src/x8-lut/scalar.c", |
| 350 | "src/x8-zip/x2-scalar.c", |
| 351 | "src/x8-zip/x3-scalar.c", |
| 352 | "src/x8-zip/x4-scalar.c", |
| 353 | "src/x8-zip/xm-scalar.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 354 | "src/requantization/precise-scalar.c", |
| 355 | "src/requantization/fp32-scalar.c", |
| 356 | "src/requantization/q31-scalar.c", |
| 357 | "src/requantization/gemmlowp-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 358 | ] |
| 359 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 360 | WASM_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 361 | "src/f32-avgpool/9p8x-minmax-wasm-c1.c", |
| 362 | "src/f32-avgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 363 | "src/f32-clamp/gen/wasm-x1.c", |
| 364 | "src/f32-clamp/gen/wasm-x2.c", |
| 365 | "src/f32-clamp/gen/wasm-x4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 366 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 367 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 368 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 369 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 370 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 371 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 372 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 373 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 374 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 375 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| 376 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 377 | "src/f32-dwconv/gen/up2x25-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 378 | "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c", |
| 379 | "src/f32-dwconv/gen/up1x4-minmax-wasm.c", |
| 380 | "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c", |
| 381 | "src/f32-dwconv/gen/up1x9-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 382 | "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c", |
| 383 | "src/f32-dwconv/gen/up1x25-minmax-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 384 | "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c", |
| 385 | "src/f32-dwconv/gen/up2x4-minmax-wasm.c", |
| 386 | "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c", |
| 387 | "src/f32-dwconv/gen/up2x9-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 388 | "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c", |
| 389 | "src/f32-dwconv/gen/up2x25-minmax-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 390 | "src/f32-gavgpool/7p7x-minmax-wasm-c1.c", |
| 391 | "src/f32-gavgpool/7x-minmax-wasm-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 392 | "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c", |
| 393 | "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c", |
| 394 | "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 395 | "src/f32-gemm/gen/1x4-wasm.c", |
| 396 | "src/f32-gemm/gen/2x4-wasm.c", |
| 397 | "src/f32-gemm/gen/4x2-wasm.c", |
| 398 | "src/f32-gemm/gen/4x4-wasm.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 399 | "src/f32-gemm/gen/1x4-relu-wasm.c", |
| 400 | "src/f32-gemm/gen/2x4-relu-wasm.c", |
| 401 | "src/f32-gemm/gen/4x2-relu-wasm.c", |
| 402 | "src/f32-gemm/gen/4x4-relu-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 403 | "src/f32-gemm/gen/1x4-minmax-wasm.c", |
| 404 | "src/f32-gemm/gen/2x4-minmax-wasm.c", |
| 405 | "src/f32-gemm/gen/4x2-minmax-wasm.c", |
| 406 | "src/f32-gemm/gen/4x4-minmax-wasm.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 407 | "src/f32-hswish/gen/wasm-x1.c", |
| 408 | "src/f32-hswish/gen/wasm-x2.c", |
| 409 | "src/f32-hswish/gen/wasm-x4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 410 | "src/f32-igemm/gen/1x4-wasm.c", |
| 411 | "src/f32-igemm/gen/2x4-wasm.c", |
| 412 | "src/f32-igemm/gen/4x2-wasm.c", |
| 413 | "src/f32-igemm/gen/4x4-wasm.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 414 | "src/f32-igemm/gen/1x4-relu-wasm.c", |
| 415 | "src/f32-igemm/gen/2x4-relu-wasm.c", |
| 416 | "src/f32-igemm/gen/4x2-relu-wasm.c", |
| 417 | "src/f32-igemm/gen/4x4-relu-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 418 | "src/f32-igemm/gen/1x4-minmax-wasm.c", |
| 419 | "src/f32-igemm/gen/2x4-minmax-wasm.c", |
| 420 | "src/f32-igemm/gen/4x2-minmax-wasm.c", |
| 421 | "src/f32-igemm/gen/4x4-minmax-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 422 | "src/f32-maxpool/9p8x-minmax-wasm-c1.c", |
| 423 | "src/f32-pavgpool/9p8x-minmax-wasm-c1.c", |
| 424 | "src/f32-pavgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 425 | "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c", |
| 426 | "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c", |
| 427 | "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c", |
| 428 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c", |
| 429 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c", |
| 430 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c", |
| 431 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c", |
| 432 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c", |
| 433 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c", |
| 434 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c", |
| 435 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c", |
| 436 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 437 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 438 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 439 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
| 440 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 441 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 442 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
| 443 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 444 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 445 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
| 446 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 447 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 448 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 449 | "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c", |
| 450 | "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c", |
| 451 | "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c", |
| 452 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c", |
| 453 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c", |
| 454 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c", |
| 455 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c", |
| 456 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c", |
| 457 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c", |
| 458 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c", |
| 459 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c", |
| 460 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c", |
| 461 | "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c", |
| 462 | "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c", |
| 463 | "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c", |
| 464 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c", |
| 465 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c", |
| 466 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 467 | "src/f32-vlrelu/gen/wasm-x1.c", |
| 468 | "src/f32-vlrelu/gen/wasm-x2.c", |
| 469 | "src/f32-vlrelu/gen/wasm-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 470 | "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c", |
| 471 | "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c", |
| 472 | "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 473 | ] |
| 474 | |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 475 | WASMSIMD_UKERNELS = [ |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame^] | 476 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c", |
| 477 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c", |
| 478 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c", |
| 479 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c", |
| 480 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c", |
| 481 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c", |
| 482 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c", |
| 483 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c", |
| 484 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c", |
| 485 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c", |
| 486 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c", |
| 487 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c", |
| 488 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c", |
| 489 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c", |
| 490 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c", |
| 491 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c", |
| 492 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c", |
| 493 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c", |
| 494 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c", |
| 495 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c", |
| 496 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c", |
| 497 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c", |
| 498 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c", |
| 499 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c", |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 500 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 501 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 502 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 503 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 504 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 505 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 506 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 507 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 508 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 509 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 510 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 511 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 512 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 513 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 514 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 515 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 516 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 517 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 518 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 519 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 520 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 521 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 522 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 523 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 524 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 525 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 526 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 527 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 528 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 529 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
| 530 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 531 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c", |
| 532 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c", |
| 533 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 534 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c", |
| 535 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c", |
| 536 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 537 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c", |
| 538 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c", |
| 539 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 540 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c", |
| 541 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c", |
| 542 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 543 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c", |
| 544 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c", |
| 545 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 546 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c", |
| 547 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c", |
| 548 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 549 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c", |
| 550 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c", |
| 551 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 552 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c", |
| 553 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c", |
| 554 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 555 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c", |
| 556 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c", |
| 557 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 558 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c", |
| 559 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c", |
| 560 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 561 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 562 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 563 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 564 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 565 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 566 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 567 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 568 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 569 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 570 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 571 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 572 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 573 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 574 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 575 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 576 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 577 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 578 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 579 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 580 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 581 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 582 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 583 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 584 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 585 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 586 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 587 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 588 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 589 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 590 | "src/requantization/fp32-wasmsimd.c", |
| 591 | ] |
| 592 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 593 | PSIMD_FASTMATH_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 594 | "src/f32-argmaxpool/4x-psimd-c4.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 595 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 596 | "src/f32-argmaxpool/9x-psimd-c4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 597 | "src/f32-avgpool/9p8x-minmax-psimd-c4.c", |
| 598 | "src/f32-avgpool/9x-minmax-psimd-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 599 | "src/f32-clamp/gen/psimd-x4.c", |
| 600 | "src/f32-clamp/gen/psimd-x8.c", |
Erich Elsen | 0a1970e | 2020-06-10 09:24:59 -0700 | [diff] [blame] | 601 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c", |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 602 | "src/f32-dwconv-chw/3x3p1-psimd.c", |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 603 | "src/f32-dwconv-chw/3x3s2p1-psimd.c", |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 604 | "src/f32-dwconv-chw/5x5p2-psimd.c", |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 605 | "src/f32-dwconv-chw/5x5s2p2-psimd.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 606 | "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c", |
| 607 | "src/f32-dwconv/gen/up4x25-minmax-psimd.c", |
| 608 | "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c", |
| 609 | "src/f32-dwconv/gen/up4x4-minmax-psimd.c", |
| 610 | "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c", |
| 611 | "src/f32-dwconv/gen/up4x9-minmax-psimd.c", |
| 612 | "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c", |
| 613 | "src/f32-dwconv/gen/up8x25-minmax-psimd.c", |
| 614 | "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c", |
| 615 | "src/f32-dwconv/gen/up8x4-minmax-psimd.c", |
| 616 | "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c", |
| 617 | "src/f32-dwconv/gen/up8x9-minmax-psimd.c", |
Erich Elsen | 6f278b5 | 2020-06-10 16:13:11 -0700 | [diff] [blame] | 618 | "src/f32-gavgpool-cw/psimd-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 619 | "src/f32-gavgpool/7p7x-minmax-psimd-c4.c", |
| 620 | "src/f32-gavgpool/7x-minmax-psimd-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 621 | "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c", |
| 622 | "src/f32-gemm/gen/1x8-minmax-psimd-splat.c", |
| 623 | "src/f32-gemm/gen/1x8s4-minmax-psimd.c", |
| 624 | "src/f32-gemm/gen/4x2c4-minmax-psimd.c", |
| 625 | "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c", |
| 626 | "src/f32-gemm/gen/4x8-minmax-psimd-splat.c", |
| 627 | "src/f32-gemm/gen/4x8s4-minmax-psimd.c", |
| 628 | "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c", |
| 629 | "src/f32-gemm/gen/6x8-minmax-psimd-splat.c", |
| 630 | "src/f32-gemm/gen/6x8s4-minmax-psimd.c", |
| 631 | "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c", |
| 632 | "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c", |
| 633 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c", |
| 634 | "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c", |
| 635 | "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c", |
| 636 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c", |
| 637 | "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c", |
| 638 | "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c", |
| 639 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 640 | "src/f32-hswish/gen/psimd-x4.c", |
| 641 | "src/f32-hswish/gen/psimd-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 642 | "src/f32-ibilinear/gen/psimd-c4.c", |
| 643 | "src/f32-ibilinear/gen/psimd-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 644 | "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c", |
| 645 | "src/f32-igemm/gen/1x8-minmax-psimd-splat.c", |
| 646 | "src/f32-igemm/gen/1x8s4-minmax-psimd.c", |
| 647 | "src/f32-igemm/gen/4x2c4-minmax-psimd.c", |
| 648 | "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c", |
| 649 | "src/f32-igemm/gen/4x8-minmax-psimd-splat.c", |
| 650 | "src/f32-igemm/gen/4x8s4-minmax-psimd.c", |
| 651 | "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c", |
| 652 | "src/f32-igemm/gen/6x8-minmax-psimd-splat.c", |
| 653 | "src/f32-igemm/gen/6x8s4-minmax-psimd.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 654 | "src/f32-maxpool/9p8x-minmax-psimd-c4.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 655 | "src/f32-spmm/gen/4x1-minmax-psimd.c", |
| 656 | "src/f32-spmm/gen/8x1-minmax-psimd.c", |
| 657 | "src/f32-spmm/gen/16x1-minmax-psimd.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 658 | "src/f32-pavgpool/9p8x-minmax-psimd-c4.c", |
| 659 | "src/f32-pavgpool/9x-minmax-psimd-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 660 | "src/f32-ppmm/gen/4x8-minmax-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 661 | "src/f32-prelu/gen/psimd-2x4.c", |
| 662 | "src/f32-prelu/gen/psimd-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 663 | "src/f32-rmax/psimd.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 664 | "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c", |
| 665 | "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c", |
| 666 | "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c", |
| 667 | "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c", |
| 668 | "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c", |
| 669 | "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c", |
| 670 | "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c", |
| 671 | "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 672 | "src/f32-vbinary/gen/vmax-psimd-x4.c", |
| 673 | "src/f32-vbinary/gen/vmax-psimd-x8.c", |
| 674 | "src/f32-vbinary/gen/vmaxc-psimd-x4.c", |
| 675 | "src/f32-vbinary/gen/vmaxc-psimd-x8.c", |
| 676 | "src/f32-vbinary/gen/vmin-psimd-x4.c", |
| 677 | "src/f32-vbinary/gen/vmin-psimd-x8.c", |
| 678 | "src/f32-vbinary/gen/vminc-psimd-x4.c", |
| 679 | "src/f32-vbinary/gen/vminc-psimd-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 680 | "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c", |
| 681 | "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c", |
| 682 | "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c", |
| 683 | "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c", |
| 684 | "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c", |
| 685 | "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 686 | "src/f32-vbinary/gen/vrsqrdiffc-psimd-x4.c", |
| 687 | "src/f32-vbinary/gen/vrsqrdiffc-psimd-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 688 | "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c", |
| 689 | "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 690 | "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c", |
| 691 | "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c", |
| 692 | "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c", |
| 693 | "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 694 | "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c", |
| 695 | "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c", |
| 696 | "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c", |
| 697 | "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 698 | "src/f32-vlrelu/gen/psimd-x4.c", |
| 699 | "src/f32-vlrelu/gen/psimd-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 700 | "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c", |
| 701 | "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 702 | "src/f32-vunary/gen/vabs-psimd-x4.c", |
| 703 | "src/f32-vunary/gen/vabs-psimd-x8.c", |
| 704 | "src/f32-vunary/gen/vneg-psimd-x4.c", |
| 705 | "src/f32-vunary/gen/vneg-psimd-x8.c", |
| 706 | "src/f32-vunary/gen/vsqr-psimd-x4.c", |
| 707 | "src/f32-vunary/gen/vsqr-psimd-x8.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 708 | "src/x32-fill/psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 709 | "src/x32-packx/x4-psimd.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 710 | "src/x32-pad/psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 711 | "src/x32-unpool/psimd.c", |
| 712 | "src/x32-zip/x2-psimd.c", |
| 713 | "src/x32-zip/x3-psimd.c", |
| 714 | "src/x32-zip/x4-psimd.c", |
| 715 | "src/x32-zip/xm-psimd.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 716 | "src/requantization/precise-psimd.c", |
| 717 | "src/requantization/fp32-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 718 | ] |
| 719 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 720 | PSIMD_ACCMATH_UKERNELS = [ |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 721 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c", |
| 722 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c", |
| 723 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c", |
| 724 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c", |
| 725 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c", |
| 726 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c", |
| 727 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c", |
| 728 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c", |
| 729 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c", |
| 730 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c", |
| 731 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c", |
| 732 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 733 | "src/f32-sigmoid/gen/psimd-p5-div-x4.c", |
| 734 | "src/f32-sigmoid/gen/psimd-p5-div-x8.c", |
| 735 | "src/f32-sigmoid/gen/psimd-p5-div-x12.c", |
| 736 | "src/f32-sigmoid/gen/psimd-p5-div-x16.c", |
| 737 | "src/f32-sigmoid/gen/psimd-p5-div-x20.c", |
| 738 | "src/f32-sigmoid/gen/psimd-p5-div-x24.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 739 | "src/f32-vrnd/gen/vrndne-psimd-x4.c", |
| 740 | "src/f32-vrnd/gen/vrndne-psimd-x8.c", |
| 741 | "src/f32-vrnd/gen/vrndz-psimd-x4.c", |
| 742 | "src/f32-vrnd/gen/vrndz-psimd-x8.c", |
| 743 | "src/f32-vrnd/gen/vrndu-psimd-x4.c", |
| 744 | "src/f32-vrnd/gen/vrndu-psimd-x8.c", |
| 745 | "src/f32-vrnd/gen/vrndd-psimd-x4.c", |
| 746 | "src/f32-vrnd/gen/vrndd-psimd-x8.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 747 | "src/math/roundne-psimd-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 748 | "src/math/roundd-psimd-addsub.c", |
| 749 | "src/math/roundu-psimd-addsub.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 750 | "src/math/roundz-psimd-addsub.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 751 | "src/math/sigmoid-psimd-p5-div.c", |
| 752 | ] |
| 753 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 754 | # ISA-specific micro-kernels |
| 755 | NEON_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 756 | "src/f32-avgpool/9p8x-minmax-neon-c4.c", |
| 757 | "src/f32-avgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 758 | "src/f32-clamp/gen/neon-x4.c", |
| 759 | "src/f32-clamp/gen/neon-x8.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 760 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c", |
| 761 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 762 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c", |
| 763 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 764 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c", |
| 765 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 766 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c", |
| 767 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 768 | "src/f32-dwconv/gen/up4x4-minmax-neon.c", |
| 769 | "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", |
| 770 | "src/f32-dwconv/gen/up8x4-minmax-neon.c", |
| 771 | "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 772 | "src/f32-dwconv/gen/up4x9-minmax-neon.c", |
| 773 | "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", |
| 774 | "src/f32-dwconv/gen/up8x9-minmax-neon.c", |
| 775 | "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 776 | "src/f32-dwconv/gen/up4x25-minmax-neon.c", |
| 777 | "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", |
| 778 | "src/f32-dwconv/gen/up8x25-minmax-neon.c", |
| 779 | "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 780 | "src/f32-gavgpool-cw/neon-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 781 | "src/f32-gavgpool/7p7x-minmax-neon-c4.c", |
| 782 | "src/f32-gavgpool/7x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 783 | "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 784 | "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 785 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 786 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 787 | "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c", |
| 788 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 789 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 790 | "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 791 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 792 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 793 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 794 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 795 | "src/f32-gemm/gen/1x8s4-minmax-neon.c", |
| 796 | "src/f32-gemm/gen/4x8s4-minmax-neon.c", |
| 797 | "src/f32-gemm/gen/6x8s4-minmax-neon.c", |
| 798 | "src/f32-gemm/gen/8x8s4-minmax-neon.c", |
| 799 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c", |
| 800 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c", |
| 801 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c", |
| 802 | "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c", |
| 803 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c", |
| 804 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c", |
| 805 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c", |
| 806 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c", |
| 807 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c", |
| 808 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c", |
| 809 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c", |
| 810 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c", |
| 811 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c", |
| 812 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c", |
| 813 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 814 | "src/f32-hswish/gen/neon-x4.c", |
| 815 | "src/f32-hswish/gen/neon-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 816 | "src/f32-ibilinear/gen/neon-c4.c", |
| 817 | "src/f32-ibilinear/gen/neon-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 818 | "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 819 | "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 820 | "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c", |
| 821 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 822 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 823 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 824 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 825 | "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 826 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 827 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 828 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 829 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 830 | "src/f32-igemm/gen/1x8s4-minmax-neon.c", |
| 831 | "src/f32-igemm/gen/4x8s4-minmax-neon.c", |
| 832 | "src/f32-igemm/gen/6x8s4-minmax-neon.c", |
| 833 | "src/f32-igemm/gen/8x8s4-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 834 | "src/f32-maxpool/9p8x-minmax-neon-c4.c", |
| 835 | "src/f32-pavgpool/9p8x-minmax-neon-c4.c", |
| 836 | "src/f32-pavgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 837 | "src/f32-ppmm/gen/4x8-minmax-neon.c", |
| 838 | "src/f32-ppmm/gen/8x8-minmax-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 839 | "src/f32-prelu/gen/neon-2x4.c", |
| 840 | "src/f32-prelu/gen/neon-2x8.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 841 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 842 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 843 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 844 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 845 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 846 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 847 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 848 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 849 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 850 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| 851 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 852 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 853 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| 854 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| 855 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| 856 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| 857 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 858 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| 859 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| 860 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 861 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| 862 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 863 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 864 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 865 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 866 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 867 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 868 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 869 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 870 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 871 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 872 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| 873 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 874 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 875 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 876 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 877 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 878 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 879 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 880 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 881 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 882 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 883 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 884 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 885 | "src/f32-vbinary/gen/vadd-minmax-neon-x4.c", |
| 886 | "src/f32-vbinary/gen/vadd-minmax-neon-x8.c", |
| 887 | "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c", |
| 888 | "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 889 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 890 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 891 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 892 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 893 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 894 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 895 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 896 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 897 | "src/f32-vbinary/gen/vmul-minmax-neon-x4.c", |
| 898 | "src/f32-vbinary/gen/vmul-minmax-neon-x8.c", |
| 899 | "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c", |
| 900 | "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 901 | "src/f32-vbinary/gen/vrsqrdiffc-neon-x4.c", |
| 902 | "src/f32-vbinary/gen/vrsqrdiffc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 903 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c", |
| 904 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 905 | "src/f32-vbinary/gen/vsqrdiff-neon-x4.c", |
| 906 | "src/f32-vbinary/gen/vsqrdiff-neon-x8.c", |
| 907 | "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c", |
| 908 | "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 909 | "src/f32-vbinary/gen/vsub-minmax-neon-x4.c", |
| 910 | "src/f32-vbinary/gen/vsub-minmax-neon-x8.c", |
| 911 | "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c", |
| 912 | "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 913 | "src/f32-vlrelu/gen/neon-x4.c", |
| 914 | "src/f32-vlrelu/gen/neon-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 915 | "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c", |
| 916 | "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 917 | "src/f32-vrnd/gen/vrndne-neon-x4.c", |
| 918 | "src/f32-vrnd/gen/vrndne-neon-x8.c", |
| 919 | "src/f32-vrnd/gen/vrndz-neon-x4.c", |
| 920 | "src/f32-vrnd/gen/vrndz-neon-x8.c", |
| 921 | "src/f32-vrnd/gen/vrndu-neon-x4.c", |
| 922 | "src/f32-vrnd/gen/vrndu-neon-x8.c", |
| 923 | "src/f32-vrnd/gen/vrndd-neon-x4.c", |
| 924 | "src/f32-vrnd/gen/vrndd-neon-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 925 | "src/f32-vunary/gen/vabs-neon-x4.c", |
| 926 | "src/f32-vunary/gen/vabs-neon-x8.c", |
| 927 | "src/f32-vunary/gen/vneg-neon-x4.c", |
| 928 | "src/f32-vunary/gen/vneg-neon-x8.c", |
| 929 | "src/f32-vunary/gen/vsqr-neon-x4.c", |
| 930 | "src/f32-vunary/gen/vsqr-neon-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 931 | "src/q8-avgpool/9p8x-minmax-neon-c8.c", |
| 932 | "src/q8-avgpool/9x-minmax-neon-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 933 | "src/q8-dwconv/up8x9-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 934 | "src/q8-gavgpool/7p7x-minmax-neon-c8.c", |
| 935 | "src/q8-gavgpool/7x-minmax-neon-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 936 | "src/q8-gemm/4x8-minmax-neon.c", |
| 937 | "src/q8-gemm/8x8-minmax-neon.c", |
| 938 | "src/q8-igemm/4x8-minmax-neon.c", |
| 939 | "src/q8-igemm/8x8-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 940 | "src/q8-vadd/minmax-neon.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 941 | "src/u8-clamp/neon-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 942 | "src/u8-maxpool/9p8x-minmax-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 943 | "src/u8-rmax/neon.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 944 | "src/x32-fill/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 945 | "src/x32-packx/x4-neon-st4.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 946 | "src/x32-pad/neon.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 947 | "src/x32-unpool/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 948 | "src/x32-zip/x2-neon.c", |
| 949 | "src/x32-zip/x3-neon.c", |
| 950 | "src/x32-zip/x4-neon.c", |
| 951 | "src/x32-zip/xm-neon.c", |
| 952 | "src/x8-zip/x2-neon.c", |
| 953 | "src/x8-zip/x3-neon.c", |
| 954 | "src/x8-zip/x4-neon.c", |
| 955 | "src/x8-zip/xm-neon.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 956 | "src/math/roundne-neon-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 957 | "src/math/roundd-neon-addsub.c", |
| 958 | "src/math/roundd-neon-cvt.c", |
| 959 | "src/math/roundu-neon-addsub.c", |
| 960 | "src/math/roundu-neon-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 961 | "src/math/roundz-neon-addsub.c", |
| 962 | "src/math/roundz-neon-cvt.c", |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 963 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 964 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 965 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 966 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 967 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 968 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 969 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 970 | "src/requantization/precise-neon.c", |
| 971 | "src/requantization/fp32-neon.c", |
| 972 | "src/requantization/q31-neon.c", |
| 973 | "src/requantization/gemmlowp-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 974 | ] |
| 975 | |
| 976 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 977 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 978 | "src/f32-ibilinear/gen/neonfma-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 979 | "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 980 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 981 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 982 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 983 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 984 | "src/f32-igemm/gen/1x8s4-minmax-neonfma.c", |
| 985 | "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", |
| 986 | "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", |
| 987 | "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 988 | "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", |
| 989 | "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", |
| 990 | "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", |
| 991 | "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 992 | "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", |
| 993 | "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", |
| 994 | "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", |
| 995 | "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 996 | "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", |
| 997 | "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", |
| 998 | "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", |
| 999 | "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1000 | "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1001 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1002 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1003 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1004 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1005 | "src/f32-gemm/gen/1x8s4-minmax-neonfma.c", |
| 1006 | "src/f32-gemm/gen/4x8s4-minmax-neonfma.c", |
| 1007 | "src/f32-gemm/gen/6x8s4-minmax-neonfma.c", |
| 1008 | "src/f32-gemm/gen/8x8s4-minmax-neonfma.c", |
| 1009 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c", |
| 1010 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c", |
| 1011 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c", |
| 1012 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c", |
| 1013 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c", |
| 1014 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c", |
| 1015 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c", |
| 1016 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c", |
| 1017 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1018 | "src/f32-hswish/gen/neonfma-x4.c", |
| 1019 | "src/f32-hswish/gen/neonfma-x8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1020 | "src/f32-ppmm/gen/4x8-minmax-neonfma.c", |
| 1021 | "src/f32-ppmm/gen/8x8-minmax-neonfma.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1022 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 1023 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 1024 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 1025 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 1026 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 1027 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 1028 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 1029 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 1030 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 1031 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| 1032 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 1033 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 1034 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| 1035 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| 1036 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| 1037 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| 1038 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 1039 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| 1040 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| 1041 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 1042 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| 1043 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 1044 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 1045 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1046 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 1047 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 1048 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 1049 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 1050 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 1051 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 1052 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 1053 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 1054 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 1055 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 1056 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 1057 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 1058 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 1059 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 1060 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 1061 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 1062 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 1063 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| 1064 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 1065 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 1066 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 1067 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 1068 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 1069 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| 1070 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 1071 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 1072 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 1073 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 1074 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 1075 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| 1076 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 1077 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 1078 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 1079 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 1080 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 1081 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| 1082 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 1083 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 1084 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 1085 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 1086 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 1087 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| 1088 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 1089 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 1090 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 1091 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 1092 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 1093 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| 1094 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 1095 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 1096 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 1097 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 1098 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 1099 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1100 | "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c", |
| 1101 | "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 1102 | "src/math/exp-neonfma-lut64-p2.c", |
| 1103 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 1104 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 1105 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1106 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1107 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 1108 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 1109 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| 1110 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 1111 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 1112 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| 1113 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 1114 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 1115 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| 1116 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 1117 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 1118 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| 1119 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 1120 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 1121 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| 1122 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 1123 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 1124 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1125 | ] |
| 1126 | |
| 1127 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1128 | "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c", |
| 1129 | "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c", |
| 1130 | "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c", |
| 1131 | "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c", |
| 1132 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c", |
| 1133 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1134 | "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1135 | "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1136 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1137 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1138 | "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c", |
| 1139 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1140 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
| 1141 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c", |
| 1142 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c", |
| 1143 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c", |
| 1144 | "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c", |
| 1145 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c", |
| 1146 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c", |
| 1147 | "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1148 | "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1149 | "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c", |
| 1150 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1151 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1152 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1153 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1154 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c", |
| 1155 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1156 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c", |
| 1157 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1158 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c", |
| 1159 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1160 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c", |
| 1161 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1162 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c", |
| 1163 | "src/f32-dwconv-chw/3x3p1-neonfma.c", |
| 1164 | "src/f32-dwconv-chw/5x5p2-neonfma.c", |
| 1165 | "src/f32-dwconv-chw/3x3s2p1-neonfma.c", |
| 1166 | "src/f32-dwconv-chw/5x5s2p2-neonfma.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1167 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 1168 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 1169 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 1170 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 1171 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 1172 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| 1173 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 1174 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 1175 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 1176 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 1177 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 1178 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 1179 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 1180 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 1181 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 1182 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 1183 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 1184 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1185 | "src/f32-spmm/gen/12x1-minmax-neonfma.c", |
| 1186 | "src/f32-spmm/gen/12x2-minmax-neonfma.c", |
| 1187 | "src/f32-spmm/gen/12x4-minmax-neonfma.c", |
| 1188 | "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c", |
| 1189 | "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c", |
| 1190 | "src/f32-spmm/gen/16x1-minmax-neonfma.c", |
| 1191 | "src/f32-spmm/gen/16x2-minmax-neonfma.c", |
| 1192 | "src/f32-spmm/gen/16x4-minmax-neonfma.c", |
| 1193 | "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c", |
| 1194 | "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c", |
| 1195 | "src/f32-spmm/gen/4x1-minmax-neonfma.c", |
| 1196 | "src/f32-spmm/gen/4x2-minmax-neonfma.c", |
| 1197 | "src/f32-spmm/gen/4x4-minmax-neonfma.c", |
| 1198 | "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c", |
| 1199 | "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c", |
| 1200 | "src/f32-spmm/gen/8x1-minmax-neonfma.c", |
| 1201 | "src/f32-spmm/gen/8x2-minmax-neonfma.c", |
| 1202 | "src/f32-spmm/gen/8x4-minmax-neonfma.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1203 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| 1204 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| 1205 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| 1206 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| 1207 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| 1208 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1209 | ] |
| 1210 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1211 | NEONV8_UKERNELS = [ |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1212 | "src/f32-vrnd/gen/vrndne-neonv8-x4.c", |
| 1213 | "src/f32-vrnd/gen/vrndne-neonv8-x8.c", |
| 1214 | "src/f32-vrnd/gen/vrndz-neonv8-x4.c", |
| 1215 | "src/f32-vrnd/gen/vrndz-neonv8-x8.c", |
| 1216 | "src/f32-vrnd/gen/vrndu-neonv8-x4.c", |
| 1217 | "src/f32-vrnd/gen/vrndu-neonv8-x8.c", |
| 1218 | "src/f32-vrnd/gen/vrndd-neonv8-x4.c", |
| 1219 | "src/f32-vrnd/gen/vrndd-neonv8-x8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1220 | "src/math/roundne-neonv8.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1221 | "src/math/roundd-neonv8.c", |
| 1222 | "src/math/roundu-neonv8.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1223 | "src/math/roundz-neonv8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1224 | ] |
| 1225 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1226 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1227 | "src/f16-clamp/gen/neonfp16arith-x16.c", |
| 1228 | "src/f16-clamp/gen/neonfp16arith-x8.c", |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 1229 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c", |
| 1230 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c", |
| 1231 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c", |
| 1232 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c", |
| 1233 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c", |
| 1234 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c", |
| 1235 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c", |
| 1236 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c", |
| 1237 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c", |
| 1238 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c", |
| 1239 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c", |
| 1240 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c", |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 1241 | "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c", |
| 1242 | "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c", |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1243 | "src/f16-hswish/gen/neonfp16arith-x16.c", |
| 1244 | "src/f16-hswish/gen/neonfp16arith-x8.c", |
| 1245 | "src/f16-prelu/gen/neonfp16arith-2x16.c", |
| 1246 | "src/f16-prelu/gen/neonfp16arith-2x8.c", |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 1247 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c", |
| 1248 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c", |
| 1249 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c", |
| 1250 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c", |
| 1251 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c", |
| 1252 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c", |
| 1253 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c", |
| 1254 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c", |
| 1255 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c", |
| 1256 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c", |
| 1257 | "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c", |
| 1258 | "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c", |
| 1259 | "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c", |
| 1260 | "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c", |
| 1261 | "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c", |
| 1262 | "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c", |
| 1263 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c", |
| 1264 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c", |
| 1265 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c", |
| 1266 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c", |
| 1267 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c", |
| 1268 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c", |
| 1269 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c", |
| 1270 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c", |
| 1271 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c", |
| 1272 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c", |
| 1273 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c", |
| 1274 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c", |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 1275 | "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c", |
| 1276 | "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c", |
Frank Barchard | 83ea939 | 2020-04-20 20:15:14 -0700 | [diff] [blame] | 1277 | "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1278 | "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1279 | "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1280 | "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1281 | "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c", |
| 1282 | "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c", |
| 1283 | "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c", |
| 1284 | "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c", |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 1285 | "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1286 | "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1287 | "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1288 | "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
Frank Barchard | 3f9f99f | 2020-05-06 01:12:04 -0700 | [diff] [blame] | 1289 | "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1290 | "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1291 | "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1292 | "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
| 1293 | "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c", |
| 1294 | "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c", |
| 1295 | "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c", |
| 1296 | "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c", |
| 1297 | "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1298 | "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1299 | "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1300 | "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1301 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c", |
| 1302 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c", |
| 1303 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c", |
| 1304 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c", |
| 1305 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c", |
| 1306 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c", |
| 1307 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c", |
| 1308 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1309 | ] |
| 1310 | |
| 1311 | SSE_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1312 | "src/f32-avgpool/9p8x-minmax-sse-c4.c", |
| 1313 | "src/f32-avgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1314 | "src/f32-clamp/gen/sse-x4.c", |
| 1315 | "src/f32-clamp/gen/sse-x8.c", |
Erich Elsen | b123340 | 2020-06-08 15:53:15 -0700 | [diff] [blame] | 1316 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c", |
| 1317 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1318 | "src/f32-dwconv-chw/3x3p1-sse.c", |
| 1319 | "src/f32-dwconv-chw/3x3s2p1-sse.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1320 | "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c", |
| 1321 | "src/f32-dwconv/gen/up4x25-minmax-sse.c", |
| 1322 | "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c", |
| 1323 | "src/f32-dwconv/gen/up4x4-minmax-sse.c", |
| 1324 | "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c", |
| 1325 | "src/f32-dwconv/gen/up4x9-minmax-sse.c", |
| 1326 | "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c", |
| 1327 | "src/f32-dwconv/gen/up8x25-minmax-sse.c", |
| 1328 | "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c", |
| 1329 | "src/f32-dwconv/gen/up8x4-minmax-sse.c", |
| 1330 | "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c", |
| 1331 | "src/f32-dwconv/gen/up8x9-minmax-sse.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1332 | "src/f32-gavgpool-cw/sse-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1333 | "src/f32-gavgpool/7p7x-minmax-sse-c4.c", |
| 1334 | "src/f32-gavgpool/7x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1335 | "src/f32-gemm/gen/1x8-minmax-sse-dup.c", |
| 1336 | "src/f32-gemm/gen/1x8-minmax-sse-load1.c", |
| 1337 | "src/f32-gemm/gen/1x8s4-minmax-sse.c", |
| 1338 | "src/f32-gemm/gen/4x2c4-minmax-sse.c", |
| 1339 | "src/f32-gemm/gen/4x8-minmax-sse-dup.c", |
| 1340 | "src/f32-gemm/gen/4x8-minmax-sse-load1.c", |
| 1341 | "src/f32-gemm/gen/4x8s4-minmax-sse.c", |
| 1342 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c", |
| 1343 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c", |
| 1344 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c", |
| 1345 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c", |
| 1346 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c", |
| 1347 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1348 | "src/f32-hswish/gen/sse-x4.c", |
| 1349 | "src/f32-hswish/gen/sse-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1350 | "src/f32-ibilinear/gen/sse-c4.c", |
| 1351 | "src/f32-ibilinear/gen/sse-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1352 | "src/f32-igemm/gen/1x8-minmax-sse-dup.c", |
| 1353 | "src/f32-igemm/gen/1x8-minmax-sse-load1.c", |
| 1354 | "src/f32-igemm/gen/1x8s4-minmax-sse.c", |
| 1355 | "src/f32-igemm/gen/4x2c4-minmax-sse.c", |
| 1356 | "src/f32-igemm/gen/4x8-minmax-sse-dup.c", |
| 1357 | "src/f32-igemm/gen/4x8-minmax-sse-load1.c", |
| 1358 | "src/f32-igemm/gen/4x8s4-minmax-sse.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1359 | "src/f32-maxpool/9p8x-minmax-sse-c4.c", |
| 1360 | "src/f32-pavgpool/9p8x-minmax-sse-c4.c", |
| 1361 | "src/f32-pavgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1362 | "src/f32-ppmm/gen/4x8-minmax-sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1363 | "src/f32-rmax/sse.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1364 | "src/f32-spmm/gen/4x1-minmax-sse.c", |
| 1365 | "src/f32-spmm/gen/8x1-minmax-sse.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 1366 | "src/f32-spmm/gen/16x1-minmax-sse.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1367 | "src/f32-vbinary/gen/vadd-minmax-sse-x4.c", |
| 1368 | "src/f32-vbinary/gen/vadd-minmax-sse-x8.c", |
| 1369 | "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c", |
| 1370 | "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c", |
| 1371 | "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c", |
| 1372 | "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c", |
| 1373 | "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c", |
| 1374 | "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1375 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 1376 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 1377 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 1378 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 1379 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 1380 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 1381 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 1382 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1383 | "src/f32-vbinary/gen/vmul-minmax-sse-x4.c", |
| 1384 | "src/f32-vbinary/gen/vmul-minmax-sse-x8.c", |
| 1385 | "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c", |
| 1386 | "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c", |
| 1387 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c", |
| 1388 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1389 | "src/f32-vbinary/gen/vrsqrdiffc-sse-x4.c", |
| 1390 | "src/f32-vbinary/gen/vrsqrdiffc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1391 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c", |
| 1392 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1393 | "src/f32-vbinary/gen/vsqrdiff-sse-x4.c", |
| 1394 | "src/f32-vbinary/gen/vsqrdiff-sse-x8.c", |
| 1395 | "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c", |
| 1396 | "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1397 | "src/f32-vbinary/gen/vsub-minmax-sse-x4.c", |
| 1398 | "src/f32-vbinary/gen/vsub-minmax-sse-x8.c", |
| 1399 | "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c", |
| 1400 | "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 1401 | "src/f32-vlrelu/gen/sse-x4.c", |
| 1402 | "src/f32-vlrelu/gen/sse-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1403 | "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c", |
| 1404 | "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1405 | "src/f32-vunary/gen/vabs-sse-x4.c", |
| 1406 | "src/f32-vunary/gen/vabs-sse-x8.c", |
| 1407 | "src/f32-vunary/gen/vneg-sse-x4.c", |
| 1408 | "src/f32-vunary/gen/vneg-sse-x8.c", |
| 1409 | "src/f32-vunary/gen/vsqr-sse-x4.c", |
| 1410 | "src/f32-vunary/gen/vsqr-sse-x8.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1411 | "src/x32-fill/sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1412 | "src/x32-packx/x4-sse.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1413 | "src/x32-pad/sse.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1414 | "src/math/roundne-sse-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1415 | "src/math/roundd-sse-addsub.c", |
| 1416 | "src/math/roundu-sse-addsub.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1417 | "src/math/roundz-sse-addsub.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1418 | ] |
| 1419 | |
| 1420 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1421 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 1422 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 1423 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1424 | "src/f32-prelu/gen/sse2-2x4.c", |
| 1425 | "src/f32-prelu/gen/sse2-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1426 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| 1427 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| 1428 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| 1429 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| 1430 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 1431 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| 1432 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| 1433 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 1434 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| 1435 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 1436 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 1437 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1438 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1439 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1440 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1441 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1442 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 1443 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1444 | "src/f32-vrnd/gen/vrndne-sse2-x4.c", |
| 1445 | "src/f32-vrnd/gen/vrndne-sse2-x8.c", |
| 1446 | "src/f32-vrnd/gen/vrndz-sse2-x4.c", |
| 1447 | "src/f32-vrnd/gen/vrndz-sse2-x8.c", |
| 1448 | "src/f32-vrnd/gen/vrndu-sse2-x4.c", |
| 1449 | "src/f32-vrnd/gen/vrndu-sse2-x8.c", |
| 1450 | "src/f32-vrnd/gen/vrndd-sse2-x4.c", |
| 1451 | "src/f32-vrnd/gen/vrndd-sse2-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1452 | "src/q8-avgpool/9p8x-minmax-sse2-c8.c", |
| 1453 | "src/q8-avgpool/9x-minmax-sse2-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1454 | "src/q8-igemm/4x4c2-minmax-sse2.c", |
| 1455 | "src/q8-dwconv/up8x9-minmax-sse2.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1456 | "src/q8-gavgpool/7p7x-minmax-sse2-c8.c", |
| 1457 | "src/q8-gavgpool/7x-minmax-sse2-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1458 | "src/q8-gemm/2x4c8-minmax-sse2.c", |
| 1459 | "src/q8-gemm/4x4c2-minmax-sse2.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1460 | "src/q8-vadd/minmax-sse2.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1461 | "src/u8-clamp/sse2-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1462 | "src/u8-maxpool/9p8x-minmax-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1463 | "src/u8-rmax/sse2.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1464 | "src/x32-unpool/sse2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1465 | "src/x32-zip/x2-sse2.c", |
| 1466 | "src/x32-zip/x3-sse2.c", |
| 1467 | "src/x32-zip/x4-sse2.c", |
| 1468 | "src/x32-zip/xm-sse2.c", |
| 1469 | "src/x8-zip/x2-sse2.c", |
| 1470 | "src/x8-zip/x3-sse2.c", |
| 1471 | "src/x8-zip/x4-sse2.c", |
| 1472 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 1473 | "src/math/exp-sse2-p5.c", |
| 1474 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1475 | "src/math/roundne-sse2-cvt.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1476 | "src/math/roundd-sse2-cvt.c", |
| 1477 | "src/math/roundu-sse2-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1478 | "src/math/roundz-sse2-cvt.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 1479 | "src/math/sigmoid-sse2-p5-div.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1480 | "src/requantization/precise-sse2.c", |
| 1481 | "src/requantization/fp32-sse2.c", |
| 1482 | "src/requantization/q31-sse2.c", |
| 1483 | "src/requantization/gemmlowp-sse2.c", |
| 1484 | ] |
| 1485 | |
| 1486 | SSSE3_UKERNELS = [ |
| 1487 | "src/requantization/precise-ssse3.c", |
| 1488 | "src/requantization/q31-ssse3.c", |
| 1489 | "src/requantization/gemmlowp-ssse3.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1490 | ] |
| 1491 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1492 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1493 | "src/f32-prelu/gen/sse41-2x4.c", |
| 1494 | "src/f32-prelu/gen/sse41-2x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1495 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1496 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1497 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1498 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1499 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 1500 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1501 | "src/f32-vrnd/gen/vrndne-sse41-x4.c", |
| 1502 | "src/f32-vrnd/gen/vrndne-sse41-x8.c", |
| 1503 | "src/f32-vrnd/gen/vrndz-sse41-x4.c", |
| 1504 | "src/f32-vrnd/gen/vrndz-sse41-x8.c", |
| 1505 | "src/f32-vrnd/gen/vrndu-sse41-x4.c", |
| 1506 | "src/f32-vrnd/gen/vrndu-sse41-x8.c", |
| 1507 | "src/f32-vrnd/gen/vrndd-sse41-x4.c", |
| 1508 | "src/f32-vrnd/gen/vrndd-sse41-x8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1509 | "src/math/roundne-sse41.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1510 | "src/math/roundd-sse41.c", |
| 1511 | "src/math/roundu-sse41.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1512 | "src/math/roundz-sse41.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1513 | "src/requantization/precise-sse4.c", |
| 1514 | "src/requantization/q31-sse4.c", |
| 1515 | "src/requantization/gemmlowp-sse4.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1516 | ] |
| 1517 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1518 | AVX_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1519 | "src/f32-clamp/gen/avx-x8.c", |
| 1520 | "src/f32-clamp/gen/avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1521 | "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c", |
| 1522 | "src/f32-dwconv/gen/up16x4-minmax-avx.c", |
| 1523 | "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c", |
| 1524 | "src/f32-dwconv/gen/up8x4-minmax-avx.c", |
| 1525 | "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c", |
| 1526 | "src/f32-dwconv/gen/up16x9-minmax-avx.c", |
| 1527 | "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c", |
| 1528 | "src/f32-dwconv/gen/up8x9-minmax-avx.c", |
| 1529 | "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c", |
| 1530 | "src/f32-dwconv/gen/up16x25-minmax-avx.c", |
| 1531 | "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c", |
| 1532 | "src/f32-dwconv/gen/up8x25-minmax-avx.c", |
| 1533 | "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c", |
| 1534 | "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c", |
| 1535 | "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c", |
| 1536 | "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c", |
| 1537 | "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c", |
| 1538 | "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c", |
| 1539 | "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c", |
| 1540 | "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c", |
| 1541 | "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c", |
| 1542 | "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c", |
| 1543 | "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c", |
| 1544 | "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c", |
| 1545 | "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c", |
| 1546 | "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c", |
| 1547 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c", |
| 1548 | "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c", |
| 1549 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c", |
| 1550 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1551 | "src/f32-hswish/gen/avx-x8.c", |
| 1552 | "src/f32-hswish/gen/avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1553 | "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c", |
| 1554 | "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c", |
| 1555 | "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c", |
| 1556 | "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c", |
| 1557 | "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c", |
| 1558 | "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c", |
| 1559 | "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c", |
| 1560 | "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c", |
| 1561 | "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 1562 | "src/f32-prelu/gen/avx-2x8.c", |
| 1563 | "src/f32-prelu/gen/avx-2x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1564 | "src/f32-rmax/avx.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1565 | "src/f32-vbinary/gen/vadd-minmax-avx-x8.c", |
| 1566 | "src/f32-vbinary/gen/vadd-minmax-avx-x16.c", |
| 1567 | "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c", |
| 1568 | "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c", |
| 1569 | "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c", |
| 1570 | "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c", |
| 1571 | "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c", |
| 1572 | "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1573 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 1574 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 1575 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 1576 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 1577 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 1578 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 1579 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 1580 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1581 | "src/f32-vbinary/gen/vmul-minmax-avx-x8.c", |
| 1582 | "src/f32-vbinary/gen/vmul-minmax-avx-x16.c", |
| 1583 | "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c", |
| 1584 | "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c", |
| 1585 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c", |
| 1586 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1587 | "src/f32-vbinary/gen/vrsqrdiffc-avx-x8.c", |
| 1588 | "src/f32-vbinary/gen/vrsqrdiffc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1589 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c", |
| 1590 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1591 | "src/f32-vbinary/gen/vsqrdiff-avx-x8.c", |
| 1592 | "src/f32-vbinary/gen/vsqrdiff-avx-x16.c", |
| 1593 | "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c", |
| 1594 | "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1595 | "src/f32-vbinary/gen/vsub-minmax-avx-x8.c", |
| 1596 | "src/f32-vbinary/gen/vsub-minmax-avx-x16.c", |
| 1597 | "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c", |
| 1598 | "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 1599 | "src/f32-vlrelu/gen/avx-x8.c", |
| 1600 | "src/f32-vlrelu/gen/avx-x16.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1601 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1602 | "src/f32-vrnd/gen/vrndne-avx-x8.c", |
| 1603 | "src/f32-vrnd/gen/vrndne-avx-x16.c", |
| 1604 | "src/f32-vrnd/gen/vrndz-avx-x8.c", |
| 1605 | "src/f32-vrnd/gen/vrndz-avx-x16.c", |
| 1606 | "src/f32-vrnd/gen/vrndu-avx-x8.c", |
| 1607 | "src/f32-vrnd/gen/vrndu-avx-x16.c", |
| 1608 | "src/f32-vrnd/gen/vrndd-avx-x8.c", |
| 1609 | "src/f32-vrnd/gen/vrndd-avx-x16.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1610 | "src/f32-vunary/gen/vabs-avx-x8.c", |
| 1611 | "src/f32-vunary/gen/vabs-avx-x16.c", |
| 1612 | "src/f32-vunary/gen/vneg-avx-x8.c", |
| 1613 | "src/f32-vunary/gen/vneg-avx-x16.c", |
| 1614 | "src/f32-vunary/gen/vsqr-avx-x8.c", |
| 1615 | "src/f32-vunary/gen/vsqr-avx-x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1616 | ] |
| 1617 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1618 | FMA3_UKERNELS = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1619 | "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c", |
| 1620 | "src/f32-dwconv/gen/up16x4-minmax-fma3.c", |
| 1621 | "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c", |
| 1622 | "src/f32-dwconv/gen/up8x4-minmax-fma3.c", |
| 1623 | "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c", |
| 1624 | "src/f32-dwconv/gen/up16x9-minmax-fma3.c", |
| 1625 | "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c", |
| 1626 | "src/f32-dwconv/gen/up8x9-minmax-fma3.c", |
| 1627 | "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c", |
| 1628 | "src/f32-dwconv/gen/up16x25-minmax-fma3.c", |
| 1629 | "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c", |
| 1630 | "src/f32-dwconv/gen/up8x25-minmax-fma3.c", |
| 1631 | "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c", |
| 1632 | "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c", |
| 1633 | "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c", |
| 1634 | "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c", |
| 1635 | "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c", |
| 1636 | "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c", |
| 1637 | "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c", |
| 1638 | "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c", |
| 1639 | "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c", |
| 1640 | "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c", |
| 1641 | "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 1642 | "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 1643 | "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 1644 | "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| 1645 | "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c", |
| 1646 | "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c", |
| 1647 | "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c", |
| 1648 | "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c", |
| 1649 | "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c", |
| 1650 | "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c", |
| 1651 | "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c", |
| 1652 | "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c", |
| 1653 | "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c", |
| 1654 | "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c", |
| 1655 | "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c", |
| 1656 | "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c", |
| 1657 | "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c", |
| 1658 | "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1659 | "src/f32-hswish/gen/fma3-x8.c", |
| 1660 | "src/f32-hswish/gen/fma3-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1661 | "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c", |
| 1662 | "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c", |
| 1663 | "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c", |
| 1664 | "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c", |
| 1665 | "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c", |
| 1666 | "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c", |
| 1667 | "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c", |
| 1668 | "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c", |
| 1669 | "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c", |
| 1670 | "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c", |
| 1671 | "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 1672 | "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 1673 | "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 1674 | "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1675 | ] |
| 1676 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1677 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1678 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 1679 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1680 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1681 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 1682 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1683 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 1684 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1685 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1686 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 1687 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1688 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1689 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 1690 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 1691 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 1692 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 1693 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 1694 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 1695 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 1696 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 1697 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 1698 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 1699 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 1700 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 1701 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 1702 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 1703 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1704 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1705 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 1706 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1707 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 1708 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1709 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1710 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 1711 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1712 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1713 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1714 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 1715 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 1716 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 1717 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 1718 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 1719 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 1720 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 1721 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 1722 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 1723 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 1724 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 1725 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 1726 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 1727 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 1728 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 1729 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 1730 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 1731 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 1732 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 1733 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 1734 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 1735 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 1736 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 1737 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 1738 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 1739 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 1740 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 1741 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 1742 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 1743 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1744 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 1745 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 1746 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 1747 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 1748 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 1749 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 1750 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 1751 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 1752 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 1753 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 1754 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 1755 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 1756 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 1757 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 1758 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 1759 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 1760 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 1761 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 1762 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 1763 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 1764 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 1765 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 1766 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 1767 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1768 | "src/math/exp-avx2-p5.c", |
| 1769 | "src/math/exp-avx2-perm-p3.c", |
| 1770 | "src/math/exp-avx2-perm-p4.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 1771 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1772 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 1773 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 1774 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 1775 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| 1776 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 1777 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 1778 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1779 | ] |
| 1780 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1781 | AVX512F_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1782 | "src/f32-clamp/gen/avx512f-x16.c", |
| 1783 | "src/f32-clamp/gen/avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1784 | "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c", |
| 1785 | "src/f32-dwconv/gen/up32x4-minmax-avx512f.c", |
| 1786 | "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c", |
| 1787 | "src/f32-dwconv/gen/up16x4-minmax-avx512f.c", |
| 1788 | "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c", |
| 1789 | "src/f32-dwconv/gen/up32x9-minmax-avx512f.c", |
| 1790 | "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c", |
| 1791 | "src/f32-dwconv/gen/up16x9-minmax-avx512f.c", |
| 1792 | "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c", |
| 1793 | "src/f32-dwconv/gen/up32x25-minmax-avx512f.c", |
| 1794 | "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c", |
| 1795 | "src/f32-dwconv/gen/up16x25-minmax-avx512f.c", |
| 1796 | "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 1797 | "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 1798 | "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 1799 | "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 1800 | "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 1801 | "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c", |
| 1802 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c", |
| 1803 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c", |
| 1804 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c", |
| 1805 | "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c", |
| 1806 | "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c", |
| 1807 | "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1808 | "src/f32-hswish/gen/avx512f-x16.c", |
| 1809 | "src/f32-hswish/gen/avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1810 | "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 1811 | "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 1812 | "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 1813 | "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 1814 | "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 1815 | "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 1816 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 1817 | "src/f32-prelu/gen/avx512f-2x32.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1818 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1819 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1820 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1821 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1822 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1823 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1824 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1825 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1826 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1827 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1828 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1829 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1830 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 1831 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1832 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1833 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 1834 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1835 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 1836 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1837 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1838 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 1839 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1840 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1841 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 1842 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1843 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 1844 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 1845 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1846 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 1847 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1848 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 1849 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 1850 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1851 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 1852 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 1853 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1854 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1855 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c", |
| 1856 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c", |
| 1857 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c", |
| 1858 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c", |
| 1859 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c", |
| 1860 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c", |
| 1861 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c", |
| 1862 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1863 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 1864 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 1865 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 1866 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 1867 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 1868 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 1869 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 1870 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1871 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c", |
| 1872 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c", |
| 1873 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c", |
| 1874 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c", |
| 1875 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c", |
| 1876 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1877 | "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x16.c", |
| 1878 | "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1879 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c", |
| 1880 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1881 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c", |
| 1882 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c", |
| 1883 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c", |
| 1884 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1885 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c", |
| 1886 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c", |
| 1887 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c", |
| 1888 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 1889 | "src/f32-vlrelu/gen/avx512f-x16.c", |
| 1890 | "src/f32-vlrelu/gen/avx512f-x32.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1891 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1892 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 1893 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 1894 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 1895 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 1896 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 1897 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 1898 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 1899 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 1900 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 1901 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 1902 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 1903 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 1904 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 1905 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 1906 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 1907 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 1908 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 1909 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 1910 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 1911 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 1912 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 1913 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 1914 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 1915 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1916 | "src/f32-vrnd/gen/vrndne-avx512f-x16.c", |
| 1917 | "src/f32-vrnd/gen/vrndne-avx512f-x32.c", |
| 1918 | "src/f32-vrnd/gen/vrndz-avx512f-x16.c", |
| 1919 | "src/f32-vrnd/gen/vrndz-avx512f-x32.c", |
| 1920 | "src/f32-vrnd/gen/vrndu-avx512f-x16.c", |
| 1921 | "src/f32-vrnd/gen/vrndu-avx512f-x32.c", |
| 1922 | "src/f32-vrnd/gen/vrndd-avx512f-x16.c", |
| 1923 | "src/f32-vrnd/gen/vrndd-avx512f-x32.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1924 | "src/f32-vunary/gen/vabs-avx512f-x16.c", |
| 1925 | "src/f32-vunary/gen/vabs-avx512f-x32.c", |
| 1926 | "src/f32-vunary/gen/vneg-avx512f-x16.c", |
| 1927 | "src/f32-vunary/gen/vneg-avx512f-x32.c", |
| 1928 | "src/f32-vunary/gen/vsqr-avx512f-x16.c", |
| 1929 | "src/f32-vunary/gen/vsqr-avx512f-x32.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1930 | "src/math/exp-avx512f-p5-scalef.c", |
| 1931 | "src/math/exp-avx512f-p5.c", |
| 1932 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | feb4923 | 2019-10-28 11:03:31 -0700 | [diff] [blame] | 1933 | "src/math/exp-avx512f-perm2-p2.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1934 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1935 | ] |
| 1936 | |
| 1937 | AARCH32_ASM_UKERNELS = [ |
Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 1938 | "src/f32-gemm/4x4-aarch32-vfp-ld64.S", |
Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 1939 | "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1940 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 1941 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
| 1942 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 1943 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
| 1944 | "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S", |
| 1945 | "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S", |
| 1946 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 1947 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
| 1948 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 1949 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1950 | ] |
| 1951 | |
| 1952 | AARCH64_ASM_UKERNELS = [ |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 1953 | "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 1954 | "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 1955 | "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 1956 | "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 1957 | "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
| 1958 | "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 1959 | "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 1960 | "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 1961 | "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 1962 | "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 1963 | "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| 1964 | "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| 1965 | "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 1966 | "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1967 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S", |
| 1968 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S", |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 1969 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1970 | "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 1971 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 1972 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 1973 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 1974 | "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 1975 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 1976 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 1977 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 1978 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 1979 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S", |
| 1980 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S", |
| 1981 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 1982 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 1983 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 1984 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 1985 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 1986 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 1987 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1988 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S", |
| 1989 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S", |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 1990 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1991 | "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 1992 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 1993 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 1994 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 1995 | "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 1996 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 1997 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 1998 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 1999 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2000 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2001 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2002 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2003 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2004 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2005 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2006 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S", |
| 2007 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2008 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2009 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2010 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2011 | "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2012 | "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2013 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2014 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2015 | "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2016 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2017 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2018 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2019 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2020 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2021 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2022 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2023 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2024 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2025 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2026 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2027 | ] |
| 2028 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2029 | INTERNAL_MICROKERNEL_HDRS = [ |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2030 | "src/requantization/gemmlowp-requantization.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2031 | "src/xnnpack/argmaxpool.h", |
| 2032 | "src/xnnpack/avgpool.h", |
| 2033 | "src/xnnpack/clamp.h", |
| 2034 | "src/xnnpack/common.h", |
| 2035 | "src/xnnpack/conv.h", |
| 2036 | "src/xnnpack/dwconv.h", |
| 2037 | "src/xnnpack/gavgpool.h", |
| 2038 | "src/xnnpack/gemm.h", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 2039 | "src/xnnpack/fill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2040 | "src/xnnpack/hswish.h", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2041 | "src/xnnpack/ibilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2042 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 2043 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2044 | "src/xnnpack/lut.h", |
| 2045 | "src/xnnpack/math.h", |
| 2046 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2047 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2048 | "src/xnnpack/packx.h", |
| 2049 | "src/xnnpack/pad.h", |
| 2050 | "src/xnnpack/params.h", |
| 2051 | "src/xnnpack/pavgpool.h", |
| 2052 | "src/xnnpack/ppmm.h", |
| 2053 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2054 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2055 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2056 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2057 | "src/xnnpack/rmax.h", |
| 2058 | "src/xnnpack/scalar-utils.h", |
| 2059 | "src/xnnpack/spmm.h", |
| 2060 | "src/xnnpack/unpool.h", |
| 2061 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2062 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2063 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2064 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2065 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2066 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2067 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2068 | "src/xnnpack/zip.h", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2069 | ] |
| 2070 | |
| 2071 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2072 | "include/xnnpack.h", |
| 2073 | "src/xnnpack/allocator.h", |
| 2074 | "src/xnnpack/compute.h", |
| 2075 | "src/xnnpack/im2col.h", |
| 2076 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2077 | "src/xnnpack/math-stubs.h", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 2078 | "src/xnnpack/memory-planner.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2079 | "src/xnnpack/operator.h", |
| 2080 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2081 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2082 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2083 | "src/xnnpack/requantization.h", |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 2084 | "src/xnnpack/subgraph.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2085 | ] |
| 2086 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2087 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2088 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2089 | ] |
| 2090 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2091 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2092 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2093 | "include/xnnpack.h", |
| 2094 | ] |
| 2095 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2096 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2097 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2098 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2099 | "src/xnnpack/requantization.h", |
| 2100 | "include/xnnpack.h", |
| 2101 | ] |
| 2102 | |
| 2103 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 2104 | "src/xnnpack/params.h", |
| 2105 | "src/xnnpack/common.h", |
| 2106 | ] |
| 2107 | |
| 2108 | WEIGHTS_PACK_HDRS = [ |
| 2109 | "src/xnnpack/pack.h", |
| 2110 | "src/xnnpack/operator.h", |
| 2111 | "src/xnnpack/compute.h", |
| 2112 | ] |
| 2113 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2114 | LOGGING_COPTS = select({ |
| 2115 | # No logging in optimized mode |
| 2116 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 2117 | # Full logging in debug mode |
| 2118 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 2119 | # Error-only logging in default (fastbuild) mode |
| 2120 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 2121 | }) |
| 2122 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2123 | LOGGING_SRCS = select({ |
| 2124 | # No logging in optimized mode |
| 2125 | ":optimized_build": [], |
| 2126 | "//conditions:default": [ |
| 2127 | "src/operator-strings.c", |
| 2128 | "src/subgraph-strings.c", |
| 2129 | ], |
| 2130 | }) |
| 2131 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2132 | LOGGING_HDRS = [ |
| 2133 | "src/xnnpack/log.h", |
| 2134 | ] |
| 2135 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2136 | xnnpack_cc_library( |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2137 | name = "tables", |
| 2138 | srcs = TABLE_SRCS, |
| 2139 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2140 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2141 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2142 | ) |
| 2143 | |
| 2144 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2145 | name = "scalar_ukernels", |
| 2146 | srcs = SCALAR_UKERNELS, |
| 2147 | hdrs = INTERNAL_HDRS, |
| 2148 | aarch32_copts = ["-marm"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2149 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2150 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2151 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2152 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2153 | "@FP16", |
| 2154 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2155 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2156 | ], |
| 2157 | ) |
| 2158 | |
| 2159 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2160 | name = "scalar_ukernels_test_mode", |
| 2161 | srcs = SCALAR_UKERNELS, |
| 2162 | hdrs = INTERNAL_HDRS, |
| 2163 | aarch32_copts = ["-marm"], |
| 2164 | copts = [ |
| 2165 | "-UNDEBUG", |
| 2166 | "-DXNN_TEST_MODE=1", |
| 2167 | ], |
| 2168 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2169 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2170 | deps = [ |
| 2171 | ":tables", |
| 2172 | "@FP16", |
| 2173 | "@FXdiv", |
| 2174 | "@pthreadpool", |
| 2175 | ], |
| 2176 | ) |
| 2177 | |
| 2178 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2179 | name = "wasm_ukernels", |
| 2180 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2181 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2182 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2183 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 2184 | wasmsimd_copts = [ |
| 2185 | "-msimd128", |
| 2186 | "-munimplemented-simd128", |
| 2187 | ], |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 2188 | wasmsimd_srcs = WASMSIMD_UKERNELS, |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2189 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2190 | ":tables", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2191 | "@FP16", |
| 2192 | "@FXdiv", |
| 2193 | "@pthreadpool", |
| 2194 | ], |
| 2195 | ) |
| 2196 | |
| 2197 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2198 | name = "wasm_ukernels_test_mode", |
| 2199 | hdrs = INTERNAL_HDRS, |
| 2200 | copts = [ |
| 2201 | "-UNDEBUG", |
| 2202 | "-DXNN_TEST_MODE=1", |
| 2203 | ], |
| 2204 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2205 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2206 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 2207 | wasmsimd_copts = [ |
| 2208 | "-msimd128", |
| 2209 | "-munimplemented-simd128", |
| 2210 | ], |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 2211 | wasmsimd_srcs = WASMSIMD_UKERNELS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2212 | deps = [ |
| 2213 | ":tables", |
| 2214 | "@FP16", |
| 2215 | "@FXdiv", |
| 2216 | "@pthreadpool", |
| 2217 | ], |
| 2218 | ) |
| 2219 | |
| 2220 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2221 | name = "psimd_fastmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2222 | hdrs = INTERNAL_HDRS, |
| 2223 | aarch32_copts = [ |
| 2224 | "-marm", |
| 2225 | "-mfpu=neon", |
| 2226 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2227 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2228 | gcc_x86_copts = ["-msse2"], |
| 2229 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2230 | optimized_copts = [ |
| 2231 | "-O3", |
| 2232 | "-ffast-math", |
| 2233 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 2234 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2235 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2236 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2237 | "@FP16", |
| 2238 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2239 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2240 | ], |
| 2241 | ) |
| 2242 | |
| 2243 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2244 | name = "psimd_fastmath_ukernels_test_mode", |
| 2245 | hdrs = INTERNAL_HDRS, |
| 2246 | aarch32_copts = [ |
| 2247 | "-marm", |
| 2248 | "-mfpu=neon", |
| 2249 | ], |
| 2250 | copts = [ |
| 2251 | "-UNDEBUG", |
| 2252 | "-DXNN_TEST_MODE=1", |
| 2253 | ], |
| 2254 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2255 | gcc_x86_copts = ["-msse2"], |
| 2256 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2257 | optimized_copts = [ |
| 2258 | "-O3", |
| 2259 | "-ffast-math", |
| 2260 | ], |
| 2261 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
| 2262 | deps = [ |
| 2263 | ":tables", |
| 2264 | "@FP16", |
| 2265 | "@psimd", |
| 2266 | "@pthreadpool", |
| 2267 | ], |
| 2268 | ) |
| 2269 | |
| 2270 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2271 | name = "psimd_accmath_ukernels", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2272 | hdrs = INTERNAL_HDRS, |
| 2273 | aarch32_copts = [ |
| 2274 | "-marm", |
| 2275 | "-mfpu=neon", |
| 2276 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2277 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2278 | gcc_x86_copts = ["-msse2"], |
| 2279 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2280 | optimized_copts = [ |
| 2281 | "-O3", |
| 2282 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 2283 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2284 | deps = [ |
| 2285 | ":tables", |
| 2286 | "@FP16", |
| 2287 | "@psimd", |
| 2288 | "@pthreadpool", |
| 2289 | ], |
| 2290 | ) |
| 2291 | |
| 2292 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2293 | name = "psimd_accmath_ukernels_test_mode", |
| 2294 | hdrs = INTERNAL_HDRS, |
| 2295 | aarch32_copts = [ |
| 2296 | "-marm", |
| 2297 | "-mfpu=neon", |
| 2298 | ], |
| 2299 | copts = [ |
| 2300 | "-UNDEBUG", |
| 2301 | "-DXNN_TEST_MODE=1", |
| 2302 | ], |
| 2303 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2304 | gcc_x86_copts = ["-msse2"], |
| 2305 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2306 | optimized_copts = [ |
| 2307 | "-O3", |
| 2308 | ], |
| 2309 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
| 2310 | deps = [ |
| 2311 | ":tables", |
| 2312 | "@FP16", |
| 2313 | "@psimd", |
| 2314 | "@pthreadpool", |
| 2315 | ], |
| 2316 | ) |
| 2317 | |
| 2318 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2319 | name = "neon_ukernels", |
| 2320 | hdrs = INTERNAL_HDRS, |
| 2321 | aarch32_copts = [ |
| 2322 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2323 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2324 | "-mfpu=neon", |
| 2325 | ], |
| 2326 | aarch32_srcs = NEON_UKERNELS, |
| 2327 | aarch64_srcs = NEON_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2328 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2329 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2330 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2331 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2332 | "@FP16", |
| 2333 | "@pthreadpool", |
| 2334 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2335 | ) |
| 2336 | |
| 2337 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2338 | name = "neon_ukernels_test_mode", |
| 2339 | hdrs = INTERNAL_HDRS, |
| 2340 | aarch32_copts = [ |
| 2341 | "-marm", |
| 2342 | "-march=armv7-a", |
| 2343 | "-mfpu=neon", |
| 2344 | ], |
| 2345 | aarch32_srcs = NEON_UKERNELS, |
| 2346 | aarch64_srcs = NEON_UKERNELS, |
| 2347 | copts = [ |
| 2348 | "-UNDEBUG", |
| 2349 | "-DXNN_TEST_MODE=1", |
| 2350 | ], |
| 2351 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2352 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2353 | deps = [ |
| 2354 | ":tables", |
| 2355 | "@FP16", |
| 2356 | "@pthreadpool", |
| 2357 | ], |
| 2358 | ) |
| 2359 | |
| 2360 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2361 | name = "neonfma_ukernels", |
| 2362 | hdrs = INTERNAL_HDRS, |
| 2363 | aarch32_copts = [ |
| 2364 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2365 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2366 | "-mfpu=neon-vfpv4", |
| 2367 | ], |
| 2368 | aarch32_srcs = NEONFMA_UKERNELS, |
| 2369 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2370 | apple_aarch32_copts = [ |
| 2371 | "-mcpu=swift", |
| 2372 | "-mtune=generic", |
| 2373 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2374 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2375 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2376 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2377 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2378 | "@FP16", |
| 2379 | "@pthreadpool", |
| 2380 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2381 | ) |
| 2382 | |
| 2383 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2384 | name = "neonfma_ukernels_test_mode", |
| 2385 | hdrs = INTERNAL_HDRS, |
| 2386 | aarch32_copts = [ |
| 2387 | "-marm", |
| 2388 | "-march=armv7-a", |
| 2389 | "-mfpu=neon-vfpv4", |
| 2390 | ], |
| 2391 | aarch32_srcs = NEONFMA_UKERNELS, |
| 2392 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2393 | apple_aarch32_copts = [ |
| 2394 | "-mcpu=swift", |
| 2395 | "-mtune=generic", |
| 2396 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2397 | copts = [ |
| 2398 | "-UNDEBUG", |
| 2399 | "-DXNN_TEST_MODE=1", |
| 2400 | ], |
| 2401 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2402 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2403 | deps = [ |
| 2404 | ":tables", |
| 2405 | "@FP16", |
| 2406 | "@pthreadpool", |
| 2407 | ], |
| 2408 | ) |
| 2409 | |
| 2410 | xnnpack_cc_library( |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2411 | name = "neonv8_ukernels", |
| 2412 | hdrs = INTERNAL_HDRS, |
| 2413 | aarch32_copts = [ |
| 2414 | "-marm", |
| 2415 | "-march=armv8-a", |
| 2416 | "-mfpu=neon-fp-armv8", |
| 2417 | ], |
| 2418 | aarch32_srcs = NEONV8_UKERNELS, |
| 2419 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2420 | apple_aarch32_copts = [ |
| 2421 | "-mcpu=cyclone", |
| 2422 | "-mtune=generic", |
| 2423 | ], |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2424 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2425 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2426 | deps = [ |
| 2427 | ":tables", |
| 2428 | "@FP16", |
| 2429 | "@pthreadpool", |
| 2430 | ], |
| 2431 | ) |
| 2432 | |
| 2433 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2434 | name = "neonv8_ukernels_test_mode", |
| 2435 | hdrs = INTERNAL_HDRS, |
| 2436 | aarch32_copts = [ |
| 2437 | "-marm", |
| 2438 | "-march=armv8-a", |
| 2439 | "-mfpu=neon-fp-armv8", |
| 2440 | ], |
| 2441 | aarch32_srcs = NEONV8_UKERNELS, |
| 2442 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2443 | apple_aarch32_copts = [ |
| 2444 | "-mcpu=cyclone", |
| 2445 | "-mtune=generic", |
| 2446 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2447 | copts = [ |
| 2448 | "-UNDEBUG", |
| 2449 | "-DXNN_TEST_MODE=1", |
| 2450 | ], |
| 2451 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2452 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2453 | deps = [ |
| 2454 | ":tables", |
| 2455 | "@FP16", |
| 2456 | "@pthreadpool", |
| 2457 | ], |
| 2458 | ) |
| 2459 | |
| 2460 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2461 | name = "neonfp16arith_ukernels", |
| 2462 | hdrs = INTERNAL_HDRS, |
| 2463 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 2464 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2465 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2466 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2467 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2468 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2469 | "@FP16", |
| 2470 | "@pthreadpool", |
| 2471 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2472 | ) |
| 2473 | |
| 2474 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2475 | name = "neonfp16arith_ukernels_test_mode", |
| 2476 | hdrs = INTERNAL_HDRS, |
| 2477 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 2478 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 2479 | copts = [ |
| 2480 | "-UNDEBUG", |
| 2481 | "-DXNN_TEST_MODE=1", |
| 2482 | ], |
| 2483 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2484 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2485 | deps = [ |
| 2486 | ":tables", |
| 2487 | "@FP16", |
| 2488 | "@pthreadpool", |
| 2489 | ], |
| 2490 | ) |
| 2491 | |
| 2492 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2493 | name = "sse2_ukernels", |
| 2494 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2495 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2496 | gcc_x86_copts = ["-msse2"], |
| 2497 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2498 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2499 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2500 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2501 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2502 | "@FP16", |
| 2503 | "@pthreadpool", |
| 2504 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2505 | ) |
| 2506 | |
| 2507 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2508 | name = "sse2_ukernels_test_mode", |
| 2509 | hdrs = INTERNAL_HDRS, |
| 2510 | copts = [ |
| 2511 | "-UNDEBUG", |
| 2512 | "-DXNN_TEST_MODE=1", |
| 2513 | ], |
| 2514 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2515 | gcc_x86_copts = ["-msse2"], |
| 2516 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2517 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 2518 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| 2519 | deps = [ |
| 2520 | ":tables", |
| 2521 | "@FP16", |
| 2522 | "@pthreadpool", |
| 2523 | ], |
| 2524 | ) |
| 2525 | |
| 2526 | xnnpack_cc_library( |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2527 | name = "ssse3_ukernels", |
| 2528 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2529 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2530 | gcc_x86_copts = ["-mssse3"], |
| 2531 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2532 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2533 | x86_srcs = SSSE3_UKERNELS, |
| 2534 | deps = [ |
| 2535 | ":tables", |
| 2536 | "@FP16", |
| 2537 | "@pthreadpool", |
| 2538 | ], |
| 2539 | ) |
| 2540 | |
| 2541 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2542 | name = "ssse3_ukernels_test_mode", |
| 2543 | hdrs = INTERNAL_HDRS, |
| 2544 | copts = [ |
| 2545 | "-UNDEBUG", |
| 2546 | "-DXNN_TEST_MODE=1", |
| 2547 | ], |
| 2548 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2549 | gcc_x86_copts = ["-mssse3"], |
| 2550 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2551 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 2552 | x86_srcs = SSSE3_UKERNELS, |
| 2553 | deps = [ |
| 2554 | ":tables", |
| 2555 | "@FP16", |
| 2556 | "@pthreadpool", |
| 2557 | ], |
| 2558 | ) |
| 2559 | |
| 2560 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2561 | name = "sse41_ukernels", |
| 2562 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2563 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2564 | gcc_x86_copts = ["-msse4.1"], |
| 2565 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2566 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2567 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2568 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2569 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2570 | "@FP16", |
| 2571 | "@pthreadpool", |
| 2572 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2573 | ) |
| 2574 | |
| 2575 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2576 | name = "sse41_ukernels_test_mode", |
| 2577 | hdrs = INTERNAL_HDRS, |
| 2578 | copts = [ |
| 2579 | "-UNDEBUG", |
| 2580 | "-DXNN_TEST_MODE=1", |
| 2581 | ], |
| 2582 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2583 | gcc_x86_copts = ["-msse4.1"], |
| 2584 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2585 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 2586 | x86_srcs = SSE41_UKERNELS, |
| 2587 | deps = [ |
| 2588 | ":tables", |
| 2589 | "@FP16", |
| 2590 | "@pthreadpool", |
| 2591 | ], |
| 2592 | ) |
| 2593 | |
| 2594 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2595 | name = "avx_ukernels", |
| 2596 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2597 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2598 | gcc_x86_copts = ["-mavx"], |
| 2599 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2600 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2601 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2602 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2603 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2604 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2605 | "@FP16", |
| 2606 | "@pthreadpool", |
| 2607 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2608 | ) |
| 2609 | |
| 2610 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2611 | name = "avx_ukernels_test_mode", |
| 2612 | hdrs = INTERNAL_HDRS, |
| 2613 | copts = [ |
| 2614 | "-UNDEBUG", |
| 2615 | "-DXNN_TEST_MODE=1", |
| 2616 | ], |
| 2617 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2618 | gcc_x86_copts = ["-mavx"], |
| 2619 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2620 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2621 | msvc_x86_64_copts = ["/arch:AVX"], |
| 2622 | x86_srcs = AVX_UKERNELS, |
| 2623 | deps = [ |
| 2624 | ":tables", |
| 2625 | "@FP16", |
| 2626 | "@pthreadpool", |
| 2627 | ], |
| 2628 | ) |
| 2629 | |
| 2630 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2631 | name = "fma3_ukernels", |
| 2632 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2633 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2634 | gcc_x86_copts = ["-mfma"], |
| 2635 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2636 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2637 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2638 | x86_srcs = FMA3_UKERNELS, |
| 2639 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2640 | ":tables", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2641 | "@FP16", |
| 2642 | "@pthreadpool", |
| 2643 | ], |
| 2644 | ) |
| 2645 | |
| 2646 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2647 | name = "fma3_ukernels_test_mode", |
| 2648 | hdrs = INTERNAL_HDRS, |
| 2649 | copts = [ |
| 2650 | "-UNDEBUG", |
| 2651 | "-DXNN_TEST_MODE=1", |
| 2652 | ], |
| 2653 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2654 | gcc_x86_copts = ["-mfma"], |
| 2655 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2656 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2657 | msvc_x86_64_copts = ["/arch:AVX"], |
| 2658 | x86_srcs = FMA3_UKERNELS, |
| 2659 | deps = [ |
| 2660 | ":tables", |
| 2661 | "@FP16", |
| 2662 | "@pthreadpool", |
| 2663 | ], |
| 2664 | ) |
| 2665 | |
| 2666 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2667 | name = "avx2_ukernels", |
| 2668 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2669 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2670 | gcc_x86_copts = [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2671 | "-mfma", |
| 2672 | "-mavx2", |
| 2673 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2674 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2675 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 2676 | msvc_x86_64_copts = ["/arch:AVX2"], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2677 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2678 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2679 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2680 | "@FP16", |
| 2681 | "@pthreadpool", |
| 2682 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2683 | ) |
| 2684 | |
| 2685 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2686 | name = "avx2_ukernels_test_mode", |
| 2687 | hdrs = INTERNAL_HDRS, |
| 2688 | copts = [ |
| 2689 | "-UNDEBUG", |
| 2690 | "-DXNN_TEST_MODE=1", |
| 2691 | ], |
| 2692 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2693 | gcc_x86_copts = [ |
| 2694 | "-mfma", |
| 2695 | "-mavx2", |
| 2696 | ], |
| 2697 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2698 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 2699 | msvc_x86_64_copts = ["/arch:AVX2"], |
| 2700 | x86_srcs = AVX2_UKERNELS, |
| 2701 | deps = [ |
| 2702 | ":tables", |
| 2703 | "@FP16", |
| 2704 | "@pthreadpool", |
| 2705 | ], |
| 2706 | ) |
| 2707 | |
| 2708 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2709 | name = "avx512f_ukernels", |
| 2710 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2711 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2712 | gcc_x86_copts = ["-mavx512f"], |
| 2713 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 2714 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2715 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 2716 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 2717 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2718 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2719 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2720 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2721 | "@FP16", |
| 2722 | "@pthreadpool", |
| 2723 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2724 | ) |
| 2725 | |
| 2726 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2727 | name = "avx512f_ukernels_test_mode", |
| 2728 | hdrs = INTERNAL_HDRS, |
| 2729 | copts = [ |
| 2730 | "-UNDEBUG", |
| 2731 | "-DXNN_TEST_MODE=1", |
| 2732 | ], |
| 2733 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2734 | gcc_x86_copts = ["-mavx512f"], |
| 2735 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 2736 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2737 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 2738 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 2739 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 2740 | x86_srcs = AVX512F_UKERNELS, |
| 2741 | deps = [ |
| 2742 | ":tables", |
| 2743 | "@FP16", |
| 2744 | "@pthreadpool", |
| 2745 | ], |
| 2746 | ) |
| 2747 | |
| 2748 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2749 | name = "asm_ukernels", |
| 2750 | hdrs = ["src/xnnpack/assembly.h"], |
| 2751 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 2752 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2753 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 2754 | ) |
| 2755 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2756 | xnnpack_cc_library( |
| 2757 | name = "logging_utils", |
| 2758 | srcs = LOGGING_SRCS, |
| 2759 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 2760 | copts = LOGGING_COPTS + [ |
| 2761 | "-Isrc", |
| 2762 | "-Iinclude", |
| 2763 | ] + select({ |
| 2764 | ":debug_build": [], |
| 2765 | "//conditions:default": xnnpack_min_size_copts(), |
| 2766 | }), |
| 2767 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2768 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2769 | visibility = xnnpack_visibility(), |
| 2770 | deps = [ |
| 2771 | "@FP16", |
| 2772 | "@clog", |
| 2773 | "@pthreadpool", |
| 2774 | ], |
| 2775 | ) |
| 2776 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2777 | xnnpack_aggregate_library( |
| 2778 | name = "ukernels", |
| 2779 | aarch32_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2780 | ":neon_ukernels", |
| 2781 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2782 | ":neonv8_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2783 | ":asm_ukernels", |
| 2784 | ], |
| 2785 | aarch64_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2786 | ":neon_ukernels", |
| 2787 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2788 | ":neonv8_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2789 | ":neonfp16arith_ukernels", |
| 2790 | ":asm_ukernels", |
| 2791 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2792 | generic_deps = [ |
| 2793 | ":scalar_ukernels", |
| 2794 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 2795 | psimd_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2796 | ":psimd_fastmath_ukernels", |
| 2797 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2798 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2799 | wasm_deps = [ |
| 2800 | ":wasm_ukernels", |
| 2801 | ], |
| 2802 | wasmsimd_deps = [ |
| 2803 | ":wasm_ukernels", |
| 2804 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2805 | x86_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2806 | ":sse2_ukernels", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2807 | ":ssse3_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2808 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2809 | ":avx_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2810 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2811 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2812 | ":avx512f_ukernels", |
| 2813 | ], |
| 2814 | ) |
| 2815 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2816 | xnnpack_aggregate_library( |
| 2817 | name = "ukernels_test_mode", |
| 2818 | aarch32_deps = [ |
| 2819 | ":neon_ukernels_test_mode", |
| 2820 | ":neonfma_ukernels_test_mode", |
| 2821 | ":neonv8_ukernels_test_mode", |
| 2822 | ":asm_ukernels", |
| 2823 | ], |
| 2824 | aarch64_deps = [ |
| 2825 | ":neon_ukernels_test_mode", |
| 2826 | ":neonfma_ukernels_test_mode", |
| 2827 | ":neonv8_ukernels_test_mode", |
| 2828 | ":neonfp16arith_ukernels_test_mode", |
| 2829 | ":asm_ukernels", |
| 2830 | ], |
| 2831 | generic_deps = [ |
| 2832 | ":scalar_ukernels_test_mode", |
| 2833 | ], |
| 2834 | psimd_deps = [ |
| 2835 | ":psimd_fastmath_ukernels_test_mode", |
| 2836 | ":psimd_accmath_ukernels_test_mode", |
| 2837 | ], |
| 2838 | wasm_deps = [ |
| 2839 | ":wasm_ukernels_test_mode", |
| 2840 | ], |
| 2841 | wasmsimd_deps = [ |
| 2842 | ":wasm_ukernels_test_mode", |
| 2843 | ], |
| 2844 | x86_deps = [ |
| 2845 | ":sse2_ukernels_test_mode", |
| 2846 | ":ssse3_ukernels_test_mode", |
| 2847 | ":sse41_ukernels_test_mode", |
| 2848 | ":avx_ukernels_test_mode", |
| 2849 | ":fma3_ukernels_test_mode", |
| 2850 | ":avx2_ukernels_test_mode", |
| 2851 | ":avx512f_ukernels_test_mode", |
| 2852 | ], |
| 2853 | ) |
| 2854 | |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 2855 | xnnpack_aggregate_library( |
| 2856 | name = "ukernels_no_f16", |
| 2857 | aarch32_deps = [ |
| 2858 | ":neon_ukernels", |
| 2859 | ":neonfma_ukernels", |
| 2860 | ":neonv8_ukernels", |
| 2861 | ":asm_ukernels", |
| 2862 | ], |
| 2863 | aarch64_deps = [ |
| 2864 | ":neon_ukernels", |
| 2865 | ":neonfma_ukernels", |
| 2866 | ":neonv8_ukernels", |
| 2867 | ":asm_ukernels", |
| 2868 | ], |
| 2869 | generic_deps = [ |
| 2870 | ":scalar_ukernels", |
| 2871 | ], |
| 2872 | psimd_deps = [ |
| 2873 | ":psimd_fastmath_ukernels", |
| 2874 | ":psimd_accmath_ukernels", |
| 2875 | ], |
| 2876 | wasm_deps = [ |
| 2877 | ":wasm_ukernels", |
| 2878 | ], |
| 2879 | wasmsimd_deps = [ |
| 2880 | ":wasm_ukernels", |
| 2881 | ], |
| 2882 | x86_deps = [ |
| 2883 | ":sse2_ukernels", |
| 2884 | ":ssse3_ukernels", |
| 2885 | ":sse41_ukernels", |
| 2886 | ":avx_ukernels", |
| 2887 | ":fma3_ukernels", |
| 2888 | ":avx2_ukernels", |
| 2889 | ":avx512f_ukernels", |
| 2890 | ], |
| 2891 | ) |
| 2892 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2893 | xnnpack_cc_library( |
| 2894 | name = "im2col", |
| 2895 | srcs = ["src/im2col.c"], |
| 2896 | hdrs = [ |
| 2897 | "src/xnnpack/common.h", |
| 2898 | "src/xnnpack/im2col.h", |
| 2899 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2900 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2901 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2902 | ) |
| 2903 | |
| 2904 | xnnpack_cc_library( |
| 2905 | name = "indirection", |
| 2906 | srcs = ["src/indirection.c"], |
| 2907 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2908 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2909 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2910 | deps = [ |
| 2911 | "@FP16", |
| 2912 | "@FXdiv", |
| 2913 | "@pthreadpool", |
| 2914 | ], |
| 2915 | ) |
| 2916 | |
| 2917 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2918 | name = "indirection_test_mode", |
| 2919 | srcs = ["src/indirection.c"], |
| 2920 | hdrs = INTERNAL_HDRS, |
| 2921 | copts = [ |
| 2922 | "-UNDEBUG", |
| 2923 | "-DXNN_TEST_MODE=1", |
| 2924 | ], |
| 2925 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2926 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2927 | deps = [ |
| 2928 | "@FP16", |
| 2929 | "@FXdiv", |
| 2930 | "@pthreadpool", |
| 2931 | ], |
| 2932 | ) |
| 2933 | |
| 2934 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2935 | name = "operator_run", |
| 2936 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2937 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2938 | copts = LOGGING_COPTS + select({ |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 2939 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 2940 | "//conditions:default": [], |
| 2941 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2942 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2943 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2944 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2945 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2946 | "@FP16", |
| 2947 | "@FXdiv", |
| 2948 | "@clog", |
| 2949 | "@pthreadpool", |
| 2950 | ], |
| 2951 | ) |
| 2952 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 2953 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2954 | name = "operator_run_test_mode", |
| 2955 | srcs = ["src/operator-run.c"], |
| 2956 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 2957 | copts = LOGGING_COPTS + [ |
| 2958 | "-UNDEBUG", |
| 2959 | "-DXNN_TEST_MODE=1", |
| 2960 | ] + select({ |
| 2961 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 2962 | "//conditions:default": [], |
| 2963 | }), |
| 2964 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2965 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2966 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2967 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2968 | "@FP16", |
| 2969 | "@FXdiv", |
| 2970 | "@clog", |
| 2971 | "@pthreadpool", |
| 2972 | ], |
| 2973 | ) |
| 2974 | |
| 2975 | xnnpack_cc_library( |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 2976 | name = "memory_planner", |
| 2977 | srcs = ["src/memory-planner.c"], |
| 2978 | hdrs = INTERNAL_HDRS, |
| 2979 | defines = select({ |
| 2980 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 2981 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 2982 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 2983 | }), |
| 2984 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2985 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2986 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2987 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 2988 | "@pthreadpool", |
| 2989 | ], |
| 2990 | ) |
| 2991 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2992 | xnnpack_cc_library( |
| 2993 | name = "memory_planner_test_mode", |
| 2994 | srcs = ["src/memory-planner.c"], |
| 2995 | hdrs = INTERNAL_HDRS, |
| 2996 | copts = [ |
| 2997 | "-UNDEBUG", |
| 2998 | "-DXNN_TEST_MODE=1", |
| 2999 | ], |
| 3000 | defines = select({ |
| 3001 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 3002 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 3003 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 3004 | }), |
| 3005 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3006 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3007 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3008 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3009 | "@pthreadpool", |
| 3010 | ], |
| 3011 | ) |
| 3012 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3013 | cc_library( |
| 3014 | name = "enable_assembly", |
| 3015 | defines = select({ |
| 3016 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 3017 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 3018 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3019 | }), |
| 3020 | ) |
| 3021 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3022 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3023 | name = "operators", |
| 3024 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3025 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3026 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3027 | ], |
| 3028 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3029 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3030 | "-Isrc", |
| 3031 | "-Iinclude", |
| 3032 | ] + select({ |
| 3033 | ":debug_build": [], |
| 3034 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3035 | }) + select({ |
| 3036 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3037 | "//conditions:default": [], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3038 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3039 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3040 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3041 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3042 | ":indirection", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3043 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3044 | "@FP16", |
| 3045 | "@FXdiv", |
| 3046 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3047 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3048 | ], |
| 3049 | ) |
| 3050 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3051 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3052 | name = "operators_test_mode", |
| 3053 | srcs = OPERATOR_SRCS + [ |
| 3054 | "src/memory.c", |
| 3055 | "src/operator-delete.c", |
| 3056 | ], |
| 3057 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3058 | copts = LOGGING_COPTS + [ |
| 3059 | "-Isrc", |
| 3060 | "-Iinclude", |
| 3061 | "-UNDEBUG", |
| 3062 | "-DXNN_TEST_MODE=1", |
| 3063 | ] + select({ |
| 3064 | ":debug_build": [], |
| 3065 | "//conditions:default": xnnpack_min_size_copts(), |
| 3066 | }) + select({ |
| 3067 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3068 | "//conditions:default": [], |
| 3069 | }), |
| 3070 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3071 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3072 | deps = [ |
| 3073 | ":indirection_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3074 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3075 | "@FP16", |
| 3076 | "@FXdiv", |
| 3077 | "@clog", |
| 3078 | "@pthreadpool", |
| 3079 | ], |
| 3080 | ) |
| 3081 | |
| 3082 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3083 | name = "XNNPACK", |
| 3084 | srcs = [ |
| 3085 | "src/init.c", |
Marat Dukhan | ccfdbd1 | 2020-02-03 14:27:45 -0800 | [diff] [blame] | 3086 | "src/runtime.c", |
| 3087 | "src/subgraph.c", |
| 3088 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 3089 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3090 | hdrs = ["include/xnnpack.h"], |
| 3091 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3092 | "-Isrc", |
| 3093 | "-Iinclude", |
| 3094 | ] + select({ |
| 3095 | ":debug_build": [], |
| 3096 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3097 | }) + select({ |
| 3098 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3099 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3100 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3101 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3102 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3103 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3104 | visibility = xnnpack_visibility(), |
| 3105 | deps = [ |
| 3106 | ":enable_assembly", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3107 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3108 | ":memory_planner", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3109 | ":operator_run", |
| 3110 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3111 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3112 | "@clog", |
Marat Dukhan | ab2946c | 2020-05-21 20:04:13 -0700 | [diff] [blame] | 3113 | "@FP16", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3114 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 3115 | ] + select({ |
| 3116 | ":emscripten": [], |
| 3117 | "//conditions:default": ["@cpuinfo"], |
| 3118 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3119 | ) |
| 3120 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3121 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3122 | name = "XNNPACK_test_mode", |
| 3123 | srcs = [ |
| 3124 | "src/init.c", |
| 3125 | "src/runtime.c", |
| 3126 | "src/subgraph.c", |
| 3127 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 3128 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3129 | hdrs = ["include/xnnpack.h"], |
| 3130 | copts = LOGGING_COPTS + [ |
| 3131 | "-Isrc", |
| 3132 | "-Iinclude", |
| 3133 | "-UNDEBUG", |
| 3134 | "-DXNN_TEST_MODE=1", |
| 3135 | ] + select({ |
| 3136 | ":debug_build": [], |
| 3137 | "//conditions:default": xnnpack_min_size_copts(), |
| 3138 | }) + select({ |
| 3139 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3140 | "//conditions:default": [], |
| 3141 | }), |
| 3142 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3143 | includes = ["include"], |
| 3144 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3145 | visibility = xnnpack_visibility(), |
| 3146 | deps = [ |
| 3147 | ":enable_assembly", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3148 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3149 | ":memory_planner_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3150 | ":operator_run_test_mode", |
| 3151 | ":operators_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3152 | ":ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3153 | "@clog", |
| 3154 | "@FP16", |
| 3155 | "@pthreadpool", |
| 3156 | ] + select({ |
| 3157 | ":emscripten": [], |
| 3158 | "//conditions:default": ["@cpuinfo"], |
| 3159 | }), |
| 3160 | ) |
| 3161 | |
| 3162 | xnnpack_cc_library( |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3163 | name = "xnnpack_f32", |
| 3164 | srcs = [ |
| 3165 | "src/init.c", |
| 3166 | "src/runtime.c", |
| 3167 | "src/subgraph.c", |
| 3168 | "src/tensor.c", |
| 3169 | ] + SUBGRAPH_SRCS, |
| 3170 | hdrs = ["include/xnnpack.h"], |
| 3171 | copts = LOGGING_COPTS + [ |
| 3172 | "-Isrc", |
| 3173 | "-Iinclude", |
| 3174 | ] + select({ |
| 3175 | ":debug_build": [], |
| 3176 | "//conditions:default": xnnpack_min_size_copts(), |
| 3177 | }) + select({ |
| 3178 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3179 | "//conditions:default": [], |
| 3180 | }), |
| 3181 | defines = [ |
| 3182 | "XNN_NO_Q8_OPERATORS", |
| 3183 | "XNN_NO_U8_OPERATORS", |
| 3184 | "XNN_NO_X8_OPERATORS", |
| 3185 | "XNN_NO_F16_OPERATORS", |
| 3186 | "XNN_NO_X16_OPERATORS", |
| 3187 | ], |
| 3188 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3189 | includes = ["include"], |
| 3190 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3191 | visibility = xnnpack_visibility(), |
| 3192 | deps = [ |
| 3193 | ":enable_assembly", |
| 3194 | ":logging_utils", |
| 3195 | ":memory_planner", |
| 3196 | ":operator_run", |
| 3197 | ":operators", |
| 3198 | ":ukernels_no_f16", |
| 3199 | "@clog", |
| 3200 | "@FP16", |
| 3201 | "@pthreadpool", |
| 3202 | ] + select({ |
| 3203 | ":emscripten": [], |
| 3204 | "//conditions:default": ["@cpuinfo"], |
| 3205 | }), |
| 3206 | ) |
| 3207 | |
| 3208 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3209 | name = "xnnpack_operators_nhwc_f32", |
| 3210 | srcs = [ |
| 3211 | "src/init.c", |
| 3212 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3213 | hdrs = ["include/xnnpack.h"], |
| 3214 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3215 | "-Isrc", |
| 3216 | "-Iinclude", |
| 3217 | ] + select({ |
| 3218 | ":debug_build": [], |
| 3219 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3220 | }) + select({ |
| 3221 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3222 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3223 | }), |
| 3224 | defines = [ |
| 3225 | "XNN_NO_Q8_OPERATORS", |
| 3226 | "XNN_NO_U8_OPERATORS", |
| 3227 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3228 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3229 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3230 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3231 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3232 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3233 | visibility = xnnpack_visibility(), |
| 3234 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3235 | ":enable_assembly", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3236 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3237 | ":operator_run", |
| 3238 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3239 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3240 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3241 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3242 | ] + select({ |
| 3243 | ":emscripten": [], |
| 3244 | "//conditions:default": ["@cpuinfo"], |
| 3245 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3246 | ) |
| 3247 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3248 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3249 | name = "bench_utils", |
| 3250 | srcs = ["bench/utils.cc"], |
| 3251 | hdrs = ["bench/utils.h"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 3252 | deps = [ |
| 3253 | "@com_google_benchmark//:benchmark", |
| 3254 | "@cpuinfo", |
| 3255 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3256 | ) |
| 3257 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 3258 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3259 | |
| 3260 | xnnpack_benchmark( |
| 3261 | name = "q8_gemm_bench", |
| 3262 | srcs = [ |
| 3263 | "bench/gemm.h", |
| 3264 | "bench/q8-gemm.cc", |
| 3265 | "src/xnnpack/AlignedAllocator.h", |
| 3266 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3267 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3268 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
| 3269 | ) |
| 3270 | |
| 3271 | xnnpack_benchmark( |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 3272 | name = "f16_igemm_bench", |
| 3273 | srcs = [ |
| 3274 | "bench/f16-igemm.cc", |
| 3275 | "bench/conv.h", |
| 3276 | "bench/google/conv.h", |
| 3277 | "src/xnnpack/AlignedAllocator.h", |
| 3278 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3279 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 3280 | ) |
| 3281 | |
| 3282 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3283 | name = "f16_gemm_bench", |
| 3284 | srcs = [ |
| 3285 | "bench/f16-gemm.cc", |
| 3286 | "bench/gemm.h", |
| 3287 | "src/xnnpack/AlignedAllocator.h", |
| 3288 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3289 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3290 | ) |
| 3291 | |
| 3292 | xnnpack_benchmark( |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3293 | name = "f16_spmm_bench", |
| 3294 | srcs = [ |
| 3295 | "bench/f16-spmm.cc", |
| 3296 | "bench/gemm.h", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3297 | "src/xnnpack/AlignedAllocator.h", |
| 3298 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3299 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3300 | ) |
| 3301 | |
| 3302 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3303 | name = "f32_igemm_bench", |
| 3304 | srcs = [ |
| 3305 | "bench/f32-igemm.cc", |
| 3306 | "bench/conv.h", |
| 3307 | "src/xnnpack/AlignedAllocator.h", |
| 3308 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 3309 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3310 | ) |
| 3311 | |
| 3312 | xnnpack_benchmark( |
| 3313 | name = "f32_conv_hwc_bench", |
| 3314 | srcs = [ |
| 3315 | "bench/f32-conv-hwc.cc", |
| 3316 | "bench/dconv.h", |
| 3317 | "src/xnnpack/AlignedAllocator.h", |
| 3318 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3319 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3320 | ) |
| 3321 | |
| 3322 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3323 | name = "f32_conv_hwc2chw_bench", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 3324 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3325 | "bench/f32-conv-hwc2chw.cc", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 3326 | "bench/dconv.h", |
| 3327 | "src/xnnpack/AlignedAllocator.h", |
| 3328 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 3329 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3330 | ) |
| 3331 | |
| 3332 | xnnpack_benchmark( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 3333 | name = "f16_dwconv_bench", |
| 3334 | srcs = [ |
| 3335 | "bench/f16-dwconv.cc", |
| 3336 | "bench/dwconv.h", |
| 3337 | "bench/google/dwconv.h", |
| 3338 | "src/xnnpack/AlignedAllocator.h", |
| 3339 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3340 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 3341 | ) |
| 3342 | |
| 3343 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3344 | name = "f32_dwconv_bench", |
| 3345 | srcs = [ |
| 3346 | "bench/f32-dwconv.cc", |
| 3347 | "bench/dwconv.h", |
| 3348 | "src/xnnpack/AlignedAllocator.h", |
| 3349 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3350 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 3351 | ) |
| 3352 | |
| 3353 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3354 | name = "f32_dwconv_chw_bench", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3355 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3356 | "bench/f32-dwconv-chw.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3357 | "bench/dwconv.h", |
| 3358 | "src/xnnpack/AlignedAllocator.h", |
| 3359 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 3360 | deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"], |
| 3361 | ) |
| 3362 | |
| 3363 | xnnpack_benchmark( |
| 3364 | name = "f32_gemm_bench", |
| 3365 | srcs = [ |
| 3366 | "bench/f32-gemm.cc", |
| 3367 | "bench/gemm.h", |
| 3368 | "src/xnnpack/AlignedAllocator.h", |
| 3369 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3370 | copts = xnnpack_optional_ruy_copts(), |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 3371 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3372 | ) |
| 3373 | |
| 3374 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 3375 | name = "f32_raddexpminusmax_bench", |
| 3376 | srcs = [ |
| 3377 | "bench/f32-raddexpminusmax.cc", |
| 3378 | "src/xnnpack/AlignedAllocator.h", |
| 3379 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3380 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3381 | ) |
| 3382 | |
| 3383 | xnnpack_benchmark( |
| 3384 | name = "f32_raddextexp_bench", |
| 3385 | srcs = [ |
| 3386 | "bench/f32-raddextexp.cc", |
| 3387 | "src/xnnpack/AlignedAllocator.h", |
| 3388 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3389 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3390 | ) |
| 3391 | |
| 3392 | xnnpack_benchmark( |
| 3393 | name = "f32_raddstoreexpminusmax_bench", |
| 3394 | srcs = [ |
| 3395 | "bench/f32-raddstoreexpminusmax.cc", |
| 3396 | "src/xnnpack/AlignedAllocator.h", |
| 3397 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3398 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3399 | ) |
| 3400 | |
| 3401 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3402 | name = "f32_rmax_bench", |
| 3403 | srcs = [ |
| 3404 | "bench/f32-rmax.cc", |
| 3405 | "src/xnnpack/AlignedAllocator.h", |
| 3406 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3407 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3408 | ) |
| 3409 | |
| 3410 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 3411 | name = "f32_sigmoid_bench", |
| 3412 | srcs = [ |
| 3413 | "bench/f32-sigmoid.cc", |
| 3414 | "src/xnnpack/AlignedAllocator.h", |
| 3415 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 3416 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3417 | ) |
| 3418 | |
| 3419 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3420 | name = "f32_spmm_bench", |
| 3421 | srcs = [ |
| 3422 | "bench/f32-spmm.cc", |
| 3423 | "bench/gemm.h", |
| 3424 | "src/xnnpack/AlignedAllocator.h", |
| 3425 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3426 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3427 | ) |
| 3428 | |
| 3429 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3430 | name = "f32_softmax_bench", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 3431 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3432 | "bench/f32-softmax.cc", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 3433 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3434 | copts = xnnpack_optional_dnnl_copts(), |
Marat Dukhan | 8d3c693 | 2020-03-06 20:27:27 -0800 | [diff] [blame] | 3435 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(), |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 3436 | ) |
| 3437 | |
| 3438 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 3439 | name = "f32_vscaleexpminusmax_bench", |
| 3440 | srcs = [ |
| 3441 | "bench/f32-vscaleexpminusmax.cc", |
| 3442 | "src/xnnpack/AlignedAllocator.h", |
| 3443 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3444 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3445 | ) |
| 3446 | |
| 3447 | xnnpack_benchmark( |
| 3448 | name = "f32_vscaleextexp_bench", |
| 3449 | srcs = [ |
| 3450 | "bench/f32-vscaleextexp.cc", |
| 3451 | "src/xnnpack/AlignedAllocator.h", |
| 3452 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3453 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3454 | ) |
| 3455 | |
| 3456 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3457 | name = "f32_im2col_gemm_bench", |
| 3458 | srcs = [ |
| 3459 | "bench/f32-im2col-gemm.cc", |
| 3460 | "bench/conv.h", |
| 3461 | "src/xnnpack/AlignedAllocator.h", |
| 3462 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 3463 | deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"], |
| 3464 | ) |
| 3465 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3466 | xnnpack_benchmark( |
| 3467 | name = "requantization_bench", |
| 3468 | srcs = [ |
| 3469 | "bench/requantization.cc", |
| 3470 | "src/xnnpack/requantization-stubs.h", |
| 3471 | "src/xnnpack/AlignedAllocator.h", |
| 3472 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3473 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3474 | ) |
| 3475 | |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 3476 | xnnpack_benchmark( |
| 3477 | name = "rounding_bench", |
| 3478 | srcs = [ |
| 3479 | "bench/rounding.cc", |
| 3480 | "src/xnnpack/math-stubs.h", |
| 3481 | "src/xnnpack/AlignedAllocator.h", |
| 3482 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3483 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3484 | ) |
| 3485 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3486 | ########################### Benchmarks for operators ########################### |
| 3487 | |
| 3488 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3489 | name = "average_pooling_bench", |
| 3490 | srcs = ["bench/average-pooling.cc"], |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 3491 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3492 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3493 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3494 | ) |
| 3495 | |
| 3496 | xnnpack_benchmark( |
| 3497 | name = "channel_shuffle_bench", |
| 3498 | srcs = ["bench/channel-shuffle.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3499 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3500 | ) |
| 3501 | |
| 3502 | xnnpack_benchmark( |
| 3503 | name = "convolution_bench", |
| 3504 | srcs = ["bench/convolution.cc"], |
| 3505 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3506 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3507 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3508 | ) |
| 3509 | |
| 3510 | xnnpack_benchmark( |
| 3511 | name = "deconvolution_bench", |
| 3512 | srcs = ["bench/deconvolution.cc"], |
| 3513 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3514 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3515 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3516 | ) |
| 3517 | |
| 3518 | xnnpack_benchmark( |
| 3519 | name = "global_average_pooling_bench", |
| 3520 | srcs = ["bench/global-average-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3521 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3522 | ) |
| 3523 | |
| 3524 | xnnpack_benchmark( |
| 3525 | name = "max_pooling_bench", |
| 3526 | srcs = ["bench/max-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3527 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3528 | ) |
| 3529 | |
| 3530 | xnnpack_benchmark( |
| 3531 | name = "sigmoid_bench", |
| 3532 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 3533 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 3534 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3535 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3536 | ) |
| 3537 | |
| 3538 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 3539 | name = "prelu_bench", |
| 3540 | srcs = ["bench/prelu.cc"], |
| 3541 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3542 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3543 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 3544 | ) |
| 3545 | |
| 3546 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3547 | name = "softmax_bench", |
| 3548 | srcs = ["bench/softmax.cc"], |
Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame] | 3549 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 3550 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3551 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3552 | ) |
| 3553 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3554 | ############################# End-to-end benchmarks ############################ |
| 3555 | |
| 3556 | cc_library( |
| 3557 | name = "mobilenet_v1", |
| 3558 | srcs = ["models/mobilenet-v1.cc"], |
| 3559 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3560 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3561 | linkstatic = True, |
| 3562 | deps = [ |
| 3563 | ":XNNPACK", |
| 3564 | "@pthreadpool", |
| 3565 | ], |
| 3566 | ) |
| 3567 | |
| 3568 | cc_library( |
| 3569 | name = "mobilenet_v2", |
| 3570 | srcs = ["models/mobilenet-v2.cc"], |
| 3571 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3572 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3573 | linkstatic = True, |
| 3574 | deps = [ |
| 3575 | ":XNNPACK", |
| 3576 | "@pthreadpool", |
| 3577 | ], |
| 3578 | ) |
| 3579 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3580 | cc_library( |
| 3581 | name = "mobilenet_v3_large", |
| 3582 | srcs = ["models/mobilenet-v3-large.cc"], |
| 3583 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3584 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3585 | linkstatic = True, |
| 3586 | deps = [ |
| 3587 | ":XNNPACK", |
| 3588 | "@pthreadpool", |
| 3589 | ], |
| 3590 | ) |
| 3591 | |
| 3592 | cc_library( |
| 3593 | name = "mobilenet_v3_small", |
| 3594 | srcs = ["models/mobilenet-v3-small.cc"], |
| 3595 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3596 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3597 | linkstatic = True, |
| 3598 | deps = [ |
| 3599 | ":XNNPACK", |
| 3600 | "@pthreadpool", |
| 3601 | ], |
| 3602 | ) |
| 3603 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3604 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 3605 | name = "f32_dwconv_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3606 | srcs = [ |
| 3607 | "bench/f32-dwconv-e2e.cc", |
| 3608 | "bench/end2end.h", |
| 3609 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 3610 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3611 | ":XNNPACK", |
| 3612 | ":mobilenet_v1", |
| 3613 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3614 | ":mobilenet_v3_large", |
| 3615 | ":mobilenet_v3_small", |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 3616 | ], |
| 3617 | ) |
| 3618 | |
| 3619 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 3620 | name = "f32_gemm_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3621 | srcs = [ |
| 3622 | "bench/f32-gemm-e2e.cc", |
| 3623 | "bench/end2end.h", |
| 3624 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 3625 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3626 | ":XNNPACK", |
| 3627 | ":mobilenet_v1", |
| 3628 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3629 | ":mobilenet_v3_large", |
| 3630 | ":mobilenet_v3_small", |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 3631 | ], |
| 3632 | ) |
| 3633 | |
| 3634 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3635 | name = "end2end_bench", |
| 3636 | srcs = ["bench/end2end.cc"], |
| 3637 | deps = [ |
| 3638 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 3639 | ":bench_utils", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3640 | ":mobilenet_v1", |
| 3641 | ":mobilenet_v2", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3642 | ":mobilenet_v3_large", |
| 3643 | ":mobilenet_v3_small", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3644 | "@pthreadpool", |
| 3645 | ], |
| 3646 | ) |
| 3647 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3648 | #################### Accuracy evaluation for math functions #################### |
| 3649 | |
| 3650 | xnnpack_benchmark( |
| 3651 | name = "f32_exp_eval", |
| 3652 | srcs = [ |
| 3653 | "eval/f32-exp.cc", |
| 3654 | "src/xnnpack/AlignedAllocator.h", |
| 3655 | ] + ACCURACY_EVAL_HDRS, |
| 3656 | deps = ACCURACY_EVAL_DEPS, |
| 3657 | ) |
| 3658 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 3659 | xnnpack_benchmark( |
| 3660 | name = "f32_expminus_eval", |
| 3661 | srcs = [ |
| 3662 | "eval/f32-expminus.cc", |
| 3663 | "src/xnnpack/AlignedAllocator.h", |
| 3664 | ] + ACCURACY_EVAL_HDRS, |
| 3665 | deps = ACCURACY_EVAL_DEPS, |
| 3666 | ) |
| 3667 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 3668 | xnnpack_benchmark( |
| 3669 | name = "f32_extexp_eval", |
| 3670 | srcs = [ |
| 3671 | "eval/f32-extexp.cc", |
| 3672 | "src/xnnpack/AlignedAllocator.h", |
| 3673 | ] + ACCURACY_EVAL_HDRS, |
| 3674 | deps = ACCURACY_EVAL_DEPS, |
| 3675 | ) |
| 3676 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3677 | xnnpack_unit_test( |
| 3678 | name = "f32_roundne_eval", |
| 3679 | srcs = [ |
| 3680 | "eval/f32-roundne.cc", |
| 3681 | "src/xnnpack/AlignedAllocator.h", |
| 3682 | "src/xnnpack/math-stubs.h", |
| 3683 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | 22eed3d | 2020-05-11 20:13:37 -0700 | [diff] [blame] | 3684 | automatic = False, |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3685 | deps = MICROKERNEL_TEST_DEPS, |
| 3686 | ) |
| 3687 | |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 3688 | xnnpack_unit_test( |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 3689 | name = "f32_roundd_eval", |
| 3690 | srcs = [ |
| 3691 | "eval/f32-roundd.cc", |
| 3692 | "src/xnnpack/AlignedAllocator.h", |
| 3693 | "src/xnnpack/math-stubs.h", |
| 3694 | ] + MICROKERNEL_TEST_HDRS, |
| 3695 | automatic = False, |
| 3696 | deps = MICROKERNEL_TEST_DEPS, |
| 3697 | ) |
| 3698 | |
| 3699 | xnnpack_unit_test( |
| 3700 | name = "f32_roundu_eval", |
| 3701 | srcs = [ |
| 3702 | "eval/f32-roundu.cc", |
| 3703 | "src/xnnpack/AlignedAllocator.h", |
| 3704 | "src/xnnpack/math-stubs.h", |
| 3705 | ] + MICROKERNEL_TEST_HDRS, |
| 3706 | automatic = False, |
| 3707 | deps = MICROKERNEL_TEST_DEPS, |
| 3708 | ) |
| 3709 | |
| 3710 | xnnpack_unit_test( |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 3711 | name = "f32_roundz_eval", |
| 3712 | srcs = [ |
| 3713 | "eval/f32-roundz.cc", |
| 3714 | "src/xnnpack/AlignedAllocator.h", |
| 3715 | "src/xnnpack/math-stubs.h", |
| 3716 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 3717 | automatic = False, |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 3718 | deps = MICROKERNEL_TEST_DEPS, |
| 3719 | ) |
| 3720 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 3721 | xnnpack_benchmark( |
| 3722 | name = "f32_sigmoid_eval", |
| 3723 | srcs = [ |
| 3724 | "eval/f32-sigmoid.cc", |
| 3725 | "src/xnnpack/AlignedAllocator.h", |
| 3726 | ] + ACCURACY_EVAL_HDRS, |
| 3727 | deps = ACCURACY_EVAL_DEPS, |
| 3728 | ) |
| 3729 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3730 | ######################### Unit tests for micro-kernels ######################### |
| 3731 | |
| 3732 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 3733 | name = "f16_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3734 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 3735 | "test/f16-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3736 | "test/gemm-microkernel-tester.h", |
| 3737 | "src/xnnpack/AlignedAllocator.h", |
| 3738 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3739 | deps = MICROKERNEL_TEST_DEPS, |
| 3740 | ) |
| 3741 | |
| 3742 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 3743 | name = "f16_spmm_minmax_test", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3744 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 3745 | "test/f16-spmm-minmax.cc", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3746 | "test/spmm-microkernel-tester.h", |
| 3747 | "src/xnnpack/AlignedAllocator.h", |
| 3748 | ] + MICROKERNEL_TEST_HDRS, |
| 3749 | deps = MICROKERNEL_TEST_DEPS, |
| 3750 | ) |
| 3751 | |
| 3752 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3753 | name = "f32_argmaxpool_test", |
| 3754 | srcs = [ |
| 3755 | "test/f32-argmaxpool.cc", |
| 3756 | "test/argmaxpool-microkernel-tester.h", |
| 3757 | "src/xnnpack/AlignedAllocator.h", |
| 3758 | ] + MICROKERNEL_TEST_HDRS, |
| 3759 | deps = MICROKERNEL_TEST_DEPS, |
| 3760 | ) |
| 3761 | |
| 3762 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3763 | name = "f32_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3764 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3765 | "test/f32-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3766 | "test/avgpool-microkernel-tester.h", |
| 3767 | "src/xnnpack/AlignedAllocator.h", |
| 3768 | ] + MICROKERNEL_TEST_HDRS, |
| 3769 | deps = MICROKERNEL_TEST_DEPS, |
| 3770 | ) |
| 3771 | |
| 3772 | xnnpack_unit_test( |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 3773 | name = "f32_ibilinear_test", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 3774 | srcs = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 3775 | "test/f32-ibilinear.cc", |
| 3776 | "test/ibilinear-microkernel-tester.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 3777 | "src/xnnpack/AlignedAllocator.h", |
| 3778 | ] + MICROKERNEL_TEST_HDRS, |
| 3779 | deps = MICROKERNEL_TEST_DEPS, |
| 3780 | ) |
| 3781 | |
| 3782 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 3783 | name = "f16_clamp_test", |
| 3784 | srcs = [ |
| 3785 | "test/f16-clamp.cc", |
| 3786 | "test/clamp-microkernel-tester.h", |
| 3787 | ] + MICROKERNEL_TEST_HDRS, |
| 3788 | deps = MICROKERNEL_TEST_DEPS, |
| 3789 | ) |
| 3790 | |
| 3791 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3792 | name = "f32_clamp_test", |
| 3793 | srcs = [ |
| 3794 | "test/f32-clamp.cc", |
| 3795 | "test/clamp-microkernel-tester.h", |
| 3796 | ] + MICROKERNEL_TEST_HDRS, |
| 3797 | deps = MICROKERNEL_TEST_DEPS, |
| 3798 | ) |
| 3799 | |
| 3800 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 3801 | name = "f32_igemm_test", |
| 3802 | srcs = [ |
| 3803 | "test/f32-igemm.cc", |
| 3804 | "test/gemm-microkernel-tester.h", |
| 3805 | "src/xnnpack/AlignedAllocator.h", |
| 3806 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3807 | deps = MICROKERNEL_TEST_DEPS, |
| 3808 | ) |
| 3809 | |
| 3810 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 3811 | name = "f32_igemm_relu_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3812 | srcs = [ |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 3813 | "test/f32-igemm-relu.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3814 | "test/gemm-microkernel-tester.h", |
| 3815 | "src/xnnpack/AlignedAllocator.h", |
| 3816 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3817 | deps = MICROKERNEL_TEST_DEPS, |
| 3818 | ) |
| 3819 | |
| 3820 | xnnpack_unit_test( |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 3821 | name = "f32_igemm_minmax_test", |
| 3822 | srcs = [ |
| 3823 | "test/f32-igemm-minmax.cc", |
| 3824 | "test/gemm-microkernel-tester.h", |
| 3825 | "src/xnnpack/AlignedAllocator.h", |
| 3826 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3827 | deps = MICROKERNEL_TEST_DEPS, |
| 3828 | ) |
| 3829 | |
| 3830 | xnnpack_unit_test( |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 3831 | name = "f16_igemm_minmax_test", |
| 3832 | srcs = [ |
| 3833 | "test/f16-igemm-minmax.cc", |
| 3834 | "test/gemm-microkernel-tester.h", |
| 3835 | "src/xnnpack/AlignedAllocator.h", |
| 3836 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3837 | deps = MICROKERNEL_TEST_DEPS, |
| 3838 | ) |
| 3839 | |
| 3840 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3841 | name = "f32_conv_hwc_test", |
| 3842 | srcs = [ |
| 3843 | "test/f32-conv-hwc.cc", |
| 3844 | "test/conv-hwc-microkernel-tester.h", |
| 3845 | "src/xnnpack/AlignedAllocator.h", |
| 3846 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3847 | deps = MICROKERNEL_TEST_DEPS, |
| 3848 | ) |
| 3849 | |
| 3850 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3851 | name = "f32_conv_hwc2chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3852 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3853 | "test/f32-conv-hwc2chw.cc", |
| 3854 | "test/conv-hwc2chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3855 | "src/xnnpack/AlignedAllocator.h", |
| 3856 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3857 | deps = MICROKERNEL_TEST_DEPS, |
| 3858 | ) |
| 3859 | |
| 3860 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 3861 | name = "f32_dwconv_test", |
| 3862 | srcs = [ |
| 3863 | "test/f32-dwconv.cc", |
| 3864 | "test/dwconv-microkernel-tester.h", |
| 3865 | "src/xnnpack/AlignedAllocator.h", |
| 3866 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3867 | deps = MICROKERNEL_TEST_DEPS, |
| 3868 | ) |
| 3869 | |
| 3870 | xnnpack_unit_test( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 3871 | name = "f16_dwconv_minmax_test", |
| 3872 | srcs = [ |
| 3873 | "test/f16-dwconv-minmax.cc", |
| 3874 | "test/dwconv-microkernel-tester.h", |
| 3875 | "src/xnnpack/AlignedAllocator.h", |
| 3876 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3877 | deps = MICROKERNEL_TEST_DEPS, |
| 3878 | ) |
| 3879 | |
| 3880 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3881 | name = "f32_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3882 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3883 | "test/f32-dwconv-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3884 | "test/dwconv-microkernel-tester.h", |
| 3885 | "src/xnnpack/AlignedAllocator.h", |
| 3886 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3887 | deps = MICROKERNEL_TEST_DEPS, |
| 3888 | ) |
| 3889 | |
| 3890 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3891 | name = "f32_dwconv_chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3892 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3893 | "test/f32-dwconv-chw.cc", |
| 3894 | "test/dwconv-chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3895 | "src/xnnpack/AlignedAllocator.h", |
| 3896 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3897 | deps = MICROKERNEL_TEST_DEPS, |
| 3898 | ) |
| 3899 | |
| 3900 | xnnpack_unit_test( |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 3901 | name = "f16_gavgpool_minmax_test", |
| 3902 | srcs = [ |
| 3903 | "test/f16-gavgpool-minmax.cc", |
| 3904 | "test/gavgpool-microkernel-tester.h", |
| 3905 | "src/xnnpack/AlignedAllocator.h", |
| 3906 | ] + MICROKERNEL_TEST_HDRS, |
| 3907 | deps = MICROKERNEL_TEST_DEPS, |
| 3908 | ) |
| 3909 | |
| 3910 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3911 | name = "f32_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3912 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3913 | "test/f32-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3914 | "test/gavgpool-microkernel-tester.h", |
| 3915 | "src/xnnpack/AlignedAllocator.h", |
| 3916 | ] + MICROKERNEL_TEST_HDRS, |
| 3917 | deps = MICROKERNEL_TEST_DEPS, |
| 3918 | ) |
| 3919 | |
| 3920 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3921 | name = "f32_gavgpool_cw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3922 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3923 | "test/f32-gavgpool-cw.cc", |
| 3924 | "test/gavgpool-cw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3925 | "src/xnnpack/AlignedAllocator.h", |
| 3926 | ] + MICROKERNEL_TEST_HDRS, |
| 3927 | deps = MICROKERNEL_TEST_DEPS, |
| 3928 | ) |
| 3929 | |
| 3930 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 3931 | name = "f32_gemm_test", |
| 3932 | srcs = [ |
| 3933 | "test/f32-gemm.cc", |
| 3934 | "test/gemm-microkernel-tester.h", |
| 3935 | "src/xnnpack/AlignedAllocator.h", |
| 3936 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3937 | deps = MICROKERNEL_TEST_DEPS, |
| 3938 | ) |
| 3939 | |
| 3940 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 3941 | name = "f32_gemm_relu_test", |
| 3942 | srcs = [ |
| 3943 | "test/f32-gemm-relu.cc", |
| 3944 | "test/gemm-microkernel-tester.h", |
| 3945 | "src/xnnpack/AlignedAllocator.h", |
| 3946 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3947 | deps = MICROKERNEL_TEST_DEPS, |
| 3948 | ) |
| 3949 | |
| 3950 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3951 | name = "f32_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3952 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3953 | "test/f32-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3954 | "test/gemm-microkernel-tester.h", |
| 3955 | "src/xnnpack/AlignedAllocator.h", |
| 3956 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3957 | deps = MICROKERNEL_TEST_DEPS, |
| 3958 | ) |
| 3959 | |
| 3960 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3961 | name = "f32_gemminc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3962 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3963 | "test/f32-gemminc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3964 | "test/gemm-microkernel-tester.h", |
| 3965 | "src/xnnpack/AlignedAllocator.h", |
| 3966 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 3967 | deps = MICROKERNEL_TEST_DEPS, |
| 3968 | ) |
| 3969 | |
| 3970 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 3971 | name = "f16_hswish_test", |
| 3972 | srcs = [ |
| 3973 | "test/f16-hswish.cc", |
| 3974 | "test/hswish-microkernel-tester.h", |
| 3975 | ] + MICROKERNEL_TEST_HDRS, |
| 3976 | deps = MICROKERNEL_TEST_DEPS, |
| 3977 | ) |
| 3978 | |
| 3979 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3980 | name = "f32_hswish_test", |
| 3981 | srcs = [ |
| 3982 | "test/f32-hswish.cc", |
| 3983 | "test/hswish-microkernel-tester.h", |
| 3984 | ] + MICROKERNEL_TEST_HDRS, |
| 3985 | deps = MICROKERNEL_TEST_DEPS, |
| 3986 | ) |
| 3987 | |
| 3988 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3989 | name = "f32_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3990 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3991 | "test/f32-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3992 | "test/maxpool-microkernel-tester.h", |
| 3993 | ] + MICROKERNEL_TEST_HDRS, |
| 3994 | deps = MICROKERNEL_TEST_DEPS, |
| 3995 | ) |
| 3996 | |
| 3997 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 3998 | name = "f32_pavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3999 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4000 | "test/f32-pavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4001 | "test/avgpool-microkernel-tester.h", |
| 4002 | "src/xnnpack/AlignedAllocator.h", |
| 4003 | ] + MICROKERNEL_TEST_HDRS, |
| 4004 | deps = MICROKERNEL_TEST_DEPS, |
| 4005 | ) |
| 4006 | |
| 4007 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4008 | name = "f32_ppmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4009 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4010 | "test/f32-ppmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4011 | "test/gemm-microkernel-tester.h", |
| 4012 | "src/xnnpack/AlignedAllocator.h", |
| 4013 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 4014 | deps = MICROKERNEL_TEST_DEPS, |
| 4015 | ) |
| 4016 | |
| 4017 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 4018 | name = "f16_prelu_test", |
| 4019 | srcs = [ |
| 4020 | "test/f16-prelu.cc", |
| 4021 | "test/prelu-microkernel-tester.h", |
| 4022 | "src/xnnpack/AlignedAllocator.h", |
| 4023 | ] + MICROKERNEL_TEST_HDRS, |
| 4024 | deps = MICROKERNEL_TEST_DEPS, |
| 4025 | ) |
| 4026 | |
| 4027 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4028 | name = "f32_prelu_test", |
| 4029 | srcs = [ |
| 4030 | "test/f32-prelu.cc", |
| 4031 | "test/prelu-microkernel-tester.h", |
| 4032 | "src/xnnpack/AlignedAllocator.h", |
| 4033 | ] + MICROKERNEL_TEST_HDRS, |
| 4034 | deps = MICROKERNEL_TEST_DEPS, |
| 4035 | ) |
| 4036 | |
| 4037 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4038 | name = "f32_raddexpminusmax_test", |
| 4039 | srcs = [ |
| 4040 | "test/f32-raddexpminusmax.cc", |
| 4041 | "test/raddexpminusmax-microkernel-tester.h", |
| 4042 | ] + MICROKERNEL_TEST_HDRS, |
| 4043 | deps = MICROKERNEL_TEST_DEPS, |
| 4044 | ) |
| 4045 | |
| 4046 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 4047 | name = "f32_raddextexp_test", |
| 4048 | srcs = [ |
| 4049 | "test/f32-raddextexp.cc", |
| 4050 | "test/raddextexp-microkernel-tester.h", |
| 4051 | ] + MICROKERNEL_TEST_HDRS, |
| 4052 | deps = MICROKERNEL_TEST_DEPS, |
| 4053 | ) |
| 4054 | |
| 4055 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4056 | name = "f32_raddstoreexpminusmax_test", |
| 4057 | srcs = [ |
| 4058 | "test/f32-raddstoreexpminusmax.cc", |
| 4059 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 4060 | ] + MICROKERNEL_TEST_HDRS, |
| 4061 | deps = MICROKERNEL_TEST_DEPS, |
| 4062 | ) |
| 4063 | |
| 4064 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4065 | name = "f32_rmax_test", |
| 4066 | srcs = [ |
| 4067 | "test/f32-rmax.cc", |
| 4068 | "test/rmax-microkernel-tester.h", |
| 4069 | ] + MICROKERNEL_TEST_HDRS, |
| 4070 | deps = MICROKERNEL_TEST_DEPS, |
| 4071 | ) |
| 4072 | |
| 4073 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 4074 | name = "f32_sigmoid_test", |
| 4075 | srcs = [ |
| 4076 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4077 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 4078 | ] + MICROKERNEL_TEST_HDRS, |
| 4079 | deps = MICROKERNEL_TEST_DEPS, |
| 4080 | ) |
| 4081 | |
| 4082 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4083 | name = "f32_spmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4084 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4085 | "test/f32-spmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4086 | "test/spmm-microkernel-tester.h", |
| 4087 | "src/xnnpack/AlignedAllocator.h", |
| 4088 | ] + MICROKERNEL_TEST_HDRS, |
| 4089 | deps = MICROKERNEL_TEST_DEPS, |
| 4090 | ) |
| 4091 | |
| 4092 | xnnpack_unit_test( |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 4093 | name = "f16_vadd_minmax_test", |
| 4094 | srcs = [ |
| 4095 | "test/f16-vadd-minmax.cc", |
| 4096 | "test/vbinary-microkernel-tester.h", |
| 4097 | ] + MICROKERNEL_TEST_HDRS, |
| 4098 | deps = MICROKERNEL_TEST_DEPS, |
| 4099 | ) |
| 4100 | |
| 4101 | xnnpack_unit_test( |
| 4102 | name = "f16_vaddc_minmax_test", |
| 4103 | srcs = [ |
| 4104 | "test/f16-vaddc-minmax.cc", |
| 4105 | "test/vbinaryc-microkernel-tester.h", |
| 4106 | ] + MICROKERNEL_TEST_HDRS, |
| 4107 | deps = MICROKERNEL_TEST_DEPS, |
| 4108 | ) |
| 4109 | |
| 4110 | xnnpack_unit_test( |
| 4111 | name = "f16_vdiv_minmax_test", |
| 4112 | srcs = [ |
| 4113 | "test/f16-vdiv-minmax.cc", |
| 4114 | "test/vbinary-microkernel-tester.h", |
| 4115 | ] + MICROKERNEL_TEST_HDRS, |
| 4116 | deps = MICROKERNEL_TEST_DEPS, |
| 4117 | ) |
| 4118 | |
| 4119 | xnnpack_unit_test( |
| 4120 | name = "f16_vdivc_minmax_test", |
| 4121 | srcs = [ |
| 4122 | "test/f16-vdivc-minmax.cc", |
| 4123 | "test/vbinaryc-microkernel-tester.h", |
| 4124 | ] + MICROKERNEL_TEST_HDRS, |
| 4125 | deps = MICROKERNEL_TEST_DEPS, |
| 4126 | ) |
| 4127 | |
| 4128 | xnnpack_unit_test( |
| 4129 | name = "f16_vrdivc_minmax_test", |
| 4130 | srcs = [ |
| 4131 | "test/f16-vrdivc-minmax.cc", |
| 4132 | "test/vbinaryc-microkernel-tester.h", |
| 4133 | ] + MICROKERNEL_TEST_HDRS, |
| 4134 | deps = MICROKERNEL_TEST_DEPS, |
| 4135 | ) |
| 4136 | |
| 4137 | xnnpack_unit_test( |
| 4138 | name = "f16_vmax_test", |
| 4139 | srcs = [ |
| 4140 | "test/f16-vmax.cc", |
| 4141 | "test/vbinary-microkernel-tester.h", |
| 4142 | ] + MICROKERNEL_TEST_HDRS, |
| 4143 | deps = MICROKERNEL_TEST_DEPS, |
| 4144 | ) |
| 4145 | |
| 4146 | xnnpack_unit_test( |
| 4147 | name = "f16_vmaxc_test", |
| 4148 | srcs = [ |
| 4149 | "test/f16-vmaxc.cc", |
| 4150 | "test/vbinaryc-microkernel-tester.h", |
| 4151 | ] + MICROKERNEL_TEST_HDRS, |
| 4152 | deps = MICROKERNEL_TEST_DEPS, |
| 4153 | ) |
| 4154 | |
| 4155 | xnnpack_unit_test( |
| 4156 | name = "f16_vmin_test", |
| 4157 | srcs = [ |
| 4158 | "test/f16-vmin.cc", |
| 4159 | "test/vbinary-microkernel-tester.h", |
| 4160 | ] + MICROKERNEL_TEST_HDRS, |
| 4161 | deps = MICROKERNEL_TEST_DEPS, |
| 4162 | ) |
| 4163 | |
| 4164 | xnnpack_unit_test( |
| 4165 | name = "f16_vminc_test", |
| 4166 | srcs = [ |
| 4167 | "test/f16-vminc.cc", |
| 4168 | "test/vbinaryc-microkernel-tester.h", |
| 4169 | ] + MICROKERNEL_TEST_HDRS, |
| 4170 | deps = MICROKERNEL_TEST_DEPS, |
| 4171 | ) |
| 4172 | |
| 4173 | xnnpack_unit_test( |
| 4174 | name = "f16_vmul_minmax_test", |
| 4175 | srcs = [ |
| 4176 | "test/f16-vmul-minmax.cc", |
| 4177 | "test/vbinary-microkernel-tester.h", |
| 4178 | ] + MICROKERNEL_TEST_HDRS, |
| 4179 | deps = MICROKERNEL_TEST_DEPS, |
| 4180 | ) |
| 4181 | |
| 4182 | xnnpack_unit_test( |
| 4183 | name = "f16_vmulc_minmax_test", |
| 4184 | srcs = [ |
| 4185 | "test/f16-vmulc-minmax.cc", |
| 4186 | "test/vbinaryc-microkernel-tester.h", |
| 4187 | ] + MICROKERNEL_TEST_HDRS, |
| 4188 | deps = MICROKERNEL_TEST_DEPS, |
| 4189 | ) |
| 4190 | |
| 4191 | xnnpack_unit_test( |
| 4192 | name = "f16_vsub_minmax_test", |
| 4193 | srcs = [ |
| 4194 | "test/f16-vsub-minmax.cc", |
| 4195 | "test/vbinary-microkernel-tester.h", |
| 4196 | ] + MICROKERNEL_TEST_HDRS, |
| 4197 | deps = MICROKERNEL_TEST_DEPS, |
| 4198 | ) |
| 4199 | |
| 4200 | xnnpack_unit_test( |
| 4201 | name = "f16_vsubc_minmax_test", |
| 4202 | srcs = [ |
| 4203 | "test/f16-vsubc-minmax.cc", |
| 4204 | "test/vbinaryc-microkernel-tester.h", |
| 4205 | ] + MICROKERNEL_TEST_HDRS, |
| 4206 | deps = MICROKERNEL_TEST_DEPS, |
| 4207 | ) |
| 4208 | |
| 4209 | xnnpack_unit_test( |
| 4210 | name = "f16_vrsubc_minmax_test", |
| 4211 | srcs = [ |
| 4212 | "test/f16-vrsubc-minmax.cc", |
| 4213 | "test/vbinaryc-microkernel-tester.h", |
| 4214 | ] + MICROKERNEL_TEST_HDRS, |
| 4215 | deps = MICROKERNEL_TEST_DEPS, |
| 4216 | ) |
| 4217 | |
| 4218 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4219 | name = "f32_vabs_test", |
| 4220 | srcs = [ |
| 4221 | "test/f32-vabs.cc", |
| 4222 | "test/vunary-microkernel-tester.h", |
| 4223 | ] + MICROKERNEL_TEST_HDRS, |
| 4224 | deps = MICROKERNEL_TEST_DEPS, |
| 4225 | ) |
| 4226 | |
| 4227 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4228 | name = "f32_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4229 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4230 | "test/f32-vadd-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4231 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4232 | ] + MICROKERNEL_TEST_HDRS, |
| 4233 | deps = MICROKERNEL_TEST_DEPS, |
| 4234 | ) |
| 4235 | |
| 4236 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4237 | name = "f32_vaddc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4238 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4239 | "test/f32-vaddc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4240 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4241 | ] + MICROKERNEL_TEST_HDRS, |
| 4242 | deps = MICROKERNEL_TEST_DEPS, |
| 4243 | ) |
| 4244 | |
| 4245 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4246 | name = "f32_vdiv_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4247 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4248 | "test/f32-vdiv-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4249 | "test/vbinary-microkernel-tester.h", |
| 4250 | ] + MICROKERNEL_TEST_HDRS, |
| 4251 | deps = MICROKERNEL_TEST_DEPS, |
| 4252 | ) |
| 4253 | |
| 4254 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4255 | name = "f32_vdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4256 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4257 | "test/f32-vdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4258 | "test/vbinaryc-microkernel-tester.h", |
| 4259 | ] + MICROKERNEL_TEST_HDRS, |
| 4260 | deps = MICROKERNEL_TEST_DEPS, |
| 4261 | ) |
| 4262 | |
| 4263 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4264 | name = "f32_vrdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4265 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4266 | "test/f32-vrdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4267 | "test/vbinaryc-microkernel-tester.h", |
| 4268 | ] + MICROKERNEL_TEST_HDRS, |
| 4269 | deps = MICROKERNEL_TEST_DEPS, |
| 4270 | ) |
| 4271 | |
| 4272 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 4273 | name = "f32_vmax_test", |
| 4274 | srcs = [ |
| 4275 | "test/f32-vmax.cc", |
| 4276 | "test/vbinary-microkernel-tester.h", |
| 4277 | ] + MICROKERNEL_TEST_HDRS, |
| 4278 | deps = MICROKERNEL_TEST_DEPS, |
| 4279 | ) |
| 4280 | |
| 4281 | xnnpack_unit_test( |
| 4282 | name = "f32_vmaxc_test", |
| 4283 | srcs = [ |
| 4284 | "test/f32-vmaxc.cc", |
| 4285 | "test/vbinaryc-microkernel-tester.h", |
| 4286 | ] + MICROKERNEL_TEST_HDRS, |
| 4287 | deps = MICROKERNEL_TEST_DEPS, |
| 4288 | ) |
| 4289 | |
| 4290 | xnnpack_unit_test( |
| 4291 | name = "f32_vmin_test", |
| 4292 | srcs = [ |
| 4293 | "test/f32-vmin.cc", |
| 4294 | "test/vbinary-microkernel-tester.h", |
| 4295 | ] + MICROKERNEL_TEST_HDRS, |
| 4296 | deps = MICROKERNEL_TEST_DEPS, |
| 4297 | ) |
| 4298 | |
| 4299 | xnnpack_unit_test( |
| 4300 | name = "f32_vminc_test", |
| 4301 | srcs = [ |
| 4302 | "test/f32-vminc.cc", |
| 4303 | "test/vbinaryc-microkernel-tester.h", |
| 4304 | ] + MICROKERNEL_TEST_HDRS, |
| 4305 | deps = MICROKERNEL_TEST_DEPS, |
| 4306 | ) |
| 4307 | |
| 4308 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4309 | name = "f32_vmul_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4310 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4311 | "test/f32-vmul-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4312 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4313 | ] + MICROKERNEL_TEST_HDRS, |
| 4314 | deps = MICROKERNEL_TEST_DEPS, |
| 4315 | ) |
| 4316 | |
| 4317 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4318 | name = "f32_vmulc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4319 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4320 | "test/f32-vmulc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4321 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4322 | ] + MICROKERNEL_TEST_HDRS, |
| 4323 | deps = MICROKERNEL_TEST_DEPS, |
| 4324 | ) |
| 4325 | |
| 4326 | xnnpack_unit_test( |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 4327 | name = "f16_vmulcaddc_minmax_test", |
| 4328 | srcs = [ |
| 4329 | "test/f16-vmulcaddc-minmax.cc", |
| 4330 | "test/vmulcaddc-microkernel-tester.h", |
| 4331 | "src/xnnpack/AlignedAllocator.h", |
| 4332 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 4333 | deps = MICROKERNEL_TEST_DEPS, |
| 4334 | ) |
| 4335 | |
| 4336 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4337 | name = "f32_vmulcaddc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4338 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4339 | "test/f32-vmulcaddc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4340 | "test/vmulcaddc-microkernel-tester.h", |
| 4341 | "src/xnnpack/AlignedAllocator.h", |
| 4342 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 4343 | deps = MICROKERNEL_TEST_DEPS, |
| 4344 | ) |
| 4345 | |
| 4346 | xnnpack_unit_test( |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 4347 | name = "f32_vlrelu_test", |
| 4348 | srcs = [ |
| 4349 | "test/f32-vlrelu.cc", |
| 4350 | "test/vunary-microkernel-tester.h", |
| 4351 | ] + MICROKERNEL_TEST_HDRS, |
| 4352 | deps = MICROKERNEL_TEST_DEPS, |
| 4353 | ) |
| 4354 | |
| 4355 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4356 | name = "f32_vneg_test", |
| 4357 | srcs = [ |
| 4358 | "test/f32-vneg.cc", |
| 4359 | "test/vunary-microkernel-tester.h", |
| 4360 | ] + MICROKERNEL_TEST_HDRS, |
| 4361 | deps = MICROKERNEL_TEST_DEPS, |
| 4362 | ) |
| 4363 | |
| 4364 | xnnpack_unit_test( |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 4365 | name = "f32_vrndne_test", |
| 4366 | srcs = [ |
| 4367 | "test/f32-vrndne.cc", |
| 4368 | "test/vunary-microkernel-tester.h", |
| 4369 | ] + MICROKERNEL_TEST_HDRS, |
| 4370 | deps = MICROKERNEL_TEST_DEPS, |
| 4371 | ) |
| 4372 | |
| 4373 | xnnpack_unit_test( |
| 4374 | name = "f32_vrndz_test", |
| 4375 | srcs = [ |
| 4376 | "test/f32-vrndz.cc", |
| 4377 | "test/vunary-microkernel-tester.h", |
| 4378 | ] + MICROKERNEL_TEST_HDRS, |
| 4379 | deps = MICROKERNEL_TEST_DEPS, |
| 4380 | ) |
| 4381 | |
| 4382 | xnnpack_unit_test( |
| 4383 | name = "f32_vrndu_test", |
| 4384 | srcs = [ |
| 4385 | "test/f32-vrndu.cc", |
| 4386 | "test/vunary-microkernel-tester.h", |
| 4387 | ] + MICROKERNEL_TEST_HDRS, |
| 4388 | deps = MICROKERNEL_TEST_DEPS, |
| 4389 | ) |
| 4390 | |
| 4391 | xnnpack_unit_test( |
| 4392 | name = "f32_vrndd_test", |
| 4393 | srcs = [ |
| 4394 | "test/f32-vrndd.cc", |
| 4395 | "test/vunary-microkernel-tester.h", |
| 4396 | ] + MICROKERNEL_TEST_HDRS, |
| 4397 | deps = MICROKERNEL_TEST_DEPS, |
| 4398 | ) |
| 4399 | |
| 4400 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 4401 | name = "f32_vscale_test", |
| 4402 | srcs = [ |
| 4403 | "test/f32-vscale.cc", |
| 4404 | "test/vscale-microkernel-tester.h", |
| 4405 | ] + MICROKERNEL_TEST_HDRS, |
| 4406 | deps = MICROKERNEL_TEST_DEPS, |
| 4407 | ) |
| 4408 | |
| 4409 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4410 | name = "f32_vscaleexpminusmax_test", |
| 4411 | srcs = [ |
| 4412 | "test/f32-vscaleexpminusmax.cc", |
| 4413 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 4414 | ] + MICROKERNEL_TEST_HDRS, |
| 4415 | deps = MICROKERNEL_TEST_DEPS, |
| 4416 | ) |
| 4417 | |
| 4418 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 4419 | name = "f32_vscaleextexp_test", |
| 4420 | srcs = [ |
| 4421 | "test/f32-vscaleextexp.cc", |
| 4422 | "test/vscaleextexp-microkernel-tester.h", |
| 4423 | ] + MICROKERNEL_TEST_HDRS, |
| 4424 | deps = MICROKERNEL_TEST_DEPS, |
| 4425 | ) |
| 4426 | |
| 4427 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4428 | name = "f32_vsqr_test", |
| 4429 | srcs = [ |
| 4430 | "test/f32-vsqr.cc", |
| 4431 | "test/vunary-microkernel-tester.h", |
| 4432 | ] + MICROKERNEL_TEST_HDRS, |
| 4433 | deps = MICROKERNEL_TEST_DEPS, |
| 4434 | ) |
| 4435 | |
| 4436 | xnnpack_unit_test( |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 4437 | name = "f32_vsqrdiff_test", |
| 4438 | srcs = [ |
| 4439 | "test/f32-vsqrdiff.cc", |
| 4440 | "test/vbinary-microkernel-tester.h", |
| 4441 | ] + MICROKERNEL_TEST_HDRS, |
| 4442 | deps = MICROKERNEL_TEST_DEPS, |
| 4443 | ) |
| 4444 | |
| 4445 | xnnpack_unit_test( |
| 4446 | name = "f32_vsqrdiffc_test", |
| 4447 | srcs = [ |
| 4448 | "test/f32-vsqrdiffc.cc", |
| 4449 | "test/vbinaryc-microkernel-tester.h", |
| 4450 | ] + MICROKERNEL_TEST_HDRS, |
| 4451 | deps = MICROKERNEL_TEST_DEPS, |
| 4452 | ) |
| 4453 | |
| 4454 | xnnpack_unit_test( |
| 4455 | name = "f32_vrsqrdiffc_test", |
| 4456 | srcs = [ |
| 4457 | "test/f32-vrsqrdiffc.cc", |
| 4458 | "test/vbinaryc-microkernel-tester.h", |
| 4459 | ] + MICROKERNEL_TEST_HDRS, |
| 4460 | deps = MICROKERNEL_TEST_DEPS, |
| 4461 | ) |
| 4462 | |
| 4463 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4464 | name = "f32_vsub_minmax_test", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4465 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4466 | "test/f32-vsub-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4467 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4468 | ] + MICROKERNEL_TEST_HDRS, |
| 4469 | deps = MICROKERNEL_TEST_DEPS, |
| 4470 | ) |
| 4471 | |
| 4472 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4473 | name = "f32_vsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4474 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4475 | "test/f32-vsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4476 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4477 | ] + MICROKERNEL_TEST_HDRS, |
| 4478 | deps = MICROKERNEL_TEST_DEPS, |
| 4479 | ) |
| 4480 | |
| 4481 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4482 | name = "f32_vrsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4483 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4484 | "test/f32-vrsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4485 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4486 | ] + MICROKERNEL_TEST_HDRS, |
| 4487 | deps = MICROKERNEL_TEST_DEPS, |
| 4488 | ) |
| 4489 | |
| 4490 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4491 | name = "q8_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4492 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4493 | "test/q8-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4494 | "test/avgpool-microkernel-tester.h", |
| 4495 | "src/xnnpack/AlignedAllocator.h", |
| 4496 | ] + MICROKERNEL_TEST_HDRS, |
| 4497 | deps = MICROKERNEL_TEST_DEPS, |
| 4498 | ) |
| 4499 | |
| 4500 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4501 | name = "q8_igemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4502 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4503 | "test/q8-igemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4504 | "test/gemm-microkernel-tester.h", |
| 4505 | "src/xnnpack/AlignedAllocator.h", |
| 4506 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 4507 | deps = MICROKERNEL_TEST_DEPS, |
| 4508 | ) |
| 4509 | |
| 4510 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4511 | name = "q8_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4512 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4513 | "test/q8-dwconv-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4514 | "test/dwconv-microkernel-tester.h", |
| 4515 | "src/xnnpack/AlignedAllocator.h", |
| 4516 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 4517 | deps = MICROKERNEL_TEST_DEPS, |
| 4518 | ) |
| 4519 | |
| 4520 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4521 | name = "q8_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4522 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4523 | "test/q8-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4524 | "test/gavgpool-microkernel-tester.h", |
| 4525 | "src/xnnpack/AlignedAllocator.h", |
| 4526 | ] + MICROKERNEL_TEST_HDRS, |
| 4527 | deps = MICROKERNEL_TEST_DEPS, |
| 4528 | ) |
| 4529 | |
| 4530 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4531 | name = "q8_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4532 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4533 | "test/q8-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4534 | "test/gemm-microkernel-tester.h", |
| 4535 | "src/xnnpack/AlignedAllocator.h", |
| 4536 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 4537 | deps = MICROKERNEL_TEST_DEPS, |
| 4538 | ) |
| 4539 | |
| 4540 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4541 | name = "q8_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4542 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4543 | "test/q8-vadd-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4544 | "test/vadd-microkernel-tester.h", |
| 4545 | ] + MICROKERNEL_TEST_HDRS, |
| 4546 | deps = MICROKERNEL_TEST_DEPS, |
| 4547 | ) |
| 4548 | |
| 4549 | xnnpack_unit_test( |
| 4550 | name = "u8_clamp_test", |
| 4551 | srcs = [ |
| 4552 | "test/u8-clamp.cc", |
| 4553 | "test/clamp-microkernel-tester.h", |
| 4554 | ] + MICROKERNEL_TEST_HDRS, |
| 4555 | deps = MICROKERNEL_TEST_DEPS, |
| 4556 | ) |
| 4557 | |
| 4558 | xnnpack_unit_test( |
| 4559 | name = "u8_lut32norm_test", |
| 4560 | srcs = [ |
| 4561 | "test/u8-lut32norm.cc", |
| 4562 | "test/lut-norm-microkernel-tester.h", |
| 4563 | ] + MICROKERNEL_TEST_HDRS, |
| 4564 | deps = MICROKERNEL_TEST_DEPS, |
| 4565 | ) |
| 4566 | |
| 4567 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4568 | name = "u8_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4569 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4570 | "test/u8-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4571 | "test/maxpool-microkernel-tester.h", |
| 4572 | ] + MICROKERNEL_TEST_HDRS, |
| 4573 | deps = MICROKERNEL_TEST_DEPS, |
| 4574 | ) |
| 4575 | |
| 4576 | xnnpack_unit_test( |
| 4577 | name = "u8_rmax_test", |
| 4578 | srcs = [ |
| 4579 | "test/u8-rmax.cc", |
| 4580 | "test/rmax-microkernel-tester.h", |
| 4581 | ] + MICROKERNEL_TEST_HDRS, |
| 4582 | deps = MICROKERNEL_TEST_DEPS, |
| 4583 | ) |
| 4584 | |
| 4585 | xnnpack_unit_test( |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 4586 | name = "x32_fill_test", |
| 4587 | srcs = [ |
| 4588 | "test/x32-fill.cc", |
| 4589 | "test/fill-microkernel-tester.h", |
| 4590 | ] + MICROKERNEL_TEST_HDRS, |
| 4591 | deps = MICROKERNEL_TEST_DEPS, |
| 4592 | ) |
| 4593 | |
| 4594 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4595 | name = "x32_packx_test", |
| 4596 | srcs = [ |
| 4597 | "test/x32-packx.cc", |
| 4598 | "test/pack-microkernel-tester.h", |
| 4599 | "src/xnnpack/AlignedAllocator.h", |
| 4600 | ] + MICROKERNEL_TEST_HDRS, |
| 4601 | deps = MICROKERNEL_TEST_DEPS, |
| 4602 | ) |
| 4603 | |
| 4604 | xnnpack_unit_test( |
| 4605 | name = "x32_pad_test", |
| 4606 | srcs = [ |
| 4607 | "test/x32-pad.cc", |
| 4608 | "test/pad-microkernel-tester.h", |
| 4609 | ] + MICROKERNEL_TEST_HDRS, |
| 4610 | deps = MICROKERNEL_TEST_DEPS, |
| 4611 | ) |
| 4612 | |
| 4613 | xnnpack_unit_test( |
| 4614 | name = "x32_unpool_test", |
| 4615 | srcs = [ |
| 4616 | "test/x32-unpool.cc", |
| 4617 | "test/unpool-microkernel-tester.h", |
| 4618 | ] + MICROKERNEL_TEST_HDRS, |
| 4619 | deps = MICROKERNEL_TEST_DEPS, |
| 4620 | ) |
| 4621 | |
| 4622 | xnnpack_unit_test( |
| 4623 | name = "x32_zip_test", |
| 4624 | srcs = [ |
| 4625 | "test/x32-zip.cc", |
| 4626 | "test/zip-microkernel-tester.h", |
| 4627 | ] + MICROKERNEL_TEST_HDRS, |
| 4628 | deps = MICROKERNEL_TEST_DEPS, |
| 4629 | ) |
| 4630 | |
| 4631 | xnnpack_unit_test( |
| 4632 | name = "x8_lut_test", |
| 4633 | srcs = [ |
| 4634 | "test/x8-lut.cc", |
| 4635 | "test/lut-microkernel-tester.h", |
| 4636 | ] + MICROKERNEL_TEST_HDRS, |
| 4637 | deps = MICROKERNEL_TEST_DEPS, |
| 4638 | ) |
| 4639 | |
| 4640 | xnnpack_unit_test( |
| 4641 | name = "x8_zip_test", |
| 4642 | srcs = [ |
| 4643 | "test/x8-zip.cc", |
| 4644 | "test/zip-microkernel-tester.h", |
| 4645 | ] + MICROKERNEL_TEST_HDRS, |
| 4646 | deps = MICROKERNEL_TEST_DEPS, |
| 4647 | ) |
| 4648 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 4649 | xnnpack_unit_test( |
| 4650 | name = "requantization_test", |
| 4651 | srcs = [ |
| 4652 | "src/xnnpack/requantization-stubs.h", |
| 4653 | "test/requantization.cc", |
| 4654 | "test/requantization-tester.h", |
| 4655 | ] + MICROKERNEL_TEST_HDRS, |
| 4656 | deps = MICROKERNEL_TEST_DEPS, |
| 4657 | ) |
| 4658 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 4659 | ########################## Size tests for the library ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4660 | |
| 4661 | xnnpack_binary( |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 4662 | name = "operator_size_test", |
| 4663 | srcs = ["test/operator-size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4664 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4665 | ) |
| 4666 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 4667 | xnnpack_binary( |
| 4668 | name = "subgraph_size_test", |
| 4669 | srcs = ["test/subgraph-size.c"], |
| 4670 | deps = [":XNNPACK"], |
| 4671 | ) |
| 4672 | |
| 4673 | ########################### Unit tests for operators ########################## |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4674 | |
| 4675 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4676 | name = "abs_nc_test", |
| 4677 | srcs = [ |
| 4678 | "test/abs-nc.cc", |
| 4679 | "test/abs-operator-tester.h", |
| 4680 | ], |
| 4681 | deps = OPERATOR_TEST_DEPS, |
| 4682 | ) |
| 4683 | |
| 4684 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 4685 | name = "add_nd_test", |
| 4686 | srcs = [ |
| 4687 | "test/add-nd.cc", |
| 4688 | "test/binary-elementwise-operator-tester.h", |
| 4689 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4690 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 4691 | ) |
| 4692 | |
| 4693 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4694 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4695 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4696 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4697 | "test/argmax-pooling-operator-tester.h", |
| 4698 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4699 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4700 | ) |
| 4701 | |
| 4702 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4703 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4704 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4705 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4706 | "test/average-pooling-operator-tester.h", |
| 4707 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4708 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4709 | ) |
| 4710 | |
| 4711 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 4712 | name = "bankers_rounding_nc_test", |
| 4713 | srcs = [ |
| 4714 | "test/bankers-rounding-nc.cc", |
| 4715 | "test/bankers-rounding-operator-tester.h", |
| 4716 | ], |
| 4717 | deps = OPERATOR_TEST_DEPS, |
| 4718 | ) |
| 4719 | |
| 4720 | xnnpack_unit_test( |
| 4721 | name = "ceiling_nc_test", |
| 4722 | srcs = [ |
| 4723 | "test/ceiling-nc.cc", |
| 4724 | "test/ceiling-operator-tester.h", |
| 4725 | ], |
| 4726 | deps = OPERATOR_TEST_DEPS, |
| 4727 | ) |
| 4728 | |
| 4729 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4730 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4731 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4732 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4733 | "test/channel-shuffle-operator-tester.h", |
| 4734 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4735 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4736 | ) |
| 4737 | |
| 4738 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4739 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4740 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4741 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4742 | "test/clamp-operator-tester.h", |
| 4743 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4744 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4745 | ) |
| 4746 | |
| 4747 | xnnpack_unit_test( |
Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 4748 | name = "constant_pad_nd_test", |
| 4749 | srcs = [ |
| 4750 | "test/constant-pad-nd.cc", |
| 4751 | "test/constant-pad-operator-tester.h", |
| 4752 | ], |
| 4753 | deps = OPERATOR_TEST_DEPS, |
| 4754 | ) |
| 4755 | |
| 4756 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4757 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4758 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4759 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4760 | "test/convolution-operator-tester.h", |
| 4761 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4762 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4763 | ) |
| 4764 | |
| 4765 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4766 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4767 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4768 | "test/convolution-nchw.cc", |
| 4769 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4770 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4771 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4772 | ) |
| 4773 | |
| 4774 | xnnpack_unit_test( |
Marat Dukhan | 4e21b27 | 2020-06-04 18:45:01 -0700 | [diff] [blame] | 4775 | name = "copy_nc_test", |
| 4776 | srcs = [ |
| 4777 | "test/copy-nc.cc", |
| 4778 | "test/copy-operator-tester.h", |
| 4779 | ], |
| 4780 | deps = OPERATOR_TEST_DEPS, |
| 4781 | ) |
| 4782 | |
| 4783 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4784 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4785 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4786 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4787 | "test/deconvolution-operator-tester.h", |
| 4788 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4789 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4790 | ) |
| 4791 | |
| 4792 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 4793 | name = "divide_nd_test", |
| 4794 | srcs = [ |
| 4795 | "test/binary-elementwise-operator-tester.h", |
| 4796 | "test/divide-nd.cc", |
| 4797 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4798 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 4799 | ) |
| 4800 | |
| 4801 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4802 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4803 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4804 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4805 | "test/fully-connected-operator-tester.h", |
| 4806 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4807 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4808 | ) |
| 4809 | |
| 4810 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 4811 | name = "floor_nc_test", |
| 4812 | srcs = [ |
| 4813 | "test/floor-nc.cc", |
| 4814 | "test/floor-operator-tester.h", |
| 4815 | ], |
| 4816 | deps = OPERATOR_TEST_DEPS, |
| 4817 | ) |
| 4818 | |
| 4819 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4820 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4821 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4822 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4823 | "test/global-average-pooling-operator-tester.h", |
| 4824 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4825 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4826 | ) |
| 4827 | |
| 4828 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4829 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4830 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4831 | "test/global-average-pooling-ncw.cc", |
| 4832 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4833 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4834 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4835 | ) |
| 4836 | |
| 4837 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4838 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4839 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4840 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4841 | "test/hardswish-operator-tester.h", |
| 4842 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4843 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4844 | ) |
| 4845 | |
| 4846 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4847 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4848 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4849 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4850 | "test/leaky-relu-operator-tester.h", |
| 4851 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4852 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4853 | ) |
| 4854 | |
| 4855 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4856 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4857 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4858 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4859 | "test/max-pooling-operator-tester.h", |
| 4860 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4861 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4862 | ) |
| 4863 | |
| 4864 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 4865 | name = "maximum_nd_test", |
| 4866 | srcs = [ |
| 4867 | "test/binary-elementwise-operator-tester.h", |
| 4868 | "test/maximum-nd.cc", |
| 4869 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4870 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 4871 | ) |
| 4872 | |
| 4873 | xnnpack_unit_test( |
| 4874 | name = "minimum_nd_test", |
| 4875 | srcs = [ |
| 4876 | "test/binary-elementwise-operator-tester.h", |
| 4877 | "test/minimum-nd.cc", |
| 4878 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4879 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 4880 | ) |
| 4881 | |
| 4882 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4883 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 4884 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 4885 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4886 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 4887 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4888 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 4889 | ) |
| 4890 | |
| 4891 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4892 | name = "negate_nc_test", |
| 4893 | srcs = [ |
| 4894 | "test/negate-nc.cc", |
| 4895 | "test/negate-operator-tester.h", |
| 4896 | ], |
| 4897 | deps = OPERATOR_TEST_DEPS, |
| 4898 | ) |
| 4899 | |
| 4900 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4901 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4902 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4903 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4904 | "test/prelu-operator-tester.h", |
| 4905 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4906 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4907 | ) |
| 4908 | |
| 4909 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4910 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 4911 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4912 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 4913 | "test/resize-bilinear-operator-tester.h", |
| 4914 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4915 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 4916 | ) |
| 4917 | |
| 4918 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4919 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4920 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4921 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4922 | "test/sigmoid-operator-tester.h", |
| 4923 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4924 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4925 | ) |
| 4926 | |
| 4927 | xnnpack_unit_test( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4928 | name = "softmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4929 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4930 | "test/softmax-nc.cc", |
| 4931 | "test/softmax-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4932 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4933 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4934 | ) |
| 4935 | |
| 4936 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4937 | name = "square_nc_test", |
| 4938 | srcs = [ |
| 4939 | "test/square-nc.cc", |
| 4940 | "test/square-operator-tester.h", |
| 4941 | ], |
| 4942 | deps = OPERATOR_TEST_DEPS, |
| 4943 | ) |
| 4944 | |
| 4945 | xnnpack_unit_test( |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 4946 | name = "squared_difference_nd_test", |
| 4947 | srcs = [ |
| 4948 | "test/binary-elementwise-operator-tester.h", |
| 4949 | "test/squared-difference-nd.cc", |
| 4950 | ], |
| 4951 | deps = OPERATOR_TEST_DEPS, |
| 4952 | ) |
| 4953 | |
| 4954 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 4955 | name = "subtract_nd_test", |
| 4956 | srcs = [ |
| 4957 | "test/binary-elementwise-operator-tester.h", |
| 4958 | "test/subtract-nd.cc", |
| 4959 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4960 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 4961 | ) |
| 4962 | |
| 4963 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 4964 | name = "truncation_nc_test", |
| 4965 | srcs = [ |
| 4966 | "test/truncation-nc.cc", |
| 4967 | "test/truncation-operator-tester.h", |
| 4968 | ], |
| 4969 | deps = OPERATOR_TEST_DEPS, |
| 4970 | ) |
| 4971 | |
| 4972 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4973 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4974 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4975 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4976 | "test/unpooling-operator-tester.h", |
| 4977 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4978 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4979 | ) |
| 4980 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4981 | ############################### Misc unit tests ############################### |
| 4982 | |
| 4983 | xnnpack_unit_test( |
| 4984 | name = "memory_planner_test", |
| 4985 | srcs = [ |
| 4986 | "test/memory-planner-test.cc", |
| 4987 | ], |
| 4988 | deps = [ |
| 4989 | ":XNNPACK", |
| 4990 | ":memory_planner", |
| 4991 | ], |
| 4992 | ) |
| 4993 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4994 | ############################# Build configurations ############################# |
| 4995 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 4996 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4997 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 4998 | name = "xnn_enable_assembly_explicit_true", |
| 4999 | define_values = {"xnn_enable_assembly": "true"}, |
| 5000 | ) |
| 5001 | |
| 5002 | # Disables usage of assembly kernels. |
| 5003 | config_setting( |
| 5004 | name = "xnn_enable_assembly_explicit_false", |
| 5005 | define_values = {"xnn_enable_assembly": "false"}, |
| 5006 | ) |
| 5007 | |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 5008 | # Disables usage of HMP-aware optimizations. |
| 5009 | config_setting( |
| 5010 | name = "xnn_enable_hmp_explicit_false", |
| 5011 | define_values = {"xnn_enable_hmp": "false"}, |
| 5012 | ) |
| 5013 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 5014 | # Enable usage of optimized memory allocation |
| 5015 | config_setting( |
| 5016 | name = "xnn_enable_memopt_explicit_true", |
| 5017 | define_values = {"xnnpack_enable_memopt": "true"}, |
| 5018 | ) |
| 5019 | |
| 5020 | # Disable usage of optimized memory allocation |
| 5021 | config_setting( |
| 5022 | name = "xnn_enable_memopt_explicit_false", |
| 5023 | define_values = {"xnnpack_enable_memopt": "false"}, |
| 5024 | ) |
| 5025 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5026 | # Builds with -c dbg |
| 5027 | config_setting( |
| 5028 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5029 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5030 | "compilation_mode": "dbg", |
| 5031 | }, |
| 5032 | ) |
| 5033 | |
| 5034 | # Builds with -c opt |
| 5035 | config_setting( |
| 5036 | name = "optimized_build", |
| 5037 | values = { |
| 5038 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5039 | }, |
| 5040 | ) |
| 5041 | |
| 5042 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5043 | name = "linux_k8", |
| 5044 | values = {"cpu": "k8"}, |
| 5045 | ) |
| 5046 | |
| 5047 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 5048 | name = "linux_arm", |
| 5049 | values = {"cpu": "arm"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 5050 | ) |
| 5051 | |
| 5052 | config_setting( |
Marat Dukhan | f0bd4de | 2020-06-15 15:53:19 -0700 | [diff] [blame] | 5053 | name = "linux_armeabi", |
| 5054 | values = {"cpu": "armeabi"}, |
| 5055 | ) |
| 5056 | |
| 5057 | config_setting( |
Terry Heo | 68eef3f | 2020-04-13 22:53:52 -0700 | [diff] [blame] | 5058 | name = "linux_armhf", |
| 5059 | values = {"cpu": "armhf"}, |
| 5060 | ) |
| 5061 | |
| 5062 | config_setting( |
Marat Dukhan | a720e93 | 2020-06-10 13:01:11 -0700 | [diff] [blame] | 5063 | name = "linux_armv7a", |
| 5064 | values = {"cpu": "armv7a"}, |
| 5065 | ) |
| 5066 | |
| 5067 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 5068 | name = "linux_aarch64", |
| 5069 | values = {"cpu": "aarch64"}, |
| 5070 | ) |
| 5071 | |
| 5072 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5073 | name = "android", |
| 5074 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 5075 | ) |
| 5076 | |
| 5077 | config_setting( |
| 5078 | name = "android_armv7", |
| 5079 | values = { |
| 5080 | "crosstool_top": "//external:android/crosstool", |
| 5081 | "cpu": "armeabi-v7a", |
| 5082 | }, |
| 5083 | ) |
| 5084 | |
| 5085 | config_setting( |
| 5086 | name = "android_arm64", |
| 5087 | values = { |
| 5088 | "crosstool_top": "//external:android/crosstool", |
| 5089 | "cpu": "arm64-v8a", |
| 5090 | }, |
| 5091 | ) |
| 5092 | |
| 5093 | config_setting( |
| 5094 | name = "android_x86", |
| 5095 | values = { |
| 5096 | "crosstool_top": "//external:android/crosstool", |
| 5097 | "cpu": "x86", |
| 5098 | }, |
| 5099 | ) |
| 5100 | |
| 5101 | config_setting( |
| 5102 | name = "android_x86_64", |
| 5103 | values = { |
| 5104 | "crosstool_top": "//external:android/crosstool", |
| 5105 | "cpu": "x86_64", |
| 5106 | }, |
| 5107 | ) |
| 5108 | |
| 5109 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 5110 | name = "windows_x86_64", |
| 5111 | values = {"cpu": "x64_windows"}, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 5112 | ) |
| 5113 | |
| 5114 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 5115 | name = "windows_x86_64_clang", |
| 5116 | values = { |
| 5117 | "compiler": "clang-cl", |
| 5118 | "cpu": "x64_windows", |
| 5119 | }, |
| 5120 | ) |
| 5121 | |
| 5122 | config_setting( |
| 5123 | name = "windows_x86_64_mingw", |
| 5124 | values = { |
| 5125 | "compiler": "mingw-gcc", |
| 5126 | "cpu": "x64_windows", |
| 5127 | }, |
| 5128 | ) |
| 5129 | |
| 5130 | config_setting( |
| 5131 | name = "windows_x86_64_msys", |
| 5132 | values = { |
| 5133 | "compiler": "msys-gcc", |
| 5134 | "cpu": "x64_windows", |
| 5135 | }, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 5136 | ) |
| 5137 | |
| 5138 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 5139 | name = "macos_x86_64", |
| 5140 | values = { |
| 5141 | "apple_platform_type": "macos", |
| 5142 | "cpu": "darwin", |
| 5143 | }, |
| 5144 | ) |
| 5145 | |
| 5146 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5147 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5148 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5149 | ) |
| 5150 | |
| 5151 | config_setting( |
| 5152 | name = "emscripten_wasm", |
| 5153 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5154 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5155 | "cpu": "wasm", |
| 5156 | }, |
| 5157 | ) |
| 5158 | |
| 5159 | config_setting( |
| 5160 | name = "emscripten_wasmsimd", |
| 5161 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5162 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5163 | "cpu": "wasm", |
Marat Dukhan | 81c6260 | 2020-05-29 13:22:49 -0700 | [diff] [blame] | 5164 | "copt": "-msimd128", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5165 | }, |
| 5166 | ) |
| 5167 | |
| 5168 | config_setting( |
| 5169 | name = "emscripten_asmjs", |
| 5170 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5171 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5172 | "cpu": "asmjs", |
| 5173 | }, |
| 5174 | ) |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5175 | |
| 5176 | config_setting( |
| 5177 | name = "ios_armv7", |
| 5178 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5179 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5180 | "cpu": "ios_armv7", |
| 5181 | }, |
| 5182 | ) |
| 5183 | |
| 5184 | config_setting( |
| 5185 | name = "ios_arm64", |
| 5186 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5187 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5188 | "cpu": "ios_arm64", |
| 5189 | }, |
| 5190 | ) |
| 5191 | |
| 5192 | config_setting( |
| 5193 | name = "ios_arm64e", |
| 5194 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5195 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5196 | "cpu": "ios_arm64e", |
| 5197 | }, |
| 5198 | ) |
| 5199 | |
| 5200 | config_setting( |
| 5201 | name = "ios_x86", |
| 5202 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5203 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5204 | "cpu": "ios_i386", |
| 5205 | }, |
| 5206 | ) |
| 5207 | |
| 5208 | config_setting( |
| 5209 | name = "ios_x86_64", |
| 5210 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5211 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5212 | "cpu": "ios_x86_64", |
| 5213 | }, |
| 5214 | ) |
| 5215 | |
| 5216 | config_setting( |
| 5217 | name = "watchos_armv7k", |
| 5218 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5219 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5220 | "cpu": "watchos_armv7k", |
| 5221 | }, |
| 5222 | ) |
| 5223 | |
| 5224 | config_setting( |
| 5225 | name = "watchos_arm64_32", |
| 5226 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5227 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5228 | "cpu": "watchos_arm64_32", |
| 5229 | }, |
| 5230 | ) |
| 5231 | |
| 5232 | config_setting( |
| 5233 | name = "watchos_x86", |
| 5234 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5235 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5236 | "cpu": "watchos_i386", |
| 5237 | }, |
| 5238 | ) |
| 5239 | |
| 5240 | config_setting( |
| 5241 | name = "watchos_x86_64", |
| 5242 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5243 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5244 | "cpu": "watchos_x86_64", |
| 5245 | }, |
| 5246 | ) |
| 5247 | |
| 5248 | config_setting( |
| 5249 | name = "tvos_arm64", |
| 5250 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5251 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5252 | "cpu": "tvos_arm64", |
| 5253 | }, |
| 5254 | ) |
| 5255 | |
| 5256 | config_setting( |
| 5257 | name = "tvos_x86_64", |
| 5258 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5259 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5260 | "cpu": "tvos_x86_64", |
| 5261 | }, |
| 5262 | ) |