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", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 232 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 233 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 234 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 235 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c", |
| 236 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c", |
| 237 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 238 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 239 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 240 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 241 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c", |
| 242 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c", |
| 243 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 244 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 245 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 246 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 247 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c", |
| 248 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c", |
| 249 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 250 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 251 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 252 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 253 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 254 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 255 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| 256 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 257 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 258 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| 259 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 260 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 261 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| 262 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 263 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 264 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 265 | "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c", |
| 266 | "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c", |
| 267 | "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 268 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 269 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 270 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 271 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c", |
| 272 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c", |
| 273 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 274 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 275 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 276 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 277 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c", |
| 278 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c", |
| 279 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 280 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 281 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 282 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 283 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c", |
| 284 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c", |
| 285 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 286 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 287 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 288 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 289 | "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c", |
| 290 | "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c", |
| 291 | "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c", |
| 292 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c", |
| 293 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c", |
| 294 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 295 | "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c", |
| 296 | "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c", |
| 297 | "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 298 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 299 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 300 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 301 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c", |
| 302 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c", |
| 303 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 304 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 305 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 306 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 307 | "src/f32-vlrelu/gen/scalar-x1.c", |
| 308 | "src/f32-vlrelu/gen/scalar-x2.c", |
| 309 | "src/f32-vlrelu/gen/scalar-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 310 | "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c", |
| 311 | "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c", |
| 312 | "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 313 | "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c", |
| 314 | "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c", |
| 315 | "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c", |
| 316 | "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c", |
| 317 | "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c", |
| 318 | "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c", |
| 319 | "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c", |
| 320 | "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c", |
| 321 | "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c", |
| 322 | "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c", |
| 323 | "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c", |
| 324 | "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 325 | "src/f32-vsqrt/gen/scalar-sqrt-x1.c", |
| 326 | "src/f32-vsqrt/gen/scalar-sqrt-x2.c", |
| 327 | "src/f32-vsqrt/gen/scalar-sqrt-x4.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 328 | "src/f32-vunary/gen/vabs-scalar-x1.c", |
| 329 | "src/f32-vunary/gen/vabs-scalar-x2.c", |
| 330 | "src/f32-vunary/gen/vabs-scalar-x4.c", |
| 331 | "src/f32-vunary/gen/vneg-scalar-x1.c", |
| 332 | "src/f32-vunary/gen/vneg-scalar-x2.c", |
| 333 | "src/f32-vunary/gen/vneg-scalar-x4.c", |
| 334 | "src/f32-vunary/gen/vsqr-scalar-x1.c", |
| 335 | "src/f32-vunary/gen/vsqr-scalar-x2.c", |
| 336 | "src/f32-vunary/gen/vsqr-scalar-x4.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 337 | "src/math/expminus-scalar-lut2048-p1.c", |
| 338 | "src/math/expminus-scalar-lut64-p2.c", |
| 339 | "src/math/expminus-scalar-p5.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 340 | "src/math/roundne-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 341 | "src/math/roundne-scalar-nearbyint.c", |
| 342 | "src/math/roundne-scalar-rint.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 343 | "src/math/roundd-scalar-addsub.c", |
| 344 | "src/math/roundd-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 345 | "src/math/roundd-scalar-floor.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 346 | "src/math/roundu-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 347 | "src/math/roundu-scalar-ceil.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 348 | "src/math/roundu-scalar-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 349 | "src/math/roundz-scalar-addsub.c", |
| 350 | "src/math/roundz-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 351 | "src/math/roundz-scalar-trunc.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 352 | "src/math/sigmoid-scalar-lut2048-p1-div.c", |
| 353 | "src/math/sigmoid-scalar-lut64-p2-div.c", |
| 354 | "src/math/sigmoid-scalar-p5-div.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 355 | "src/q8-avgpool/9p8x-minmax-scalar-c1.c", |
| 356 | "src/q8-avgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 357 | "src/q8-dwconv/up1x9-minmax-scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 358 | "src/q8-gavgpool/7p7x-minmax-scalar-c1.c", |
| 359 | "src/q8-gavgpool/7x-minmax-scalar-c1.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 360 | "src/q8-gemm/2x2-minmax-scalar.c", |
| 361 | "src/q8-igemm/2x2-minmax-scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 362 | "src/q8-vadd/minmax-scalar.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 363 | "src/u8-clamp/scalar-x4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 364 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 365 | "src/u8-maxpool/9p8x-minmax-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 366 | "src/u8-rmax/scalar.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 367 | "src/x32-fill/scalar-float.c", |
| 368 | "src/x32-fill/scalar-int.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 369 | "src/x32-packx/x2-scalar.c", |
| 370 | "src/x32-packx/x3-scalar.c", |
| 371 | "src/x32-packx/x4-scalar.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 372 | "src/x32-pad/scalar-int.c", |
| 373 | "src/x32-pad/scalar-float.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 374 | "src/x32-unpool/scalar.c", |
| 375 | "src/x32-zip/x2-scalar.c", |
| 376 | "src/x32-zip/x3-scalar.c", |
| 377 | "src/x32-zip/x4-scalar.c", |
| 378 | "src/x32-zip/xm-scalar.c", |
| 379 | "src/x8-lut/scalar.c", |
| 380 | "src/x8-zip/x2-scalar.c", |
| 381 | "src/x8-zip/x3-scalar.c", |
| 382 | "src/x8-zip/x4-scalar.c", |
| 383 | "src/x8-zip/xm-scalar.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 384 | "src/requantization/precise-scalar.c", |
| 385 | "src/requantization/fp32-scalar.c", |
| 386 | "src/requantization/q31-scalar.c", |
| 387 | "src/requantization/gemmlowp-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 388 | ] |
| 389 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 390 | WASM_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 391 | "src/f32-avgpool/9p8x-minmax-wasm-c1.c", |
| 392 | "src/f32-avgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 393 | "src/f32-clamp/gen/wasm-x1.c", |
| 394 | "src/f32-clamp/gen/wasm-x2.c", |
| 395 | "src/f32-clamp/gen/wasm-x4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 396 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 397 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 398 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 399 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 400 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 401 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 402 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 403 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 404 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 405 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| 406 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 407 | "src/f32-dwconv/gen/up2x25-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 408 | "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c", |
| 409 | "src/f32-dwconv/gen/up1x4-minmax-wasm.c", |
| 410 | "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c", |
| 411 | "src/f32-dwconv/gen/up1x9-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 412 | "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c", |
| 413 | "src/f32-dwconv/gen/up1x25-minmax-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 414 | "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c", |
| 415 | "src/f32-dwconv/gen/up2x4-minmax-wasm.c", |
| 416 | "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c", |
| 417 | "src/f32-dwconv/gen/up2x9-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 418 | "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c", |
| 419 | "src/f32-dwconv/gen/up2x25-minmax-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 420 | "src/f32-gavgpool/7p7x-minmax-wasm-c1.c", |
| 421 | "src/f32-gavgpool/7x-minmax-wasm-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 422 | "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c", |
| 423 | "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c", |
| 424 | "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 425 | "src/f32-gemm/gen/1x4-wasm.c", |
| 426 | "src/f32-gemm/gen/2x4-wasm.c", |
| 427 | "src/f32-gemm/gen/4x2-wasm.c", |
| 428 | "src/f32-gemm/gen/4x4-wasm.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 429 | "src/f32-gemm/gen/1x4-relu-wasm.c", |
| 430 | "src/f32-gemm/gen/2x4-relu-wasm.c", |
| 431 | "src/f32-gemm/gen/4x2-relu-wasm.c", |
| 432 | "src/f32-gemm/gen/4x4-relu-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 433 | "src/f32-gemm/gen/1x4-minmax-wasm.c", |
| 434 | "src/f32-gemm/gen/2x4-minmax-wasm.c", |
| 435 | "src/f32-gemm/gen/4x2-minmax-wasm.c", |
| 436 | "src/f32-gemm/gen/4x4-minmax-wasm.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 437 | "src/f32-hswish/gen/wasm-x1.c", |
| 438 | "src/f32-hswish/gen/wasm-x2.c", |
| 439 | "src/f32-hswish/gen/wasm-x4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 440 | "src/f32-igemm/gen/1x4-wasm.c", |
| 441 | "src/f32-igemm/gen/2x4-wasm.c", |
| 442 | "src/f32-igemm/gen/4x2-wasm.c", |
| 443 | "src/f32-igemm/gen/4x4-wasm.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 444 | "src/f32-igemm/gen/1x4-relu-wasm.c", |
| 445 | "src/f32-igemm/gen/2x4-relu-wasm.c", |
| 446 | "src/f32-igemm/gen/4x2-relu-wasm.c", |
| 447 | "src/f32-igemm/gen/4x4-relu-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 448 | "src/f32-igemm/gen/1x4-minmax-wasm.c", |
| 449 | "src/f32-igemm/gen/2x4-minmax-wasm.c", |
| 450 | "src/f32-igemm/gen/4x2-minmax-wasm.c", |
| 451 | "src/f32-igemm/gen/4x4-minmax-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 452 | "src/f32-maxpool/9p8x-minmax-wasm-c1.c", |
| 453 | "src/f32-pavgpool/9p8x-minmax-wasm-c1.c", |
| 454 | "src/f32-pavgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 7c1f808 | 2020-06-25 13:26:20 -0700 | [diff] [blame] | 455 | "src/f32-prelu/gen/wasm-2x1.c", |
| 456 | "src/f32-prelu/gen/wasm-2x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 457 | "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c", |
| 458 | "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c", |
| 459 | "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c", |
| 460 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c", |
| 461 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c", |
| 462 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c", |
| 463 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c", |
| 464 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c", |
| 465 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c", |
| 466 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c", |
| 467 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c", |
| 468 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 469 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 470 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 471 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
| 472 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 473 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 474 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
| 475 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 476 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 477 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
| 478 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 479 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 480 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 481 | "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c", |
| 482 | "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c", |
| 483 | "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c", |
| 484 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c", |
| 485 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c", |
| 486 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c", |
| 487 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c", |
| 488 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c", |
| 489 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c", |
| 490 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c", |
| 491 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c", |
| 492 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c", |
| 493 | "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c", |
| 494 | "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c", |
| 495 | "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c", |
| 496 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c", |
| 497 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c", |
| 498 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 499 | "src/f32-vlrelu/gen/wasm-x1.c", |
| 500 | "src/f32-vlrelu/gen/wasm-x2.c", |
| 501 | "src/f32-vlrelu/gen/wasm-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 502 | "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c", |
| 503 | "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c", |
| 504 | "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 505 | ] |
| 506 | |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 507 | WASMSIMD_UKERNELS = [ |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame^] | 508 | "src/f32-clamp/gen/wasmsimd-arm-x4.c", |
| 509 | "src/f32-clamp/gen/wasmsimd-arm-x8.c", |
| 510 | "src/f32-clamp/gen/wasmsimd-x86-x4.c", |
| 511 | "src/f32-clamp/gen/wasmsimd-x86-x8.c", |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 512 | "src/f32-dwconv/gen/up4x25-wasmsimd.c", |
| 513 | "src/f32-dwconv/gen/up4x4-wasmsimd.c", |
| 514 | "src/f32-dwconv/gen/up4x9-wasmsimd.c", |
| 515 | "src/f32-dwconv/gen/up8x25-wasmsimd.c", |
| 516 | "src/f32-dwconv/gen/up8x4-wasmsimd.c", |
| 517 | "src/f32-dwconv/gen/up8x9-wasmsimd.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 518 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c", |
| 519 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c", |
| 520 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c", |
| 521 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c", |
| 522 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c", |
| 523 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c", |
| 524 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c", |
| 525 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c", |
| 526 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c", |
| 527 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c", |
| 528 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c", |
| 529 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c", |
| 530 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c", |
| 531 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c", |
| 532 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c", |
| 533 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c", |
| 534 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c", |
| 535 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c", |
| 536 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c", |
| 537 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c", |
| 538 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c", |
| 539 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c", |
| 540 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c", |
| 541 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c", |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 542 | "src/f32-gemm/gen/1x8-wasmsimd-splat.c", |
| 543 | "src/f32-gemm/gen/4x8-wasmsimd-splat.c", |
| 544 | "src/f32-gemm/gen/5x8-wasmsimd-splat.c", |
| 545 | "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c", |
| 546 | "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c", |
| 547 | "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c", |
| 548 | "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c", |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 549 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 550 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 551 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 552 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 553 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 554 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 555 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 556 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 557 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 558 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 559 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 560 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 561 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 562 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 563 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 564 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 565 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 566 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 567 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 568 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 569 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 570 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 571 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 572 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 573 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 574 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 575 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 576 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 577 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 578 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
| 579 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 580 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c", |
| 581 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c", |
| 582 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 583 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c", |
| 584 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c", |
| 585 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 586 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c", |
| 587 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c", |
| 588 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 589 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c", |
| 590 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c", |
| 591 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 592 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c", |
| 593 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c", |
| 594 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 595 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c", |
| 596 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c", |
| 597 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 598 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c", |
| 599 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c", |
| 600 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 601 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c", |
| 602 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c", |
| 603 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 604 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c", |
| 605 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c", |
| 606 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 607 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c", |
| 608 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c", |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 609 | "src/f32-igemm/gen/1x8-wasmsimd-splat.c", |
| 610 | "src/f32-igemm/gen/4x8-wasmsimd-splat.c", |
| 611 | "src/f32-igemm/gen/5x8-wasmsimd-splat.c", |
| 612 | "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c", |
| 613 | "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c", |
| 614 | "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c", |
| 615 | "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c", |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 616 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 617 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 618 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 619 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 620 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 621 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 622 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 623 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 624 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 625 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 626 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 627 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 628 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 629 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 630 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 631 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 632 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 633 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 634 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 635 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 636 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 637 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 638 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 639 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 640 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 641 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 642 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 643 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 644 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 645 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 646 | "src/f32-hswish/gen/wasmsimd-arm-x4.c", |
| 647 | "src/f32-hswish/gen/wasmsimd-arm-x8.c", |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 648 | "src/f32-hswish/gen/wasmsimd-arm-x16.c", |
Marat Dukhan | 9baec80 | 2020-06-25 21:34:35 -0700 | [diff] [blame] | 649 | "src/f32-hswish/gen/wasmsimd-x86-x4.c", |
| 650 | "src/f32-hswish/gen/wasmsimd-x86-x8.c", |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 651 | "src/f32-hswish/gen/wasmsimd-x86-x16.c", |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 652 | "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c", |
| 653 | "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c", |
| 654 | "src/f32-prelu/gen/wasmsimd-minmax-2x4.c", |
| 655 | "src/f32-prelu/gen/wasmsimd-minmax-2x8.c", |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 656 | "src/f32-rmax/wasmsimd-arm.c", |
| 657 | "src/f32-rmax/wasmsimd-x86.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 658 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c", |
| 659 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c", |
| 660 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c", |
| 661 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 662 | "src/f32-vbinary/gen/vadd-wasmsimd-x4.c", |
| 663 | "src/f32-vbinary/gen/vadd-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 664 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c", |
| 665 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c", |
| 666 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c", |
| 667 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 668 | "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c", |
| 669 | "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 670 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c", |
| 671 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c", |
| 672 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c", |
| 673 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 674 | "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c", |
| 675 | "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 676 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c", |
| 677 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c", |
| 678 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c", |
| 679 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 680 | "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c", |
| 681 | "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 682 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c", |
| 683 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c", |
| 684 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c", |
| 685 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 686 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c", |
| 687 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c", |
| 688 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c", |
| 689 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 690 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c", |
| 691 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c", |
| 692 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c", |
| 693 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 694 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c", |
| 695 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c", |
| 696 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c", |
| 697 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 698 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c", |
| 699 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c", |
| 700 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c", |
| 701 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 702 | "src/f32-vbinary/gen/vmul-wasmsimd-x4.c", |
| 703 | "src/f32-vbinary/gen/vmul-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 704 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c", |
| 705 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c", |
| 706 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c", |
| 707 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 708 | "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c", |
| 709 | "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 710 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c", |
| 711 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c", |
| 712 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c", |
| 713 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 714 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c", |
| 715 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 716 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c", |
| 717 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c", |
| 718 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c", |
| 719 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 720 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c", |
| 721 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 722 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c", |
| 723 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 724 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c", |
| 725 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 726 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c", |
| 727 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c", |
| 728 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c", |
| 729 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 730 | "src/f32-vbinary/gen/vsub-wasmsimd-x4.c", |
| 731 | "src/f32-vbinary/gen/vsub-wasmsimd-x8.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 732 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c", |
| 733 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c", |
| 734 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c", |
| 735 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 736 | "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c", |
| 737 | "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c", |
Marat Dukhan | 72b399a | 2020-06-28 22:41:38 -0700 | [diff] [blame] | 738 | "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c", |
| 739 | "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c", |
| 740 | "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c", |
| 741 | "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c", |
| 742 | "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c", |
| 743 | "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c", |
| 744 | "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c", |
| 745 | "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 746 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c", |
| 747 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c", |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 748 | "src/f32-vunary/gen/vabs-wasmsimd-x4.c", |
| 749 | "src/f32-vunary/gen/vabs-wasmsimd-x8.c", |
| 750 | "src/f32-vunary/gen/vneg-wasmsimd-x4.c", |
| 751 | "src/f32-vunary/gen/vneg-wasmsimd-x8.c", |
| 752 | "src/f32-vunary/gen/vsqr-wasmsimd-x4.c", |
| 753 | "src/f32-vunary/gen/vsqr-wasmsimd-x8.c", |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 754 | "src/math/roundne-wasmsimd-addsub.c", |
| 755 | "src/math/roundd-wasmsimd-addsub.c", |
| 756 | "src/math/roundu-wasmsimd-addsub.c", |
| 757 | "src/math/roundz-wasmsimd-addsub.c", |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 758 | "src/requantization/fp32-wasmsimd.c", |
| 759 | ] |
| 760 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 761 | PSIMD_FASTMATH_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 762 | "src/f32-argmaxpool/4x-psimd-c4.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 763 | "src/f32-argmaxpool/9p8x-psimd-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 764 | "src/f32-argmaxpool/9x-psimd-c4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 765 | "src/f32-avgpool/9p8x-minmax-psimd-c4.c", |
| 766 | "src/f32-avgpool/9x-minmax-psimd-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 767 | "src/f32-clamp/gen/psimd-x4.c", |
| 768 | "src/f32-clamp/gen/psimd-x8.c", |
Erich Elsen | 0a1970e | 2020-06-10 09:24:59 -0700 | [diff] [blame] | 769 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c", |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 770 | "src/f32-dwconv-chw/3x3p1-psimd.c", |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 771 | "src/f32-dwconv-chw/3x3s2p1-psimd.c", |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 772 | "src/f32-dwconv-chw/5x5p2-psimd.c", |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 773 | "src/f32-dwconv-chw/5x5s2p2-psimd.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 774 | "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c", |
| 775 | "src/f32-dwconv/gen/up4x25-minmax-psimd.c", |
| 776 | "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c", |
| 777 | "src/f32-dwconv/gen/up4x4-minmax-psimd.c", |
| 778 | "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c", |
| 779 | "src/f32-dwconv/gen/up4x9-minmax-psimd.c", |
| 780 | "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c", |
| 781 | "src/f32-dwconv/gen/up8x25-minmax-psimd.c", |
| 782 | "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c", |
| 783 | "src/f32-dwconv/gen/up8x4-minmax-psimd.c", |
| 784 | "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c", |
| 785 | "src/f32-dwconv/gen/up8x9-minmax-psimd.c", |
Erich Elsen | 6f278b5 | 2020-06-10 16:13:11 -0700 | [diff] [blame] | 786 | "src/f32-gavgpool-cw/psimd-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 787 | "src/f32-gavgpool/7p7x-minmax-psimd-c4.c", |
| 788 | "src/f32-gavgpool/7x-minmax-psimd-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 789 | "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c", |
| 790 | "src/f32-gemm/gen/1x8-minmax-psimd-splat.c", |
| 791 | "src/f32-gemm/gen/1x8s4-minmax-psimd.c", |
| 792 | "src/f32-gemm/gen/4x2c4-minmax-psimd.c", |
| 793 | "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c", |
| 794 | "src/f32-gemm/gen/4x8-minmax-psimd-splat.c", |
| 795 | "src/f32-gemm/gen/4x8s4-minmax-psimd.c", |
| 796 | "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c", |
| 797 | "src/f32-gemm/gen/6x8-minmax-psimd-splat.c", |
| 798 | "src/f32-gemm/gen/6x8s4-minmax-psimd.c", |
| 799 | "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c", |
| 800 | "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c", |
| 801 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c", |
| 802 | "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c", |
| 803 | "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c", |
| 804 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c", |
| 805 | "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c", |
| 806 | "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c", |
| 807 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 808 | "src/f32-hswish/gen/psimd-x4.c", |
| 809 | "src/f32-hswish/gen/psimd-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 810 | "src/f32-ibilinear/gen/psimd-c4.c", |
| 811 | "src/f32-ibilinear/gen/psimd-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 812 | "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c", |
| 813 | "src/f32-igemm/gen/1x8-minmax-psimd-splat.c", |
| 814 | "src/f32-igemm/gen/1x8s4-minmax-psimd.c", |
| 815 | "src/f32-igemm/gen/4x2c4-minmax-psimd.c", |
| 816 | "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c", |
| 817 | "src/f32-igemm/gen/4x8-minmax-psimd-splat.c", |
| 818 | "src/f32-igemm/gen/4x8s4-minmax-psimd.c", |
| 819 | "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c", |
| 820 | "src/f32-igemm/gen/6x8-minmax-psimd-splat.c", |
| 821 | "src/f32-igemm/gen/6x8s4-minmax-psimd.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 822 | "src/f32-maxpool/9p8x-minmax-psimd-c4.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 823 | "src/f32-spmm/gen/4x1-minmax-psimd.c", |
| 824 | "src/f32-spmm/gen/8x1-minmax-psimd.c", |
| 825 | "src/f32-spmm/gen/16x1-minmax-psimd.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 826 | "src/f32-pavgpool/9p8x-minmax-psimd-c4.c", |
| 827 | "src/f32-pavgpool/9x-minmax-psimd-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 828 | "src/f32-ppmm/gen/4x8-minmax-psimd.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 829 | "src/f32-prelu/gen/psimd-2x4.c", |
| 830 | "src/f32-prelu/gen/psimd-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 831 | "src/f32-rmax/psimd.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 832 | "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c", |
| 833 | "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c", |
| 834 | "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c", |
| 835 | "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c", |
| 836 | "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c", |
| 837 | "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c", |
| 838 | "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c", |
| 839 | "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 840 | "src/f32-vbinary/gen/vmax-psimd-x4.c", |
| 841 | "src/f32-vbinary/gen/vmax-psimd-x8.c", |
| 842 | "src/f32-vbinary/gen/vmaxc-psimd-x4.c", |
| 843 | "src/f32-vbinary/gen/vmaxc-psimd-x8.c", |
| 844 | "src/f32-vbinary/gen/vmin-psimd-x4.c", |
| 845 | "src/f32-vbinary/gen/vmin-psimd-x8.c", |
| 846 | "src/f32-vbinary/gen/vminc-psimd-x4.c", |
| 847 | "src/f32-vbinary/gen/vminc-psimd-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 848 | "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c", |
| 849 | "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c", |
| 850 | "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c", |
| 851 | "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c", |
| 852 | "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c", |
| 853 | "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c", |
| 854 | "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c", |
| 855 | "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 856 | "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c", |
| 857 | "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c", |
| 858 | "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c", |
| 859 | "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 860 | "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c", |
| 861 | "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c", |
| 862 | "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c", |
| 863 | "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 864 | "src/f32-vlrelu/gen/psimd-x4.c", |
| 865 | "src/f32-vlrelu/gen/psimd-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 866 | "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c", |
| 867 | "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 868 | "src/f32-vunary/gen/vabs-psimd-x4.c", |
| 869 | "src/f32-vunary/gen/vabs-psimd-x8.c", |
| 870 | "src/f32-vunary/gen/vneg-psimd-x4.c", |
| 871 | "src/f32-vunary/gen/vneg-psimd-x8.c", |
| 872 | "src/f32-vunary/gen/vsqr-psimd-x4.c", |
| 873 | "src/f32-vunary/gen/vsqr-psimd-x8.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 874 | "src/x32-fill/psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 875 | "src/x32-packx/x4-psimd.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 876 | "src/x32-pad/psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 877 | "src/x32-unpool/psimd.c", |
| 878 | "src/x32-zip/x2-psimd.c", |
| 879 | "src/x32-zip/x3-psimd.c", |
| 880 | "src/x32-zip/x4-psimd.c", |
| 881 | "src/x32-zip/xm-psimd.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 882 | "src/requantization/precise-psimd.c", |
| 883 | "src/requantization/fp32-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 884 | ] |
| 885 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 886 | PSIMD_ACCMATH_UKERNELS = [ |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 887 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c", |
| 888 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c", |
| 889 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c", |
| 890 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c", |
| 891 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c", |
| 892 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c", |
| 893 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c", |
| 894 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c", |
| 895 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c", |
| 896 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c", |
| 897 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c", |
| 898 | "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 899 | "src/f32-sigmoid/gen/psimd-p5-div-x4.c", |
| 900 | "src/f32-sigmoid/gen/psimd-p5-div-x8.c", |
| 901 | "src/f32-sigmoid/gen/psimd-p5-div-x12.c", |
| 902 | "src/f32-sigmoid/gen/psimd-p5-div-x16.c", |
| 903 | "src/f32-sigmoid/gen/psimd-p5-div-x20.c", |
| 904 | "src/f32-sigmoid/gen/psimd-p5-div-x24.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 905 | "src/f32-vrnd/gen/vrndne-psimd-x4.c", |
| 906 | "src/f32-vrnd/gen/vrndne-psimd-x8.c", |
| 907 | "src/f32-vrnd/gen/vrndz-psimd-x4.c", |
| 908 | "src/f32-vrnd/gen/vrndz-psimd-x8.c", |
| 909 | "src/f32-vrnd/gen/vrndu-psimd-x4.c", |
| 910 | "src/f32-vrnd/gen/vrndu-psimd-x8.c", |
| 911 | "src/f32-vrnd/gen/vrndd-psimd-x4.c", |
| 912 | "src/f32-vrnd/gen/vrndd-psimd-x8.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 913 | "src/math/roundne-psimd-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 914 | "src/math/roundd-psimd-addsub.c", |
| 915 | "src/math/roundu-psimd-addsub.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 916 | "src/math/roundz-psimd-addsub.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 917 | "src/math/sigmoid-psimd-p5-div.c", |
| 918 | ] |
| 919 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 920 | # ISA-specific micro-kernels |
| 921 | NEON_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 922 | "src/f32-avgpool/9p8x-minmax-neon-c4.c", |
| 923 | "src/f32-avgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 924 | "src/f32-clamp/gen/neon-x4.c", |
| 925 | "src/f32-clamp/gen/neon-x8.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 926 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c", |
| 927 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 928 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c", |
| 929 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 930 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c", |
| 931 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 932 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c", |
| 933 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 934 | "src/f32-dwconv/gen/up4x4-minmax-neon.c", |
| 935 | "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", |
| 936 | "src/f32-dwconv/gen/up8x4-minmax-neon.c", |
| 937 | "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 938 | "src/f32-dwconv/gen/up4x9-minmax-neon.c", |
| 939 | "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", |
| 940 | "src/f32-dwconv/gen/up8x9-minmax-neon.c", |
| 941 | "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 942 | "src/f32-dwconv/gen/up4x25-minmax-neon.c", |
| 943 | "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", |
| 944 | "src/f32-dwconv/gen/up8x25-minmax-neon.c", |
| 945 | "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 946 | "src/f32-gavgpool-cw/neon-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 947 | "src/f32-gavgpool/7p7x-minmax-neon-c4.c", |
| 948 | "src/f32-gavgpool/7x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 949 | "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 950 | "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 951 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 952 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 953 | "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c", |
| 954 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 955 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 956 | "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 957 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 958 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 959 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 960 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 961 | "src/f32-gemm/gen/1x8s4-minmax-neon.c", |
| 962 | "src/f32-gemm/gen/4x8s4-minmax-neon.c", |
| 963 | "src/f32-gemm/gen/6x8s4-minmax-neon.c", |
| 964 | "src/f32-gemm/gen/8x8s4-minmax-neon.c", |
| 965 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c", |
| 966 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c", |
| 967 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c", |
| 968 | "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c", |
| 969 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c", |
| 970 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c", |
| 971 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c", |
| 972 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c", |
| 973 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c", |
| 974 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c", |
| 975 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c", |
| 976 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c", |
| 977 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c", |
| 978 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c", |
| 979 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 980 | "src/f32-hswish/gen/neon-x4.c", |
| 981 | "src/f32-hswish/gen/neon-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 982 | "src/f32-ibilinear/gen/neon-c4.c", |
| 983 | "src/f32-ibilinear/gen/neon-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 984 | "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 985 | "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 986 | "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c", |
| 987 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 988 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 989 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 990 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 991 | "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 992 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 993 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 994 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 995 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 996 | "src/f32-igemm/gen/1x8s4-minmax-neon.c", |
| 997 | "src/f32-igemm/gen/4x8s4-minmax-neon.c", |
| 998 | "src/f32-igemm/gen/6x8s4-minmax-neon.c", |
| 999 | "src/f32-igemm/gen/8x8s4-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1000 | "src/f32-maxpool/9p8x-minmax-neon-c4.c", |
| 1001 | "src/f32-pavgpool/9p8x-minmax-neon-c4.c", |
| 1002 | "src/f32-pavgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1003 | "src/f32-ppmm/gen/4x8-minmax-neon.c", |
| 1004 | "src/f32-ppmm/gen/8x8-minmax-neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1005 | "src/f32-prelu/gen/neon-2x4.c", |
| 1006 | "src/f32-prelu/gen/neon-2x8.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1007 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 1008 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 1009 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 1010 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 1011 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 1012 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 1013 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 1014 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 1015 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 1016 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| 1017 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 1018 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 1019 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| 1020 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| 1021 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| 1022 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| 1023 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 1024 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| 1025 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| 1026 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 1027 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| 1028 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 1029 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 1030 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1031 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1032 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1033 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 1034 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 1035 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 1036 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 1037 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 1038 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| 1039 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 1040 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 1041 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 1042 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 1043 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 1044 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 1045 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 1046 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 1047 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 1048 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 1049 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 1050 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1051 | "src/f32-vbinary/gen/vadd-minmax-neon-x4.c", |
| 1052 | "src/f32-vbinary/gen/vadd-minmax-neon-x8.c", |
| 1053 | "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c", |
| 1054 | "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1055 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 1056 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 1057 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 1058 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 1059 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 1060 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 1061 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 1062 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1063 | "src/f32-vbinary/gen/vmul-minmax-neon-x4.c", |
| 1064 | "src/f32-vbinary/gen/vmul-minmax-neon-x8.c", |
| 1065 | "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c", |
| 1066 | "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c", |
| 1067 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c", |
| 1068 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1069 | "src/f32-vbinary/gen/vsqrdiff-neon-x4.c", |
| 1070 | "src/f32-vbinary/gen/vsqrdiff-neon-x8.c", |
| 1071 | "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c", |
| 1072 | "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1073 | "src/f32-vbinary/gen/vsub-minmax-neon-x4.c", |
| 1074 | "src/f32-vbinary/gen/vsub-minmax-neon-x8.c", |
| 1075 | "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c", |
| 1076 | "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 1077 | "src/f32-vlrelu/gen/neon-x4.c", |
| 1078 | "src/f32-vlrelu/gen/neon-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1079 | "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c", |
| 1080 | "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1081 | "src/f32-vrnd/gen/vrndne-neon-x4.c", |
| 1082 | "src/f32-vrnd/gen/vrndne-neon-x8.c", |
| 1083 | "src/f32-vrnd/gen/vrndz-neon-x4.c", |
| 1084 | "src/f32-vrnd/gen/vrndz-neon-x8.c", |
| 1085 | "src/f32-vrnd/gen/vrndu-neon-x4.c", |
| 1086 | "src/f32-vrnd/gen/vrndu-neon-x8.c", |
| 1087 | "src/f32-vrnd/gen/vrndd-neon-x4.c", |
| 1088 | "src/f32-vrnd/gen/vrndd-neon-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1089 | "src/f32-vunary/gen/vabs-neon-x4.c", |
| 1090 | "src/f32-vunary/gen/vabs-neon-x8.c", |
| 1091 | "src/f32-vunary/gen/vneg-neon-x4.c", |
| 1092 | "src/f32-vunary/gen/vneg-neon-x8.c", |
| 1093 | "src/f32-vunary/gen/vsqr-neon-x4.c", |
| 1094 | "src/f32-vunary/gen/vsqr-neon-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1095 | "src/q8-avgpool/9p8x-minmax-neon-c8.c", |
| 1096 | "src/q8-avgpool/9x-minmax-neon-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1097 | "src/q8-dwconv/up8x9-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1098 | "src/q8-gavgpool/7p7x-minmax-neon-c8.c", |
| 1099 | "src/q8-gavgpool/7x-minmax-neon-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1100 | "src/q8-gemm/4x8-minmax-neon.c", |
| 1101 | "src/q8-gemm/8x8-minmax-neon.c", |
| 1102 | "src/q8-igemm/4x8-minmax-neon.c", |
| 1103 | "src/q8-igemm/8x8-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1104 | "src/q8-vadd/minmax-neon.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1105 | "src/u8-clamp/neon-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1106 | "src/u8-maxpool/9p8x-minmax-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1107 | "src/u8-rmax/neon.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1108 | "src/x32-fill/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1109 | "src/x32-packx/x4-neon-st4.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1110 | "src/x32-pad/neon.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1111 | "src/x32-unpool/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1112 | "src/x32-zip/x2-neon.c", |
| 1113 | "src/x32-zip/x3-neon.c", |
| 1114 | "src/x32-zip/x4-neon.c", |
| 1115 | "src/x32-zip/xm-neon.c", |
| 1116 | "src/x8-zip/x2-neon.c", |
| 1117 | "src/x8-zip/x3-neon.c", |
| 1118 | "src/x8-zip/x4-neon.c", |
| 1119 | "src/x8-zip/xm-neon.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1120 | "src/math/roundne-neon-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1121 | "src/math/roundd-neon-addsub.c", |
| 1122 | "src/math/roundd-neon-cvt.c", |
| 1123 | "src/math/roundu-neon-addsub.c", |
| 1124 | "src/math/roundu-neon-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1125 | "src/math/roundz-neon-addsub.c", |
| 1126 | "src/math/roundz-neon-cvt.c", |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 1127 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1128 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 1129 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 1130 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 1131 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 1132 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 1133 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1134 | "src/math/sqrt-neon-nr1rsqrts.c", |
| 1135 | "src/math/sqrt-neon-nr2rsqrts.c", |
| 1136 | "src/math/sqrt-neon-nr3rsqrts.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1137 | "src/requantization/precise-neon.c", |
| 1138 | "src/requantization/fp32-neon.c", |
| 1139 | "src/requantization/q31-neon.c", |
| 1140 | "src/requantization/gemmlowp-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1141 | ] |
| 1142 | |
| 1143 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1144 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 1145 | "src/f32-ibilinear/gen/neonfma-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1146 | "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1147 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1148 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1149 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1150 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1151 | "src/f32-igemm/gen/1x8s4-minmax-neonfma.c", |
| 1152 | "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", |
| 1153 | "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", |
| 1154 | "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1155 | "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", |
| 1156 | "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", |
| 1157 | "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", |
| 1158 | "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1159 | "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", |
| 1160 | "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", |
| 1161 | "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", |
| 1162 | "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1163 | "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", |
| 1164 | "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", |
| 1165 | "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", |
| 1166 | "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1167 | "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1168 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1169 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1170 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1171 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1172 | "src/f32-gemm/gen/1x8s4-minmax-neonfma.c", |
| 1173 | "src/f32-gemm/gen/4x8s4-minmax-neonfma.c", |
| 1174 | "src/f32-gemm/gen/6x8s4-minmax-neonfma.c", |
| 1175 | "src/f32-gemm/gen/8x8s4-minmax-neonfma.c", |
| 1176 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c", |
| 1177 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c", |
| 1178 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c", |
| 1179 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c", |
| 1180 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c", |
| 1181 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c", |
| 1182 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c", |
| 1183 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c", |
| 1184 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1185 | "src/f32-hswish/gen/neonfma-x4.c", |
| 1186 | "src/f32-hswish/gen/neonfma-x8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1187 | "src/f32-ppmm/gen/4x8-minmax-neonfma.c", |
| 1188 | "src/f32-ppmm/gen/8x8-minmax-neonfma.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1189 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 1190 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 1191 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 1192 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 1193 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 1194 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 1195 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 1196 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 1197 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 1198 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| 1199 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 1200 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 1201 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| 1202 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| 1203 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| 1204 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| 1205 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 1206 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| 1207 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| 1208 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 1209 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| 1210 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 1211 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 1212 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1213 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 1214 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 1215 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 1216 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 1217 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 1218 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 1219 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 1220 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 1221 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 1222 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 1223 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 1224 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 1225 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 1226 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 1227 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 1228 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 1229 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 1230 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| 1231 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 1232 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 1233 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 1234 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 1235 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 1236 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| 1237 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 1238 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 1239 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 1240 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 1241 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 1242 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| 1243 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 1244 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 1245 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 1246 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 1247 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 1248 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| 1249 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 1250 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 1251 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 1252 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 1253 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 1254 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| 1255 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 1256 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 1257 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 1258 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 1259 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 1260 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| 1261 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 1262 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 1263 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 1264 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 1265 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 1266 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1267 | "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c", |
| 1268 | "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1269 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c", |
| 1270 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c", |
| 1271 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c", |
| 1272 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c", |
| 1273 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c", |
| 1274 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c", |
| 1275 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c", |
| 1276 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c", |
| 1277 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c", |
| 1278 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c", |
| 1279 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c", |
| 1280 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c", |
| 1281 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c", |
| 1282 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c", |
| 1283 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c", |
| 1284 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c", |
| 1285 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c", |
| 1286 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c", |
| 1287 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c", |
| 1288 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 1289 | "src/math/exp-neonfma-lut64-p2.c", |
| 1290 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 1291 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 1292 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1293 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1294 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 1295 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 1296 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| 1297 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 1298 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 1299 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| 1300 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 1301 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 1302 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| 1303 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 1304 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 1305 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| 1306 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 1307 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 1308 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| 1309 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 1310 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 1311 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1312 | "src/math/sqrt-neonfma-nr1fma.c", |
| 1313 | "src/math/sqrt-neonfma-nr2fma.c", |
| 1314 | "src/math/sqrt-neonfma-nr3fma.c", |
| 1315 | "src/math/sqrt-neonfma-nr2fma1adj.c", |
| 1316 | "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1317 | ] |
| 1318 | |
| 1319 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1320 | "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c", |
| 1321 | "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c", |
| 1322 | "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c", |
| 1323 | "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c", |
| 1324 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c", |
| 1325 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1326 | "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1327 | "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1328 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1329 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1330 | "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c", |
| 1331 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1332 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
| 1333 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c", |
| 1334 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c", |
| 1335 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c", |
| 1336 | "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c", |
| 1337 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c", |
| 1338 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c", |
| 1339 | "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1340 | "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1341 | "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c", |
| 1342 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1343 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1344 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1345 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1346 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c", |
| 1347 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1348 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c", |
| 1349 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1350 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c", |
| 1351 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1352 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c", |
| 1353 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1354 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c", |
| 1355 | "src/f32-dwconv-chw/3x3p1-neonfma.c", |
| 1356 | "src/f32-dwconv-chw/5x5p2-neonfma.c", |
| 1357 | "src/f32-dwconv-chw/3x3s2p1-neonfma.c", |
| 1358 | "src/f32-dwconv-chw/5x5s2p2-neonfma.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1359 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 1360 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 1361 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 1362 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 1363 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 1364 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| 1365 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 1366 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 1367 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 1368 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 1369 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 1370 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 1371 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 1372 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 1373 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 1374 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 1375 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 1376 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1377 | "src/f32-spmm/gen/12x1-minmax-neonfma.c", |
| 1378 | "src/f32-spmm/gen/12x2-minmax-neonfma.c", |
| 1379 | "src/f32-spmm/gen/12x4-minmax-neonfma.c", |
| 1380 | "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c", |
| 1381 | "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c", |
| 1382 | "src/f32-spmm/gen/16x1-minmax-neonfma.c", |
| 1383 | "src/f32-spmm/gen/16x2-minmax-neonfma.c", |
| 1384 | "src/f32-spmm/gen/16x4-minmax-neonfma.c", |
| 1385 | "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c", |
| 1386 | "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c", |
| 1387 | "src/f32-spmm/gen/4x1-minmax-neonfma.c", |
| 1388 | "src/f32-spmm/gen/4x2-minmax-neonfma.c", |
| 1389 | "src/f32-spmm/gen/4x4-minmax-neonfma.c", |
| 1390 | "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c", |
| 1391 | "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c", |
| 1392 | "src/f32-spmm/gen/8x1-minmax-neonfma.c", |
| 1393 | "src/f32-spmm/gen/8x2-minmax-neonfma.c", |
| 1394 | "src/f32-spmm/gen/8x4-minmax-neonfma.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1395 | "src/f32-vsqrt/gen/neon-sqrt-x4.c", |
| 1396 | "src/f32-vsqrt/gen/neon-sqrt-x8.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1397 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| 1398 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| 1399 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| 1400 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| 1401 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| 1402 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1403 | ] |
| 1404 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1405 | NEONV8_UKERNELS = [ |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1406 | "src/f32-vrnd/gen/vrndne-neonv8-x4.c", |
| 1407 | "src/f32-vrnd/gen/vrndne-neonv8-x8.c", |
| 1408 | "src/f32-vrnd/gen/vrndz-neonv8-x4.c", |
| 1409 | "src/f32-vrnd/gen/vrndz-neonv8-x8.c", |
| 1410 | "src/f32-vrnd/gen/vrndu-neonv8-x4.c", |
| 1411 | "src/f32-vrnd/gen/vrndu-neonv8-x8.c", |
| 1412 | "src/f32-vrnd/gen/vrndd-neonv8-x4.c", |
| 1413 | "src/f32-vrnd/gen/vrndd-neonv8-x8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1414 | "src/math/roundne-neonv8.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1415 | "src/math/roundd-neonv8.c", |
| 1416 | "src/math/roundu-neonv8.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1417 | "src/math/roundz-neonv8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1418 | ] |
| 1419 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1420 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1421 | "src/f16-clamp/gen/neonfp16arith-x16.c", |
| 1422 | "src/f16-clamp/gen/neonfp16arith-x8.c", |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 1423 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c", |
| 1424 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c", |
| 1425 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c", |
| 1426 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c", |
| 1427 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c", |
| 1428 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c", |
| 1429 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c", |
| 1430 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c", |
| 1431 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c", |
| 1432 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c", |
| 1433 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c", |
| 1434 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c", |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 1435 | "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c", |
| 1436 | "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c", |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1437 | "src/f16-hswish/gen/neonfp16arith-x16.c", |
| 1438 | "src/f16-hswish/gen/neonfp16arith-x8.c", |
| 1439 | "src/f16-prelu/gen/neonfp16arith-2x16.c", |
| 1440 | "src/f16-prelu/gen/neonfp16arith-2x8.c", |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 1441 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c", |
| 1442 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c", |
| 1443 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c", |
| 1444 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c", |
| 1445 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c", |
| 1446 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c", |
| 1447 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c", |
| 1448 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c", |
| 1449 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c", |
| 1450 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c", |
| 1451 | "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c", |
| 1452 | "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c", |
| 1453 | "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c", |
| 1454 | "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c", |
| 1455 | "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c", |
| 1456 | "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c", |
| 1457 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c", |
| 1458 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c", |
| 1459 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c", |
| 1460 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c", |
| 1461 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c", |
| 1462 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c", |
| 1463 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c", |
| 1464 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c", |
| 1465 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c", |
| 1466 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c", |
| 1467 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c", |
| 1468 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c", |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 1469 | "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c", |
| 1470 | "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c", |
Frank Barchard | 83ea939 | 2020-04-20 20:15:14 -0700 | [diff] [blame] | 1471 | "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1472 | "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1473 | "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1474 | "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1475 | "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c", |
| 1476 | "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c", |
| 1477 | "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c", |
| 1478 | "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c", |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 1479 | "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1480 | "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1481 | "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1482 | "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
Frank Barchard | 3f9f99f | 2020-05-06 01:12:04 -0700 | [diff] [blame] | 1483 | "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1484 | "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1485 | "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1486 | "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
| 1487 | "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c", |
| 1488 | "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c", |
| 1489 | "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c", |
| 1490 | "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c", |
| 1491 | "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1492 | "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1493 | "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1494 | "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1495 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c", |
| 1496 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c", |
| 1497 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c", |
| 1498 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c", |
| 1499 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c", |
| 1500 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c", |
| 1501 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c", |
| 1502 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1503 | ] |
| 1504 | |
| 1505 | SSE_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1506 | "src/f32-avgpool/9p8x-minmax-sse-c4.c", |
| 1507 | "src/f32-avgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1508 | "src/f32-clamp/gen/sse-x4.c", |
| 1509 | "src/f32-clamp/gen/sse-x8.c", |
Erich Elsen | b123340 | 2020-06-08 15:53:15 -0700 | [diff] [blame] | 1510 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c", |
| 1511 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1512 | "src/f32-dwconv-chw/3x3p1-sse.c", |
| 1513 | "src/f32-dwconv-chw/3x3s2p1-sse.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1514 | "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c", |
| 1515 | "src/f32-dwconv/gen/up4x25-minmax-sse.c", |
| 1516 | "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c", |
| 1517 | "src/f32-dwconv/gen/up4x4-minmax-sse.c", |
| 1518 | "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c", |
| 1519 | "src/f32-dwconv/gen/up4x9-minmax-sse.c", |
| 1520 | "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c", |
| 1521 | "src/f32-dwconv/gen/up8x25-minmax-sse.c", |
| 1522 | "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c", |
| 1523 | "src/f32-dwconv/gen/up8x4-minmax-sse.c", |
| 1524 | "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c", |
| 1525 | "src/f32-dwconv/gen/up8x9-minmax-sse.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1526 | "src/f32-gavgpool-cw/sse-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1527 | "src/f32-gavgpool/7p7x-minmax-sse-c4.c", |
| 1528 | "src/f32-gavgpool/7x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1529 | "src/f32-gemm/gen/1x8-minmax-sse-dup.c", |
| 1530 | "src/f32-gemm/gen/1x8-minmax-sse-load1.c", |
| 1531 | "src/f32-gemm/gen/1x8s4-minmax-sse.c", |
| 1532 | "src/f32-gemm/gen/4x2c4-minmax-sse.c", |
| 1533 | "src/f32-gemm/gen/4x8-minmax-sse-dup.c", |
| 1534 | "src/f32-gemm/gen/4x8-minmax-sse-load1.c", |
| 1535 | "src/f32-gemm/gen/4x8s4-minmax-sse.c", |
| 1536 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c", |
| 1537 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c", |
| 1538 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c", |
| 1539 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c", |
| 1540 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c", |
| 1541 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1542 | "src/f32-hswish/gen/sse-x4.c", |
| 1543 | "src/f32-hswish/gen/sse-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1544 | "src/f32-ibilinear/gen/sse-c4.c", |
| 1545 | "src/f32-ibilinear/gen/sse-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1546 | "src/f32-igemm/gen/1x8-minmax-sse-dup.c", |
| 1547 | "src/f32-igemm/gen/1x8-minmax-sse-load1.c", |
| 1548 | "src/f32-igemm/gen/1x8s4-minmax-sse.c", |
| 1549 | "src/f32-igemm/gen/4x2c4-minmax-sse.c", |
| 1550 | "src/f32-igemm/gen/4x8-minmax-sse-dup.c", |
| 1551 | "src/f32-igemm/gen/4x8-minmax-sse-load1.c", |
| 1552 | "src/f32-igemm/gen/4x8s4-minmax-sse.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1553 | "src/f32-maxpool/9p8x-minmax-sse-c4.c", |
| 1554 | "src/f32-pavgpool/9p8x-minmax-sse-c4.c", |
| 1555 | "src/f32-pavgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1556 | "src/f32-ppmm/gen/4x8-minmax-sse.c", |
Marat Dukhan | 39b5e94 | 2020-06-24 15:03:48 -0700 | [diff] [blame] | 1557 | "src/f32-prelu/gen/sse-2x4.c", |
| 1558 | "src/f32-prelu/gen/sse-2x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1559 | "src/f32-rmax/sse.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1560 | "src/f32-spmm/gen/4x1-minmax-sse.c", |
| 1561 | "src/f32-spmm/gen/8x1-minmax-sse.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 1562 | "src/f32-spmm/gen/16x1-minmax-sse.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1563 | "src/f32-vbinary/gen/vadd-minmax-sse-x4.c", |
| 1564 | "src/f32-vbinary/gen/vadd-minmax-sse-x8.c", |
| 1565 | "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c", |
| 1566 | "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c", |
| 1567 | "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c", |
| 1568 | "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c", |
| 1569 | "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c", |
| 1570 | "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1571 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 1572 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 1573 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 1574 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 1575 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 1576 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 1577 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 1578 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1579 | "src/f32-vbinary/gen/vmul-minmax-sse-x4.c", |
| 1580 | "src/f32-vbinary/gen/vmul-minmax-sse-x8.c", |
| 1581 | "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c", |
| 1582 | "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c", |
| 1583 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c", |
| 1584 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c", |
| 1585 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c", |
| 1586 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1587 | "src/f32-vbinary/gen/vsqrdiff-sse-x4.c", |
| 1588 | "src/f32-vbinary/gen/vsqrdiff-sse-x8.c", |
| 1589 | "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c", |
| 1590 | "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1591 | "src/f32-vbinary/gen/vsub-minmax-sse-x4.c", |
| 1592 | "src/f32-vbinary/gen/vsub-minmax-sse-x8.c", |
| 1593 | "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c", |
| 1594 | "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 1595 | "src/f32-vlrelu/gen/sse-x4.c", |
| 1596 | "src/f32-vlrelu/gen/sse-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1597 | "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c", |
| 1598 | "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1599 | "src/f32-vsqrt/gen/sse-sqrt-x4.c", |
| 1600 | "src/f32-vsqrt/gen/sse-sqrt-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1601 | "src/f32-vunary/gen/vabs-sse-x4.c", |
| 1602 | "src/f32-vunary/gen/vabs-sse-x8.c", |
| 1603 | "src/f32-vunary/gen/vneg-sse-x4.c", |
| 1604 | "src/f32-vunary/gen/vneg-sse-x8.c", |
| 1605 | "src/f32-vunary/gen/vsqr-sse-x4.c", |
| 1606 | "src/f32-vunary/gen/vsqr-sse-x8.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1607 | "src/x32-fill/sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1608 | "src/x32-packx/x4-sse.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1609 | "src/x32-pad/sse.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1610 | "src/math/roundne-sse-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1611 | "src/math/roundd-sse-addsub.c", |
| 1612 | "src/math/roundu-sse-addsub.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1613 | "src/math/roundz-sse-addsub.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1614 | "src/math/sqrt-sse-hh1mac.c", |
| 1615 | "src/math/sqrt-sse-nr1mac.c", |
| 1616 | "src/math/sqrt-sse-nr2mac.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1617 | ] |
| 1618 | |
| 1619 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1620 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 1621 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 1622 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1623 | "src/f32-prelu/gen/sse2-2x4.c", |
| 1624 | "src/f32-prelu/gen/sse2-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1625 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| 1626 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| 1627 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| 1628 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| 1629 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 1630 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| 1631 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| 1632 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 1633 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| 1634 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 1635 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 1636 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1637 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1638 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1639 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1640 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1641 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 1642 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Marat Dukhan | 0d3f467 | 2020-06-25 16:42:58 -0700 | [diff] [blame] | 1643 | "src/f32-vlrelu/gen/sse2-x4.c", |
| 1644 | "src/f32-vlrelu/gen/sse2-x8.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1645 | "src/f32-vrnd/gen/vrndne-sse2-x4.c", |
| 1646 | "src/f32-vrnd/gen/vrndne-sse2-x8.c", |
| 1647 | "src/f32-vrnd/gen/vrndz-sse2-x4.c", |
| 1648 | "src/f32-vrnd/gen/vrndz-sse2-x8.c", |
| 1649 | "src/f32-vrnd/gen/vrndu-sse2-x4.c", |
| 1650 | "src/f32-vrnd/gen/vrndu-sse2-x8.c", |
| 1651 | "src/f32-vrnd/gen/vrndd-sse2-x4.c", |
| 1652 | "src/f32-vrnd/gen/vrndd-sse2-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1653 | "src/q8-avgpool/9p8x-minmax-sse2-c8.c", |
| 1654 | "src/q8-avgpool/9x-minmax-sse2-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1655 | "src/q8-igemm/4x4c2-minmax-sse2.c", |
| 1656 | "src/q8-dwconv/up8x9-minmax-sse2.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1657 | "src/q8-gavgpool/7p7x-minmax-sse2-c8.c", |
| 1658 | "src/q8-gavgpool/7x-minmax-sse2-c8.c", |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1659 | "src/q8-gemm/2x4c8-minmax-sse2.c", |
| 1660 | "src/q8-gemm/4x4c2-minmax-sse2.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1661 | "src/q8-vadd/minmax-sse2.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1662 | "src/u8-clamp/sse2-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1663 | "src/u8-maxpool/9p8x-minmax-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1664 | "src/u8-rmax/sse2.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1665 | "src/x32-unpool/sse2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1666 | "src/x32-zip/x2-sse2.c", |
| 1667 | "src/x32-zip/x3-sse2.c", |
| 1668 | "src/x32-zip/x4-sse2.c", |
| 1669 | "src/x32-zip/xm-sse2.c", |
| 1670 | "src/x8-zip/x2-sse2.c", |
| 1671 | "src/x8-zip/x3-sse2.c", |
| 1672 | "src/x8-zip/x4-sse2.c", |
| 1673 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 1674 | "src/math/exp-sse2-p5.c", |
| 1675 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1676 | "src/math/roundne-sse2-cvt.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1677 | "src/math/roundd-sse2-cvt.c", |
| 1678 | "src/math/roundu-sse2-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1679 | "src/math/roundz-sse2-cvt.c", |
Marat Dukhan | 80bafd2 | 2019-11-18 10:16:01 -0800 | [diff] [blame] | 1680 | "src/math/sigmoid-sse2-p5-div.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1681 | "src/requantization/precise-sse2.c", |
| 1682 | "src/requantization/fp32-sse2.c", |
| 1683 | "src/requantization/q31-sse2.c", |
| 1684 | "src/requantization/gemmlowp-sse2.c", |
| 1685 | ] |
| 1686 | |
| 1687 | SSSE3_UKERNELS = [ |
| 1688 | "src/requantization/precise-ssse3.c", |
| 1689 | "src/requantization/q31-ssse3.c", |
| 1690 | "src/requantization/gemmlowp-ssse3.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1691 | ] |
| 1692 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1693 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1694 | "src/f32-prelu/gen/sse41-2x4.c", |
| 1695 | "src/f32-prelu/gen/sse41-2x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1696 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1697 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1698 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1699 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1700 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 1701 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Marat Dukhan | 0d3f467 | 2020-06-25 16:42:58 -0700 | [diff] [blame] | 1702 | "src/f32-vlrelu/gen/sse41-x4.c", |
| 1703 | "src/f32-vlrelu/gen/sse41-x8.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1704 | "src/f32-vrnd/gen/vrndne-sse41-x4.c", |
| 1705 | "src/f32-vrnd/gen/vrndne-sse41-x8.c", |
| 1706 | "src/f32-vrnd/gen/vrndz-sse41-x4.c", |
| 1707 | "src/f32-vrnd/gen/vrndz-sse41-x8.c", |
| 1708 | "src/f32-vrnd/gen/vrndu-sse41-x4.c", |
| 1709 | "src/f32-vrnd/gen/vrndu-sse41-x8.c", |
| 1710 | "src/f32-vrnd/gen/vrndd-sse41-x4.c", |
| 1711 | "src/f32-vrnd/gen/vrndd-sse41-x8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1712 | "src/math/roundne-sse41.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1713 | "src/math/roundd-sse41.c", |
| 1714 | "src/math/roundu-sse41.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1715 | "src/math/roundz-sse41.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 1716 | "src/requantization/precise-sse4.c", |
| 1717 | "src/requantization/q31-sse4.c", |
| 1718 | "src/requantization/gemmlowp-sse4.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1719 | ] |
| 1720 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1721 | AVX_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1722 | "src/f32-clamp/gen/avx-x8.c", |
| 1723 | "src/f32-clamp/gen/avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1724 | "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c", |
| 1725 | "src/f32-dwconv/gen/up16x4-minmax-avx.c", |
| 1726 | "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c", |
| 1727 | "src/f32-dwconv/gen/up8x4-minmax-avx.c", |
| 1728 | "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c", |
| 1729 | "src/f32-dwconv/gen/up16x9-minmax-avx.c", |
| 1730 | "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c", |
| 1731 | "src/f32-dwconv/gen/up8x9-minmax-avx.c", |
| 1732 | "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c", |
| 1733 | "src/f32-dwconv/gen/up16x25-minmax-avx.c", |
| 1734 | "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c", |
| 1735 | "src/f32-dwconv/gen/up8x25-minmax-avx.c", |
| 1736 | "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c", |
| 1737 | "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c", |
| 1738 | "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c", |
| 1739 | "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c", |
| 1740 | "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c", |
| 1741 | "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c", |
| 1742 | "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c", |
| 1743 | "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c", |
| 1744 | "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c", |
| 1745 | "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c", |
| 1746 | "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c", |
| 1747 | "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c", |
| 1748 | "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c", |
| 1749 | "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c", |
| 1750 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c", |
| 1751 | "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c", |
| 1752 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c", |
| 1753 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1754 | "src/f32-hswish/gen/avx-x8.c", |
| 1755 | "src/f32-hswish/gen/avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1756 | "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c", |
| 1757 | "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c", |
| 1758 | "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c", |
| 1759 | "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c", |
| 1760 | "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c", |
| 1761 | "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c", |
| 1762 | "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c", |
| 1763 | "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c", |
| 1764 | "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 1765 | "src/f32-prelu/gen/avx-2x8.c", |
| 1766 | "src/f32-prelu/gen/avx-2x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1767 | "src/f32-rmax/avx.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1768 | "src/f32-vbinary/gen/vadd-minmax-avx-x8.c", |
| 1769 | "src/f32-vbinary/gen/vadd-minmax-avx-x16.c", |
| 1770 | "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c", |
| 1771 | "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c", |
| 1772 | "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c", |
| 1773 | "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c", |
| 1774 | "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c", |
| 1775 | "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1776 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 1777 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 1778 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 1779 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 1780 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 1781 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 1782 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 1783 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1784 | "src/f32-vbinary/gen/vmul-minmax-avx-x8.c", |
| 1785 | "src/f32-vbinary/gen/vmul-minmax-avx-x16.c", |
| 1786 | "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c", |
| 1787 | "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c", |
| 1788 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c", |
| 1789 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c", |
| 1790 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c", |
| 1791 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1792 | "src/f32-vbinary/gen/vsqrdiff-avx-x8.c", |
| 1793 | "src/f32-vbinary/gen/vsqrdiff-avx-x16.c", |
| 1794 | "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c", |
| 1795 | "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1796 | "src/f32-vbinary/gen/vsub-minmax-avx-x8.c", |
| 1797 | "src/f32-vbinary/gen/vsub-minmax-avx-x16.c", |
| 1798 | "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c", |
| 1799 | "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 1800 | "src/f32-vlrelu/gen/avx-x8.c", |
| 1801 | "src/f32-vlrelu/gen/avx-x16.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 1802 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1803 | "src/f32-vrnd/gen/vrndne-avx-x8.c", |
| 1804 | "src/f32-vrnd/gen/vrndne-avx-x16.c", |
| 1805 | "src/f32-vrnd/gen/vrndz-avx-x8.c", |
| 1806 | "src/f32-vrnd/gen/vrndz-avx-x16.c", |
| 1807 | "src/f32-vrnd/gen/vrndu-avx-x8.c", |
| 1808 | "src/f32-vrnd/gen/vrndu-avx-x16.c", |
| 1809 | "src/f32-vrnd/gen/vrndd-avx-x8.c", |
| 1810 | "src/f32-vrnd/gen/vrndd-avx-x16.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1811 | "src/f32-vsqrt/gen/avx-sqrt-x8.c", |
| 1812 | "src/f32-vsqrt/gen/avx-sqrt-x16.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1813 | "src/f32-vunary/gen/vabs-avx-x8.c", |
| 1814 | "src/f32-vunary/gen/vabs-avx-x16.c", |
| 1815 | "src/f32-vunary/gen/vneg-avx-x8.c", |
| 1816 | "src/f32-vunary/gen/vneg-avx-x16.c", |
| 1817 | "src/f32-vunary/gen/vsqr-avx-x8.c", |
| 1818 | "src/f32-vunary/gen/vsqr-avx-x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1819 | ] |
| 1820 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1821 | FMA3_UKERNELS = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1822 | "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c", |
| 1823 | "src/f32-dwconv/gen/up16x4-minmax-fma3.c", |
| 1824 | "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c", |
| 1825 | "src/f32-dwconv/gen/up8x4-minmax-fma3.c", |
| 1826 | "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c", |
| 1827 | "src/f32-dwconv/gen/up16x9-minmax-fma3.c", |
| 1828 | "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c", |
| 1829 | "src/f32-dwconv/gen/up8x9-minmax-fma3.c", |
| 1830 | "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c", |
| 1831 | "src/f32-dwconv/gen/up16x25-minmax-fma3.c", |
| 1832 | "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c", |
| 1833 | "src/f32-dwconv/gen/up8x25-minmax-fma3.c", |
| 1834 | "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c", |
| 1835 | "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c", |
| 1836 | "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c", |
| 1837 | "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c", |
| 1838 | "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c", |
| 1839 | "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c", |
| 1840 | "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c", |
| 1841 | "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c", |
| 1842 | "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c", |
| 1843 | "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c", |
| 1844 | "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 1845 | "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 1846 | "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 1847 | "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| 1848 | "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c", |
| 1849 | "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c", |
| 1850 | "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c", |
| 1851 | "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c", |
| 1852 | "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c", |
| 1853 | "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c", |
| 1854 | "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c", |
| 1855 | "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c", |
| 1856 | "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c", |
| 1857 | "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c", |
| 1858 | "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c", |
| 1859 | "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c", |
| 1860 | "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c", |
| 1861 | "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1862 | "src/f32-hswish/gen/fma3-x8.c", |
| 1863 | "src/f32-hswish/gen/fma3-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1864 | "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c", |
| 1865 | "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c", |
| 1866 | "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c", |
| 1867 | "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c", |
| 1868 | "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c", |
| 1869 | "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c", |
| 1870 | "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c", |
| 1871 | "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c", |
| 1872 | "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c", |
| 1873 | "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c", |
| 1874 | "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 1875 | "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 1876 | "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 1877 | "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1878 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c", |
| 1879 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c", |
| 1880 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c", |
| 1881 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c", |
| 1882 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c", |
| 1883 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c", |
| 1884 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c", |
| 1885 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1886 | "src/math/sqrt-fma3-nr1fma.c", |
| 1887 | "src/math/sqrt-fma3-nr2fma.c", |
| 1888 | "src/math/sqrt-fma3-nr1fma1adj.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 1889 | ] |
| 1890 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1891 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1892 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 1893 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1894 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1895 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 1896 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1897 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 1898 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1899 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1900 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 1901 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1902 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1903 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 1904 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 1905 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 1906 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 1907 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 1908 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 1909 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 1910 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 1911 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 1912 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 1913 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 1914 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 1915 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 1916 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 1917 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1918 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 1919 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 1920 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 1921 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 1922 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1923 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 1924 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 1925 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1926 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1927 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1928 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 1929 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 1930 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 1931 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 1932 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 1933 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 1934 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 1935 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 1936 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 1937 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 1938 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 1939 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 1940 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 1941 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 1942 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 1943 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 1944 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 1945 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 1946 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 1947 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 1948 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 1949 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 1950 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 1951 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 1952 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 1953 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 1954 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 1955 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 1956 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 1957 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 1958 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 1959 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 1960 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 1961 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 1962 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 1963 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 1964 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 1965 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 1966 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 1967 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 1968 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 1969 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 1970 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 1971 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 1972 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 1973 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 1974 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 1975 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 1976 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 1977 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 1978 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 1979 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 1980 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 1981 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1982 | "src/math/exp-avx2-p5.c", |
| 1983 | "src/math/exp-avx2-perm-p3.c", |
| 1984 | "src/math/exp-avx2-perm-p4.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 1985 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 1986 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 1987 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 1988 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 1989 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| 1990 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 1991 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 1992 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1993 | ] |
| 1994 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1995 | AVX512F_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1996 | "src/f32-clamp/gen/avx512f-x16.c", |
| 1997 | "src/f32-clamp/gen/avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1998 | "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c", |
| 1999 | "src/f32-dwconv/gen/up32x4-minmax-avx512f.c", |
| 2000 | "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c", |
| 2001 | "src/f32-dwconv/gen/up16x4-minmax-avx512f.c", |
| 2002 | "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c", |
| 2003 | "src/f32-dwconv/gen/up32x9-minmax-avx512f.c", |
| 2004 | "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c", |
| 2005 | "src/f32-dwconv/gen/up16x9-minmax-avx512f.c", |
| 2006 | "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c", |
| 2007 | "src/f32-dwconv/gen/up32x25-minmax-avx512f.c", |
| 2008 | "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c", |
| 2009 | "src/f32-dwconv/gen/up16x25-minmax-avx512f.c", |
| 2010 | "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2011 | "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2012 | "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2013 | "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2014 | "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2015 | "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c", |
| 2016 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c", |
| 2017 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c", |
| 2018 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c", |
| 2019 | "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c", |
| 2020 | "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c", |
| 2021 | "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c", |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 2022 | "src/f32-hswish/gen/avx512f-x16.c", |
| 2023 | "src/f32-hswish/gen/avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2024 | "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2025 | "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2026 | "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2027 | "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2028 | "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2029 | "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2030 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 2031 | "src/f32-prelu/gen/avx512f-2x32.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2032 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2033 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2034 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 2035 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2036 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 2037 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2038 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2039 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 2040 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2041 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2042 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2043 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 2044 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 2045 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2046 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 2047 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 2048 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 2049 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 2050 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2051 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 2052 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 2053 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2054 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2055 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 2056 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2057 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2058 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 2059 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2060 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 2061 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2062 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2063 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 2064 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2065 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2066 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2067 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2068 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2069 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c", |
| 2070 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c", |
| 2071 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c", |
| 2072 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c", |
| 2073 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c", |
| 2074 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c", |
| 2075 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c", |
| 2076 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2077 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 2078 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 2079 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 2080 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 2081 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 2082 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 2083 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 2084 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2085 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c", |
| 2086 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c", |
| 2087 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c", |
| 2088 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c", |
| 2089 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c", |
| 2090 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c", |
| 2091 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c", |
| 2092 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2093 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c", |
| 2094 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c", |
| 2095 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c", |
| 2096 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2097 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c", |
| 2098 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c", |
| 2099 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c", |
| 2100 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 2101 | "src/f32-vlrelu/gen/avx512f-x16.c", |
| 2102 | "src/f32-vlrelu/gen/avx512f-x32.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2103 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2104 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 2105 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 2106 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 2107 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 2108 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 2109 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 2110 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 2111 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2112 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2113 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2114 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 2115 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2116 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 2117 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 2118 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 2119 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 2120 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 2121 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 2122 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 2123 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 2124 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 2125 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 2126 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 2127 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2128 | "src/f32-vrnd/gen/vrndne-avx512f-x16.c", |
| 2129 | "src/f32-vrnd/gen/vrndne-avx512f-x32.c", |
| 2130 | "src/f32-vrnd/gen/vrndz-avx512f-x16.c", |
| 2131 | "src/f32-vrnd/gen/vrndz-avx512f-x32.c", |
| 2132 | "src/f32-vrnd/gen/vrndu-avx512f-x16.c", |
| 2133 | "src/f32-vrnd/gen/vrndu-avx512f-x32.c", |
| 2134 | "src/f32-vrnd/gen/vrndd-avx512f-x16.c", |
| 2135 | "src/f32-vrnd/gen/vrndd-avx512f-x32.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2136 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c", |
| 2137 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c", |
| 2138 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c", |
| 2139 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c", |
| 2140 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c", |
| 2141 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c", |
| 2142 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c", |
| 2143 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2144 | "src/f32-vunary/gen/vabs-avx512f-x16.c", |
| 2145 | "src/f32-vunary/gen/vabs-avx512f-x32.c", |
| 2146 | "src/f32-vunary/gen/vneg-avx512f-x16.c", |
| 2147 | "src/f32-vunary/gen/vneg-avx512f-x32.c", |
| 2148 | "src/f32-vunary/gen/vsqr-avx512f-x16.c", |
| 2149 | "src/f32-vunary/gen/vsqr-avx512f-x32.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2150 | "src/math/exp-avx512f-p5-scalef.c", |
| 2151 | "src/math/exp-avx512f-p5.c", |
| 2152 | "src/math/exp-avx512f-perm-p3.c", |
Marat Dukhan | feb4923 | 2019-10-28 11:03:31 -0700 | [diff] [blame] | 2153 | "src/math/exp-avx512f-perm2-p2.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2154 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2155 | "src/math/sqrt-avx512f-nr1fma.c", |
| 2156 | "src/math/sqrt-avx512f-nr2fma.c", |
| 2157 | "src/math/sqrt-avx512f-nr1fma1adj.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2158 | ] |
| 2159 | |
| 2160 | AARCH32_ASM_UKERNELS = [ |
Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 2161 | "src/f32-gemm/4x4-aarch32-vfp-ld64.S", |
Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 2162 | "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2163 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2164 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
| 2165 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 2166 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 2167 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| 2168 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S", |
| 2169 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| 2170 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2171 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 2172 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
| 2173 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2174 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2175 | ] |
| 2176 | |
| 2177 | AARCH64_ASM_UKERNELS = [ |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 2178 | "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 2179 | "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2180 | "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 2181 | "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 2182 | "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
| 2183 | "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2184 | "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2185 | "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2186 | "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 2187 | "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2188 | "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| 2189 | "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| 2190 | "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 2191 | "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2192 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S", |
| 2193 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S", |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 2194 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2195 | "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2196 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2197 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2198 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2199 | "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2200 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2201 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2202 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2203 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2204 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S", |
| 2205 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S", |
| 2206 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2207 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2208 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2209 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2210 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2211 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2212 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2213 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S", |
| 2214 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S", |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 2215 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2216 | "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2217 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2218 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2219 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2220 | "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2221 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2222 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2223 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2224 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2225 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2226 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2227 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2228 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2229 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2230 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2231 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S", |
| 2232 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2233 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2234 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2235 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2236 | "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2237 | "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2238 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2239 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2240 | "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2241 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2242 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2243 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2244 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2245 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2246 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2247 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2248 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2249 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2250 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2251 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2252 | ] |
| 2253 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2254 | INTERNAL_MICROKERNEL_HDRS = [ |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2255 | "src/requantization/gemmlowp-requantization.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2256 | "src/xnnpack/argmaxpool.h", |
| 2257 | "src/xnnpack/avgpool.h", |
| 2258 | "src/xnnpack/clamp.h", |
| 2259 | "src/xnnpack/common.h", |
| 2260 | "src/xnnpack/conv.h", |
| 2261 | "src/xnnpack/dwconv.h", |
| 2262 | "src/xnnpack/gavgpool.h", |
| 2263 | "src/xnnpack/gemm.h", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 2264 | "src/xnnpack/fill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2265 | "src/xnnpack/hswish.h", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2266 | "src/xnnpack/ibilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2267 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 2268 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2269 | "src/xnnpack/lut.h", |
| 2270 | "src/xnnpack/math.h", |
| 2271 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2272 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2273 | "src/xnnpack/packx.h", |
| 2274 | "src/xnnpack/pad.h", |
| 2275 | "src/xnnpack/params.h", |
| 2276 | "src/xnnpack/pavgpool.h", |
| 2277 | "src/xnnpack/ppmm.h", |
| 2278 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2279 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2280 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2281 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2282 | "src/xnnpack/rmax.h", |
| 2283 | "src/xnnpack/scalar-utils.h", |
| 2284 | "src/xnnpack/spmm.h", |
| 2285 | "src/xnnpack/unpool.h", |
| 2286 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2287 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2288 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2289 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2290 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2291 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2292 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2293 | "src/xnnpack/zip.h", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2294 | ] |
| 2295 | |
| 2296 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2297 | "include/xnnpack.h", |
| 2298 | "src/xnnpack/allocator.h", |
| 2299 | "src/xnnpack/compute.h", |
| 2300 | "src/xnnpack/im2col.h", |
| 2301 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2302 | "src/xnnpack/math-stubs.h", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 2303 | "src/xnnpack/memory-planner.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2304 | "src/xnnpack/operator.h", |
| 2305 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2306 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2307 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2308 | "src/xnnpack/requantization.h", |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 2309 | "src/xnnpack/subgraph.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2310 | ] |
| 2311 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2312 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2313 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2314 | ] |
| 2315 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2316 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2317 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2318 | "include/xnnpack.h", |
| 2319 | ] |
| 2320 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2321 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2322 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2323 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2324 | "src/xnnpack/requantization.h", |
| 2325 | "include/xnnpack.h", |
| 2326 | ] |
| 2327 | |
| 2328 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 2329 | "src/xnnpack/params.h", |
| 2330 | "src/xnnpack/common.h", |
| 2331 | ] |
| 2332 | |
| 2333 | WEIGHTS_PACK_HDRS = [ |
| 2334 | "src/xnnpack/pack.h", |
| 2335 | "src/xnnpack/operator.h", |
| 2336 | "src/xnnpack/compute.h", |
| 2337 | ] |
| 2338 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2339 | LOGGING_COPTS = select({ |
| 2340 | # No logging in optimized mode |
| 2341 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 2342 | # Full logging in debug mode |
| 2343 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 2344 | # Error-only logging in default (fastbuild) mode |
| 2345 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 2346 | }) |
| 2347 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2348 | LOGGING_SRCS = select({ |
| 2349 | # No logging in optimized mode |
| 2350 | ":optimized_build": [], |
| 2351 | "//conditions:default": [ |
| 2352 | "src/operator-strings.c", |
| 2353 | "src/subgraph-strings.c", |
| 2354 | ], |
| 2355 | }) |
| 2356 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2357 | LOGGING_HDRS = [ |
| 2358 | "src/xnnpack/log.h", |
| 2359 | ] |
| 2360 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2361 | xnnpack_cc_library( |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2362 | name = "tables", |
| 2363 | srcs = TABLE_SRCS, |
| 2364 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2365 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2366 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2367 | ) |
| 2368 | |
| 2369 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2370 | name = "scalar_ukernels", |
| 2371 | srcs = SCALAR_UKERNELS, |
| 2372 | hdrs = INTERNAL_HDRS, |
| 2373 | aarch32_copts = ["-marm"], |
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 | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2376 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2377 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2378 | "@FP16", |
| 2379 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2380 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2381 | ], |
| 2382 | ) |
| 2383 | |
| 2384 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2385 | name = "scalar_ukernels_test_mode", |
| 2386 | srcs = SCALAR_UKERNELS, |
| 2387 | hdrs = INTERNAL_HDRS, |
| 2388 | aarch32_copts = ["-marm"], |
| 2389 | copts = [ |
| 2390 | "-UNDEBUG", |
| 2391 | "-DXNN_TEST_MODE=1", |
| 2392 | ], |
| 2393 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2394 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2395 | deps = [ |
| 2396 | ":tables", |
| 2397 | "@FP16", |
| 2398 | "@FXdiv", |
| 2399 | "@pthreadpool", |
| 2400 | ], |
| 2401 | ) |
| 2402 | |
| 2403 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2404 | name = "wasm_ukernels", |
| 2405 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2406 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2407 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2408 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 1c6e389 | 2020-06-25 23:56:10 -0700 | [diff] [blame] | 2409 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2410 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2411 | ":tables", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2412 | "@FP16", |
| 2413 | "@FXdiv", |
| 2414 | "@pthreadpool", |
| 2415 | ], |
| 2416 | ) |
| 2417 | |
| 2418 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2419 | name = "wasm_ukernels_test_mode", |
| 2420 | hdrs = INTERNAL_HDRS, |
| 2421 | copts = [ |
| 2422 | "-UNDEBUG", |
| 2423 | "-DXNN_TEST_MODE=1", |
| 2424 | ], |
| 2425 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2426 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2427 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 2428 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2429 | deps = [ |
| 2430 | ":tables", |
| 2431 | "@FP16", |
| 2432 | "@FXdiv", |
| 2433 | "@pthreadpool", |
| 2434 | ], |
| 2435 | ) |
| 2436 | |
| 2437 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2438 | name = "psimd_fastmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2439 | hdrs = INTERNAL_HDRS, |
| 2440 | aarch32_copts = [ |
| 2441 | "-marm", |
| 2442 | "-mfpu=neon", |
| 2443 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2444 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2445 | gcc_x86_copts = ["-msse2"], |
| 2446 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2447 | optimized_copts = [ |
| 2448 | "-O3", |
| 2449 | "-ffast-math", |
| 2450 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 2451 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2452 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2453 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2454 | "@FP16", |
| 2455 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2456 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2457 | ], |
| 2458 | ) |
| 2459 | |
| 2460 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2461 | name = "psimd_fastmath_ukernels_test_mode", |
| 2462 | hdrs = INTERNAL_HDRS, |
| 2463 | aarch32_copts = [ |
| 2464 | "-marm", |
| 2465 | "-mfpu=neon", |
| 2466 | ], |
| 2467 | copts = [ |
| 2468 | "-UNDEBUG", |
| 2469 | "-DXNN_TEST_MODE=1", |
| 2470 | ], |
| 2471 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2472 | gcc_x86_copts = ["-msse2"], |
| 2473 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2474 | optimized_copts = [ |
| 2475 | "-O3", |
| 2476 | "-ffast-math", |
| 2477 | ], |
| 2478 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
| 2479 | deps = [ |
| 2480 | ":tables", |
| 2481 | "@FP16", |
| 2482 | "@psimd", |
| 2483 | "@pthreadpool", |
| 2484 | ], |
| 2485 | ) |
| 2486 | |
| 2487 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2488 | name = "psimd_accmath_ukernels", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2489 | hdrs = INTERNAL_HDRS, |
| 2490 | aarch32_copts = [ |
| 2491 | "-marm", |
| 2492 | "-mfpu=neon", |
| 2493 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2494 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2495 | gcc_x86_copts = ["-msse2"], |
| 2496 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2497 | optimized_copts = [ |
| 2498 | "-O3", |
| 2499 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 2500 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2501 | deps = [ |
| 2502 | ":tables", |
| 2503 | "@FP16", |
| 2504 | "@psimd", |
| 2505 | "@pthreadpool", |
| 2506 | ], |
| 2507 | ) |
| 2508 | |
| 2509 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2510 | name = "psimd_accmath_ukernels_test_mode", |
| 2511 | hdrs = INTERNAL_HDRS, |
| 2512 | aarch32_copts = [ |
| 2513 | "-marm", |
| 2514 | "-mfpu=neon", |
| 2515 | ], |
| 2516 | copts = [ |
| 2517 | "-UNDEBUG", |
| 2518 | "-DXNN_TEST_MODE=1", |
| 2519 | ], |
| 2520 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2521 | gcc_x86_copts = ["-msse2"], |
| 2522 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2523 | optimized_copts = [ |
| 2524 | "-O3", |
| 2525 | ], |
| 2526 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
| 2527 | deps = [ |
| 2528 | ":tables", |
| 2529 | "@FP16", |
| 2530 | "@psimd", |
| 2531 | "@pthreadpool", |
| 2532 | ], |
| 2533 | ) |
| 2534 | |
| 2535 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2536 | name = "neon_ukernels", |
| 2537 | hdrs = INTERNAL_HDRS, |
| 2538 | aarch32_copts = [ |
| 2539 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2540 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2541 | "-mfpu=neon", |
| 2542 | ], |
| 2543 | aarch32_srcs = NEON_UKERNELS, |
| 2544 | aarch64_srcs = NEON_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2545 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2546 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2547 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2548 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2549 | "@FP16", |
| 2550 | "@pthreadpool", |
| 2551 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2552 | ) |
| 2553 | |
| 2554 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2555 | name = "neon_ukernels_test_mode", |
| 2556 | hdrs = INTERNAL_HDRS, |
| 2557 | aarch32_copts = [ |
| 2558 | "-marm", |
| 2559 | "-march=armv7-a", |
| 2560 | "-mfpu=neon", |
| 2561 | ], |
| 2562 | aarch32_srcs = NEON_UKERNELS, |
| 2563 | aarch64_srcs = NEON_UKERNELS, |
| 2564 | copts = [ |
| 2565 | "-UNDEBUG", |
| 2566 | "-DXNN_TEST_MODE=1", |
| 2567 | ], |
| 2568 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2569 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2570 | deps = [ |
| 2571 | ":tables", |
| 2572 | "@FP16", |
| 2573 | "@pthreadpool", |
| 2574 | ], |
| 2575 | ) |
| 2576 | |
| 2577 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2578 | name = "neonfma_ukernels", |
| 2579 | hdrs = INTERNAL_HDRS, |
| 2580 | aarch32_copts = [ |
| 2581 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2582 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2583 | "-mfpu=neon-vfpv4", |
| 2584 | ], |
| 2585 | aarch32_srcs = NEONFMA_UKERNELS, |
| 2586 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2587 | apple_aarch32_copts = [ |
| 2588 | "-mcpu=swift", |
| 2589 | "-mtune=generic", |
| 2590 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2591 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2592 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2593 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2594 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2595 | "@FP16", |
| 2596 | "@pthreadpool", |
| 2597 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2598 | ) |
| 2599 | |
| 2600 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2601 | name = "neonfma_ukernels_test_mode", |
| 2602 | hdrs = INTERNAL_HDRS, |
| 2603 | aarch32_copts = [ |
| 2604 | "-marm", |
| 2605 | "-march=armv7-a", |
| 2606 | "-mfpu=neon-vfpv4", |
| 2607 | ], |
| 2608 | aarch32_srcs = NEONFMA_UKERNELS, |
| 2609 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2610 | apple_aarch32_copts = [ |
| 2611 | "-mcpu=swift", |
| 2612 | "-mtune=generic", |
| 2613 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2614 | copts = [ |
| 2615 | "-UNDEBUG", |
| 2616 | "-DXNN_TEST_MODE=1", |
| 2617 | ], |
| 2618 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2619 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2620 | deps = [ |
| 2621 | ":tables", |
| 2622 | "@FP16", |
| 2623 | "@pthreadpool", |
| 2624 | ], |
| 2625 | ) |
| 2626 | |
| 2627 | xnnpack_cc_library( |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2628 | name = "neonv8_ukernels", |
| 2629 | hdrs = INTERNAL_HDRS, |
| 2630 | aarch32_copts = [ |
| 2631 | "-marm", |
| 2632 | "-march=armv8-a", |
| 2633 | "-mfpu=neon-fp-armv8", |
| 2634 | ], |
| 2635 | aarch32_srcs = NEONV8_UKERNELS, |
| 2636 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2637 | apple_aarch32_copts = [ |
| 2638 | "-mcpu=cyclone", |
| 2639 | "-mtune=generic", |
| 2640 | ], |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2641 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2642 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2643 | deps = [ |
| 2644 | ":tables", |
| 2645 | "@FP16", |
| 2646 | "@pthreadpool", |
| 2647 | ], |
| 2648 | ) |
| 2649 | |
| 2650 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2651 | name = "neonv8_ukernels_test_mode", |
| 2652 | hdrs = INTERNAL_HDRS, |
| 2653 | aarch32_copts = [ |
| 2654 | "-marm", |
| 2655 | "-march=armv8-a", |
| 2656 | "-mfpu=neon-fp-armv8", |
| 2657 | ], |
| 2658 | aarch32_srcs = NEONV8_UKERNELS, |
| 2659 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 2660 | apple_aarch32_copts = [ |
| 2661 | "-mcpu=cyclone", |
| 2662 | "-mtune=generic", |
| 2663 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2664 | copts = [ |
| 2665 | "-UNDEBUG", |
| 2666 | "-DXNN_TEST_MODE=1", |
| 2667 | ], |
| 2668 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2669 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2670 | deps = [ |
| 2671 | ":tables", |
| 2672 | "@FP16", |
| 2673 | "@pthreadpool", |
| 2674 | ], |
| 2675 | ) |
| 2676 | |
| 2677 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2678 | name = "neonfp16arith_ukernels", |
| 2679 | hdrs = INTERNAL_HDRS, |
| 2680 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 2681 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2682 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2683 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2684 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2685 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2686 | "@FP16", |
| 2687 | "@pthreadpool", |
| 2688 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2689 | ) |
| 2690 | |
| 2691 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2692 | name = "neonfp16arith_ukernels_test_mode", |
| 2693 | hdrs = INTERNAL_HDRS, |
| 2694 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 2695 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 2696 | copts = [ |
| 2697 | "-UNDEBUG", |
| 2698 | "-DXNN_TEST_MODE=1", |
| 2699 | ], |
| 2700 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2701 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2702 | deps = [ |
| 2703 | ":tables", |
| 2704 | "@FP16", |
| 2705 | "@pthreadpool", |
| 2706 | ], |
| 2707 | ) |
| 2708 | |
| 2709 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2710 | name = "sse2_ukernels", |
| 2711 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2712 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2713 | gcc_x86_copts = ["-msse2"], |
| 2714 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2715 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2716 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2717 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2718 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2719 | "@FP16", |
| 2720 | "@pthreadpool", |
| 2721 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2722 | ) |
| 2723 | |
| 2724 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2725 | name = "sse2_ukernels_test_mode", |
| 2726 | hdrs = INTERNAL_HDRS, |
| 2727 | copts = [ |
| 2728 | "-UNDEBUG", |
| 2729 | "-DXNN_TEST_MODE=1", |
| 2730 | ], |
| 2731 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2732 | gcc_x86_copts = ["-msse2"], |
| 2733 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2734 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 2735 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| 2736 | deps = [ |
| 2737 | ":tables", |
| 2738 | "@FP16", |
| 2739 | "@pthreadpool", |
| 2740 | ], |
| 2741 | ) |
| 2742 | |
| 2743 | xnnpack_cc_library( |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2744 | name = "ssse3_ukernels", |
| 2745 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2746 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2747 | gcc_x86_copts = ["-mssse3"], |
| 2748 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2749 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2750 | x86_srcs = SSSE3_UKERNELS, |
| 2751 | deps = [ |
| 2752 | ":tables", |
| 2753 | "@FP16", |
| 2754 | "@pthreadpool", |
| 2755 | ], |
| 2756 | ) |
| 2757 | |
| 2758 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2759 | name = "ssse3_ukernels_test_mode", |
| 2760 | hdrs = INTERNAL_HDRS, |
| 2761 | copts = [ |
| 2762 | "-UNDEBUG", |
| 2763 | "-DXNN_TEST_MODE=1", |
| 2764 | ], |
| 2765 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2766 | gcc_x86_copts = ["-mssse3"], |
| 2767 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2768 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 2769 | x86_srcs = SSSE3_UKERNELS, |
| 2770 | deps = [ |
| 2771 | ":tables", |
| 2772 | "@FP16", |
| 2773 | "@pthreadpool", |
| 2774 | ], |
| 2775 | ) |
| 2776 | |
| 2777 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2778 | name = "sse41_ukernels", |
| 2779 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2780 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2781 | gcc_x86_copts = ["-msse4.1"], |
| 2782 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2783 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2784 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2785 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2786 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2787 | "@FP16", |
| 2788 | "@pthreadpool", |
| 2789 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2790 | ) |
| 2791 | |
| 2792 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2793 | name = "sse41_ukernels_test_mode", |
| 2794 | hdrs = INTERNAL_HDRS, |
| 2795 | copts = [ |
| 2796 | "-UNDEBUG", |
| 2797 | "-DXNN_TEST_MODE=1", |
| 2798 | ], |
| 2799 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2800 | gcc_x86_copts = ["-msse4.1"], |
| 2801 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2802 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 2803 | x86_srcs = SSE41_UKERNELS, |
| 2804 | deps = [ |
| 2805 | ":tables", |
| 2806 | "@FP16", |
| 2807 | "@pthreadpool", |
| 2808 | ], |
| 2809 | ) |
| 2810 | |
| 2811 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2812 | name = "avx_ukernels", |
| 2813 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2814 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2815 | gcc_x86_copts = ["-mavx"], |
| 2816 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2817 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2818 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2819 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2820 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2821 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2822 | "@FP16", |
| 2823 | "@pthreadpool", |
| 2824 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2825 | ) |
| 2826 | |
| 2827 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2828 | name = "avx_ukernels_test_mode", |
| 2829 | hdrs = INTERNAL_HDRS, |
| 2830 | copts = [ |
| 2831 | "-UNDEBUG", |
| 2832 | "-DXNN_TEST_MODE=1", |
| 2833 | ], |
| 2834 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2835 | gcc_x86_copts = ["-mavx"], |
| 2836 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2837 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2838 | msvc_x86_64_copts = ["/arch:AVX"], |
| 2839 | x86_srcs = AVX_UKERNELS, |
| 2840 | deps = [ |
| 2841 | ":tables", |
| 2842 | "@FP16", |
| 2843 | "@pthreadpool", |
| 2844 | ], |
| 2845 | ) |
| 2846 | |
| 2847 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2848 | name = "fma3_ukernels", |
| 2849 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2850 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2851 | gcc_x86_copts = ["-mfma"], |
| 2852 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2853 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2854 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2855 | x86_srcs = FMA3_UKERNELS, |
| 2856 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2857 | ":tables", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2858 | "@FP16", |
| 2859 | "@pthreadpool", |
| 2860 | ], |
| 2861 | ) |
| 2862 | |
| 2863 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2864 | name = "fma3_ukernels_test_mode", |
| 2865 | hdrs = INTERNAL_HDRS, |
| 2866 | copts = [ |
| 2867 | "-UNDEBUG", |
| 2868 | "-DXNN_TEST_MODE=1", |
| 2869 | ], |
| 2870 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2871 | gcc_x86_copts = ["-mfma"], |
| 2872 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2873 | msvc_x86_32_copts = ["/arch:AVX"], |
| 2874 | msvc_x86_64_copts = ["/arch:AVX"], |
| 2875 | x86_srcs = FMA3_UKERNELS, |
| 2876 | deps = [ |
| 2877 | ":tables", |
| 2878 | "@FP16", |
| 2879 | "@pthreadpool", |
| 2880 | ], |
| 2881 | ) |
| 2882 | |
| 2883 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2884 | name = "avx2_ukernels", |
| 2885 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2886 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2887 | gcc_x86_copts = [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2888 | "-mfma", |
| 2889 | "-mavx2", |
| 2890 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2891 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2892 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 2893 | msvc_x86_64_copts = ["/arch:AVX2"], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2894 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2895 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2896 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2897 | "@FP16", |
| 2898 | "@pthreadpool", |
| 2899 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2900 | ) |
| 2901 | |
| 2902 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2903 | name = "avx2_ukernels_test_mode", |
| 2904 | hdrs = INTERNAL_HDRS, |
| 2905 | copts = [ |
| 2906 | "-UNDEBUG", |
| 2907 | "-DXNN_TEST_MODE=1", |
| 2908 | ], |
| 2909 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2910 | gcc_x86_copts = [ |
| 2911 | "-mfma", |
| 2912 | "-mavx2", |
| 2913 | ], |
| 2914 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2915 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 2916 | msvc_x86_64_copts = ["/arch:AVX2"], |
| 2917 | x86_srcs = AVX2_UKERNELS, |
| 2918 | deps = [ |
| 2919 | ":tables", |
| 2920 | "@FP16", |
| 2921 | "@pthreadpool", |
| 2922 | ], |
| 2923 | ) |
| 2924 | |
| 2925 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2926 | name = "avx512f_ukernels", |
| 2927 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2928 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2929 | gcc_x86_copts = ["-mavx512f"], |
| 2930 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 2931 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2932 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 2933 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 2934 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2935 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2936 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2937 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2938 | "@FP16", |
| 2939 | "@pthreadpool", |
| 2940 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2941 | ) |
| 2942 | |
| 2943 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2944 | name = "avx512f_ukernels_test_mode", |
| 2945 | hdrs = INTERNAL_HDRS, |
| 2946 | copts = [ |
| 2947 | "-UNDEBUG", |
| 2948 | "-DXNN_TEST_MODE=1", |
| 2949 | ], |
| 2950 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2951 | gcc_x86_copts = ["-mavx512f"], |
| 2952 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 2953 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2954 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 2955 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 2956 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 2957 | x86_srcs = AVX512F_UKERNELS, |
| 2958 | deps = [ |
| 2959 | ":tables", |
| 2960 | "@FP16", |
| 2961 | "@pthreadpool", |
| 2962 | ], |
| 2963 | ) |
| 2964 | |
| 2965 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2966 | name = "asm_ukernels", |
| 2967 | hdrs = ["src/xnnpack/assembly.h"], |
| 2968 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 2969 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2970 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| 2971 | ) |
| 2972 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2973 | xnnpack_cc_library( |
| 2974 | name = "logging_utils", |
| 2975 | srcs = LOGGING_SRCS, |
| 2976 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 2977 | copts = LOGGING_COPTS + [ |
| 2978 | "-Isrc", |
| 2979 | "-Iinclude", |
| 2980 | ] + select({ |
| 2981 | ":debug_build": [], |
| 2982 | "//conditions:default": xnnpack_min_size_copts(), |
| 2983 | }), |
| 2984 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2985 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2986 | visibility = xnnpack_visibility(), |
| 2987 | deps = [ |
| 2988 | "@FP16", |
| 2989 | "@clog", |
| 2990 | "@pthreadpool", |
| 2991 | ], |
| 2992 | ) |
| 2993 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2994 | xnnpack_aggregate_library( |
| 2995 | name = "ukernels", |
| 2996 | aarch32_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2997 | ":neon_ukernels", |
| 2998 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2999 | ":neonv8_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3000 | ":asm_ukernels", |
| 3001 | ], |
| 3002 | aarch64_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3003 | ":neon_ukernels", |
| 3004 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3005 | ":neonv8_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3006 | ":neonfp16arith_ukernels", |
| 3007 | ":asm_ukernels", |
| 3008 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3009 | generic_deps = [ |
| 3010 | ":scalar_ukernels", |
| 3011 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3012 | psimd_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3013 | ":psimd_fastmath_ukernels", |
| 3014 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3015 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3016 | wasm_deps = [ |
| 3017 | ":wasm_ukernels", |
| 3018 | ], |
| 3019 | wasmsimd_deps = [ |
| 3020 | ":wasm_ukernels", |
| 3021 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3022 | x86_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3023 | ":sse2_ukernels", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3024 | ":ssse3_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3025 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3026 | ":avx_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3027 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3028 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3029 | ":avx512f_ukernels", |
| 3030 | ], |
| 3031 | ) |
| 3032 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3033 | xnnpack_aggregate_library( |
| 3034 | name = "ukernels_test_mode", |
| 3035 | aarch32_deps = [ |
| 3036 | ":neon_ukernels_test_mode", |
| 3037 | ":neonfma_ukernels_test_mode", |
| 3038 | ":neonv8_ukernels_test_mode", |
| 3039 | ":asm_ukernels", |
| 3040 | ], |
| 3041 | aarch64_deps = [ |
| 3042 | ":neon_ukernels_test_mode", |
| 3043 | ":neonfma_ukernels_test_mode", |
| 3044 | ":neonv8_ukernels_test_mode", |
| 3045 | ":neonfp16arith_ukernels_test_mode", |
| 3046 | ":asm_ukernels", |
| 3047 | ], |
| 3048 | generic_deps = [ |
| 3049 | ":scalar_ukernels_test_mode", |
| 3050 | ], |
| 3051 | psimd_deps = [ |
| 3052 | ":psimd_fastmath_ukernels_test_mode", |
| 3053 | ":psimd_accmath_ukernels_test_mode", |
| 3054 | ], |
| 3055 | wasm_deps = [ |
| 3056 | ":wasm_ukernels_test_mode", |
| 3057 | ], |
| 3058 | wasmsimd_deps = [ |
| 3059 | ":wasm_ukernels_test_mode", |
| 3060 | ], |
| 3061 | x86_deps = [ |
| 3062 | ":sse2_ukernels_test_mode", |
| 3063 | ":ssse3_ukernels_test_mode", |
| 3064 | ":sse41_ukernels_test_mode", |
| 3065 | ":avx_ukernels_test_mode", |
| 3066 | ":fma3_ukernels_test_mode", |
| 3067 | ":avx2_ukernels_test_mode", |
| 3068 | ":avx512f_ukernels_test_mode", |
| 3069 | ], |
| 3070 | ) |
| 3071 | |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3072 | # This target must be compatible with Android NDK r17 and compile with Clang 6. |
| 3073 | # It is similar to :ukernels target, but excludes microkernels which require |
| 3074 | # NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics. |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3075 | xnnpack_aggregate_library( |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3076 | name = "ukernels_max_compatibility", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3077 | aarch32_deps = [ |
| 3078 | ":neon_ukernels", |
| 3079 | ":neonfma_ukernels", |
| 3080 | ":neonv8_ukernels", |
| 3081 | ":asm_ukernels", |
| 3082 | ], |
| 3083 | aarch64_deps = [ |
| 3084 | ":neon_ukernels", |
| 3085 | ":neonfma_ukernels", |
| 3086 | ":neonv8_ukernels", |
| 3087 | ":asm_ukernels", |
| 3088 | ], |
| 3089 | generic_deps = [ |
| 3090 | ":scalar_ukernels", |
| 3091 | ], |
| 3092 | psimd_deps = [ |
| 3093 | ":psimd_fastmath_ukernels", |
| 3094 | ":psimd_accmath_ukernels", |
| 3095 | ], |
| 3096 | wasm_deps = [ |
| 3097 | ":wasm_ukernels", |
| 3098 | ], |
| 3099 | wasmsimd_deps = [ |
| 3100 | ":wasm_ukernels", |
| 3101 | ], |
| 3102 | x86_deps = [ |
| 3103 | ":sse2_ukernels", |
| 3104 | ":ssse3_ukernels", |
| 3105 | ":sse41_ukernels", |
| 3106 | ":avx_ukernels", |
| 3107 | ":fma3_ukernels", |
| 3108 | ":avx2_ukernels", |
| 3109 | ":avx512f_ukernels", |
| 3110 | ], |
| 3111 | ) |
| 3112 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3113 | xnnpack_cc_library( |
| 3114 | name = "im2col", |
| 3115 | srcs = ["src/im2col.c"], |
| 3116 | hdrs = [ |
| 3117 | "src/xnnpack/common.h", |
| 3118 | "src/xnnpack/im2col.h", |
| 3119 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3120 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3121 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3122 | ) |
| 3123 | |
| 3124 | xnnpack_cc_library( |
| 3125 | name = "indirection", |
| 3126 | srcs = ["src/indirection.c"], |
| 3127 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3128 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3129 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3130 | deps = [ |
| 3131 | "@FP16", |
| 3132 | "@FXdiv", |
| 3133 | "@pthreadpool", |
| 3134 | ], |
| 3135 | ) |
| 3136 | |
| 3137 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3138 | name = "indirection_test_mode", |
| 3139 | srcs = ["src/indirection.c"], |
| 3140 | hdrs = INTERNAL_HDRS, |
| 3141 | copts = [ |
| 3142 | "-UNDEBUG", |
| 3143 | "-DXNN_TEST_MODE=1", |
| 3144 | ], |
| 3145 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3146 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3147 | deps = [ |
| 3148 | "@FP16", |
| 3149 | "@FXdiv", |
| 3150 | "@pthreadpool", |
| 3151 | ], |
| 3152 | ) |
| 3153 | |
| 3154 | xnnpack_cc_library( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3155 | name = "packing", |
| 3156 | srcs = ["src/packing.c"], |
| 3157 | hdrs = INTERNAL_HDRS, |
| 3158 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3159 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3160 | deps = [ |
| 3161 | "@FP16", |
| 3162 | "@FXdiv", |
| 3163 | "@pthreadpool", |
| 3164 | ], |
| 3165 | ) |
| 3166 | |
| 3167 | xnnpack_cc_library( |
| 3168 | name = "packing_test_mode", |
| 3169 | srcs = ["src/packing.c"], |
| 3170 | hdrs = INTERNAL_HDRS, |
| 3171 | copts = [ |
| 3172 | "-UNDEBUG", |
| 3173 | "-DXNN_TEST_MODE=1", |
| 3174 | ], |
| 3175 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3176 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3177 | deps = [ |
| 3178 | "@FP16", |
| 3179 | "@FXdiv", |
| 3180 | "@pthreadpool", |
| 3181 | ], |
| 3182 | ) |
| 3183 | |
| 3184 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3185 | name = "operator_run", |
| 3186 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 3187 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3188 | copts = LOGGING_COPTS + select({ |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3189 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3190 | "//conditions:default": [], |
| 3191 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3192 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3193 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3194 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3195 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3196 | "@FP16", |
| 3197 | "@FXdiv", |
| 3198 | "@clog", |
| 3199 | "@pthreadpool", |
| 3200 | ], |
| 3201 | ) |
| 3202 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3203 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3204 | name = "operator_run_test_mode", |
| 3205 | srcs = ["src/operator-run.c"], |
| 3206 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3207 | copts = LOGGING_COPTS + [ |
| 3208 | "-UNDEBUG", |
| 3209 | "-DXNN_TEST_MODE=1", |
| 3210 | ] + select({ |
| 3211 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3212 | "//conditions:default": [], |
| 3213 | }), |
| 3214 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3215 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3216 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3217 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3218 | "@FP16", |
| 3219 | "@FXdiv", |
| 3220 | "@clog", |
| 3221 | "@pthreadpool", |
| 3222 | ], |
| 3223 | ) |
| 3224 | |
| 3225 | xnnpack_cc_library( |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3226 | name = "memory_planner", |
| 3227 | srcs = ["src/memory-planner.c"], |
| 3228 | hdrs = INTERNAL_HDRS, |
| 3229 | defines = select({ |
| 3230 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 3231 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 3232 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 3233 | }), |
| 3234 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3235 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3236 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3237 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3238 | "@pthreadpool", |
| 3239 | ], |
| 3240 | ) |
| 3241 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3242 | xnnpack_cc_library( |
| 3243 | name = "memory_planner_test_mode", |
| 3244 | srcs = ["src/memory-planner.c"], |
| 3245 | hdrs = INTERNAL_HDRS, |
| 3246 | copts = [ |
| 3247 | "-UNDEBUG", |
| 3248 | "-DXNN_TEST_MODE=1", |
| 3249 | ], |
| 3250 | defines = select({ |
| 3251 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 3252 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 3253 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 3254 | }), |
| 3255 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3256 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3257 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3258 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3259 | "@pthreadpool", |
| 3260 | ], |
| 3261 | ) |
| 3262 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3263 | cc_library( |
| 3264 | name = "enable_assembly", |
| 3265 | defines = select({ |
| 3266 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 3267 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 3268 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3269 | }), |
| 3270 | ) |
| 3271 | |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3272 | cc_library( |
| 3273 | name = "enable_sparse", |
| 3274 | defines = select({ |
| 3275 | ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"], |
| 3276 | ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"], |
Marat Dukhan | 8b283aa | 2020-06-18 16:48:57 -0700 | [diff] [blame] | 3277 | "//conditions:default": ["XNN_ENABLE_SPARSE=0"], |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3278 | }), |
| 3279 | ) |
| 3280 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3281 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3282 | name = "operators", |
| 3283 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3284 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3285 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3286 | ], |
| 3287 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3288 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3289 | "-Isrc", |
| 3290 | "-Iinclude", |
| 3291 | ] + select({ |
| 3292 | ":debug_build": [], |
| 3293 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3294 | }) + select({ |
| 3295 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3296 | "//conditions:default": [], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3297 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3298 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3299 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3300 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3301 | ":indirection", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3302 | ":logging_utils", |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3303 | ":packing", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3304 | "@FP16", |
| 3305 | "@FXdiv", |
| 3306 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3307 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3308 | ], |
| 3309 | ) |
| 3310 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3311 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3312 | name = "operators_test_mode", |
| 3313 | srcs = OPERATOR_SRCS + [ |
| 3314 | "src/memory.c", |
| 3315 | "src/operator-delete.c", |
| 3316 | ], |
| 3317 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3318 | copts = LOGGING_COPTS + [ |
| 3319 | "-Isrc", |
| 3320 | "-Iinclude", |
| 3321 | "-UNDEBUG", |
| 3322 | "-DXNN_TEST_MODE=1", |
| 3323 | ] + select({ |
| 3324 | ":debug_build": [], |
| 3325 | "//conditions:default": xnnpack_min_size_copts(), |
| 3326 | }) + select({ |
| 3327 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3328 | "//conditions:default": [], |
| 3329 | }), |
| 3330 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3331 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3332 | deps = [ |
| 3333 | ":indirection_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3334 | ":logging_utils", |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3335 | ":packing_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3336 | "@FP16", |
| 3337 | "@FXdiv", |
| 3338 | "@clog", |
| 3339 | "@pthreadpool", |
| 3340 | ], |
| 3341 | ) |
| 3342 | |
| 3343 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3344 | name = "XNNPACK", |
| 3345 | srcs = [ |
| 3346 | "src/init.c", |
Marat Dukhan | ccfdbd1 | 2020-02-03 14:27:45 -0800 | [diff] [blame] | 3347 | "src/runtime.c", |
| 3348 | "src/subgraph.c", |
| 3349 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 3350 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3351 | hdrs = ["include/xnnpack.h"], |
| 3352 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3353 | "-Isrc", |
| 3354 | "-Iinclude", |
| 3355 | ] + select({ |
| 3356 | ":debug_build": [], |
| 3357 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3358 | }) + select({ |
| 3359 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3360 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3361 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3362 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3363 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3364 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3365 | visibility = xnnpack_visibility(), |
| 3366 | deps = [ |
| 3367 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3368 | ":enable_sparse", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3369 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3370 | ":memory_planner", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3371 | ":operator_run", |
| 3372 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3373 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3374 | "@clog", |
Marat Dukhan | ab2946c | 2020-05-21 20:04:13 -0700 | [diff] [blame] | 3375 | "@FP16", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3376 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 3377 | ] + select({ |
| 3378 | ":emscripten": [], |
| 3379 | "//conditions:default": ["@cpuinfo"], |
| 3380 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3381 | ) |
| 3382 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3383 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3384 | name = "XNNPACK_test_mode", |
| 3385 | srcs = [ |
| 3386 | "src/init.c", |
| 3387 | "src/runtime.c", |
| 3388 | "src/subgraph.c", |
| 3389 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 3390 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3391 | hdrs = ["include/xnnpack.h"], |
| 3392 | copts = LOGGING_COPTS + [ |
| 3393 | "-Isrc", |
| 3394 | "-Iinclude", |
| 3395 | "-UNDEBUG", |
| 3396 | "-DXNN_TEST_MODE=1", |
| 3397 | ] + select({ |
| 3398 | ":debug_build": [], |
| 3399 | "//conditions:default": xnnpack_min_size_copts(), |
| 3400 | }) + select({ |
| 3401 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3402 | "//conditions:default": [], |
| 3403 | }), |
| 3404 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3405 | includes = ["include"], |
| 3406 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3407 | visibility = xnnpack_visibility(), |
| 3408 | deps = [ |
| 3409 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3410 | ":enable_sparse", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3411 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3412 | ":memory_planner_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3413 | ":operator_run_test_mode", |
| 3414 | ":operators_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3415 | ":ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3416 | "@clog", |
| 3417 | "@FP16", |
| 3418 | "@pthreadpool", |
| 3419 | ] + select({ |
| 3420 | ":emscripten": [], |
| 3421 | "//conditions:default": ["@cpuinfo"], |
| 3422 | }), |
| 3423 | ) |
| 3424 | |
| 3425 | xnnpack_cc_library( |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3426 | name = "xnnpack_f32", |
| 3427 | srcs = [ |
| 3428 | "src/init.c", |
| 3429 | "src/runtime.c", |
| 3430 | "src/subgraph.c", |
| 3431 | "src/tensor.c", |
| 3432 | ] + SUBGRAPH_SRCS, |
| 3433 | hdrs = ["include/xnnpack.h"], |
| 3434 | copts = LOGGING_COPTS + [ |
| 3435 | "-Isrc", |
| 3436 | "-Iinclude", |
| 3437 | ] + select({ |
| 3438 | ":debug_build": [], |
| 3439 | "//conditions:default": xnnpack_min_size_copts(), |
| 3440 | }) + select({ |
| 3441 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3442 | "//conditions:default": [], |
| 3443 | }), |
| 3444 | defines = [ |
| 3445 | "XNN_NO_Q8_OPERATORS", |
| 3446 | "XNN_NO_U8_OPERATORS", |
| 3447 | "XNN_NO_X8_OPERATORS", |
| 3448 | "XNN_NO_F16_OPERATORS", |
| 3449 | "XNN_NO_X16_OPERATORS", |
| 3450 | ], |
| 3451 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3452 | includes = ["include"], |
| 3453 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3454 | visibility = xnnpack_visibility(), |
| 3455 | deps = [ |
| 3456 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3457 | ":enable_sparse", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3458 | ":logging_utils", |
| 3459 | ":memory_planner", |
| 3460 | ":operator_run", |
| 3461 | ":operators", |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3462 | ":ukernels_max_compatibility", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3463 | "@clog", |
| 3464 | "@FP16", |
| 3465 | "@pthreadpool", |
| 3466 | ] + select({ |
| 3467 | ":emscripten": [], |
| 3468 | "//conditions:default": ["@cpuinfo"], |
| 3469 | }), |
| 3470 | ) |
| 3471 | |
| 3472 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3473 | name = "xnnpack_operators_nhwc_f32", |
| 3474 | srcs = [ |
| 3475 | "src/init.c", |
| 3476 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3477 | hdrs = ["include/xnnpack.h"], |
| 3478 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3479 | "-Isrc", |
| 3480 | "-Iinclude", |
| 3481 | ] + select({ |
| 3482 | ":debug_build": [], |
| 3483 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3484 | }) + select({ |
| 3485 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3486 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3487 | }), |
| 3488 | defines = [ |
| 3489 | "XNN_NO_Q8_OPERATORS", |
| 3490 | "XNN_NO_U8_OPERATORS", |
| 3491 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 3492 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3493 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3494 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3495 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3496 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3497 | visibility = xnnpack_visibility(), |
| 3498 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3499 | ":enable_assembly", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3500 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3501 | ":operator_run", |
| 3502 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3503 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3504 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3505 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 3506 | ] + select({ |
| 3507 | ":emscripten": [], |
| 3508 | "//conditions:default": ["@cpuinfo"], |
| 3509 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3510 | ) |
| 3511 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3512 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3513 | name = "bench_utils", |
| 3514 | srcs = ["bench/utils.cc"], |
| 3515 | hdrs = ["bench/utils.h"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 3516 | deps = [ |
| 3517 | "@com_google_benchmark//:benchmark", |
| 3518 | "@cpuinfo", |
| 3519 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3520 | ) |
| 3521 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 3522 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3523 | |
| 3524 | xnnpack_benchmark( |
| 3525 | name = "q8_gemm_bench", |
| 3526 | srcs = [ |
| 3527 | "bench/gemm.h", |
| 3528 | "bench/q8-gemm.cc", |
| 3529 | "src/xnnpack/AlignedAllocator.h", |
| 3530 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3531 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3532 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3533 | ) |
| 3534 | |
| 3535 | xnnpack_benchmark( |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 3536 | name = "f16_igemm_bench", |
| 3537 | srcs = [ |
| 3538 | "bench/f16-igemm.cc", |
| 3539 | "bench/conv.h", |
| 3540 | "bench/google/conv.h", |
| 3541 | "src/xnnpack/AlignedAllocator.h", |
| 3542 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3543 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3544 | ":indirection", |
| 3545 | ":packing", |
| 3546 | ], |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 3547 | ) |
| 3548 | |
| 3549 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3550 | name = "f16_gemm_bench", |
| 3551 | srcs = [ |
| 3552 | "bench/f16-gemm.cc", |
| 3553 | "bench/gemm.h", |
| 3554 | "src/xnnpack/AlignedAllocator.h", |
| 3555 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3556 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3557 | ":packing", |
| 3558 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3559 | ) |
| 3560 | |
| 3561 | xnnpack_benchmark( |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3562 | name = "f16_spmm_bench", |
| 3563 | srcs = [ |
| 3564 | "bench/f16-spmm.cc", |
| 3565 | "bench/gemm.h", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3566 | "src/xnnpack/AlignedAllocator.h", |
| 3567 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 3568 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3569 | ) |
| 3570 | |
| 3571 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3572 | name = "f32_igemm_bench", |
| 3573 | srcs = [ |
| 3574 | "bench/f32-igemm.cc", |
| 3575 | "bench/conv.h", |
| 3576 | "src/xnnpack/AlignedAllocator.h", |
| 3577 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3578 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3579 | ":indirection", |
| 3580 | ":packing", |
| 3581 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3582 | ) |
| 3583 | |
| 3584 | xnnpack_benchmark( |
| 3585 | name = "f32_conv_hwc_bench", |
| 3586 | srcs = [ |
| 3587 | "bench/f32-conv-hwc.cc", |
| 3588 | "bench/dconv.h", |
| 3589 | "src/xnnpack/AlignedAllocator.h", |
| 3590 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3591 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3592 | ":packing", |
| 3593 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3594 | ) |
| 3595 | |
| 3596 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3597 | name = "f32_conv_hwc2chw_bench", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 3598 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3599 | "bench/f32-conv-hwc2chw.cc", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 3600 | "bench/dconv.h", |
| 3601 | "src/xnnpack/AlignedAllocator.h", |
| 3602 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3603 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3604 | ":packing", |
| 3605 | ], |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 3606 | ) |
| 3607 | |
| 3608 | xnnpack_benchmark( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 3609 | name = "f16_dwconv_bench", |
| 3610 | srcs = [ |
| 3611 | "bench/f16-dwconv.cc", |
| 3612 | "bench/dwconv.h", |
| 3613 | "bench/google/dwconv.h", |
| 3614 | "src/xnnpack/AlignedAllocator.h", |
| 3615 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3616 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3617 | ":indirection", |
| 3618 | ":packing", |
| 3619 | ], |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 3620 | ) |
| 3621 | |
| 3622 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3623 | name = "f32_dwconv_bench", |
| 3624 | srcs = [ |
| 3625 | "bench/f32-dwconv.cc", |
| 3626 | "bench/dwconv.h", |
| 3627 | "src/xnnpack/AlignedAllocator.h", |
| 3628 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3629 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3630 | ":indirection", |
| 3631 | ":packing", |
| 3632 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3633 | ) |
| 3634 | |
| 3635 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3636 | name = "f32_dwconv_chw_bench", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3637 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 3638 | "bench/f32-dwconv-chw.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3639 | "bench/dwconv.h", |
| 3640 | "src/xnnpack/AlignedAllocator.h", |
| 3641 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3642 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3643 | ":indirection", |
| 3644 | ":packing", |
| 3645 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3646 | ) |
| 3647 | |
| 3648 | xnnpack_benchmark( |
| 3649 | name = "f32_gemm_bench", |
| 3650 | srcs = [ |
| 3651 | "bench/f32-gemm.cc", |
| 3652 | "bench/gemm.h", |
| 3653 | "src/xnnpack/AlignedAllocator.h", |
| 3654 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3655 | copts = xnnpack_optional_ruy_copts(), |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3656 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3657 | ) |
| 3658 | |
| 3659 | xnnpack_benchmark( |
Marat Dukhan | a11ca34 | 2020-06-25 23:45:07 -0700 | [diff] [blame] | 3660 | name = "f32_hswish_bench", |
| 3661 | srcs = [ |
| 3662 | "bench/f32-hswish.cc", |
| 3663 | "src/xnnpack/AlignedAllocator.h", |
| 3664 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3665 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3666 | ) |
| 3667 | |
| 3668 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 3669 | name = "f32_raddexpminusmax_bench", |
| 3670 | srcs = [ |
| 3671 | "bench/f32-raddexpminusmax.cc", |
| 3672 | "src/xnnpack/AlignedAllocator.h", |
| 3673 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3674 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3675 | ) |
| 3676 | |
| 3677 | xnnpack_benchmark( |
| 3678 | name = "f32_raddextexp_bench", |
| 3679 | srcs = [ |
| 3680 | "bench/f32-raddextexp.cc", |
| 3681 | "src/xnnpack/AlignedAllocator.h", |
| 3682 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3683 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3684 | ) |
| 3685 | |
| 3686 | xnnpack_benchmark( |
| 3687 | name = "f32_raddstoreexpminusmax_bench", |
| 3688 | srcs = [ |
| 3689 | "bench/f32-raddstoreexpminusmax.cc", |
| 3690 | "src/xnnpack/AlignedAllocator.h", |
| 3691 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3692 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3693 | ) |
| 3694 | |
| 3695 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3696 | name = "f32_rmax_bench", |
| 3697 | srcs = [ |
| 3698 | "bench/f32-rmax.cc", |
| 3699 | "src/xnnpack/AlignedAllocator.h", |
| 3700 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3701 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3702 | ) |
| 3703 | |
| 3704 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 3705 | name = "f32_sigmoid_bench", |
| 3706 | srcs = [ |
| 3707 | "bench/f32-sigmoid.cc", |
| 3708 | "src/xnnpack/AlignedAllocator.h", |
| 3709 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 3710 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3711 | ) |
| 3712 | |
| 3713 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3714 | name = "f32_spmm_bench", |
| 3715 | srcs = [ |
| 3716 | "bench/f32-spmm.cc", |
| 3717 | "bench/gemm.h", |
| 3718 | "src/xnnpack/AlignedAllocator.h", |
| 3719 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3720 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3721 | ) |
| 3722 | |
| 3723 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3724 | name = "f32_softmax_bench", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 3725 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3726 | "bench/f32-softmax.cc", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 3727 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3728 | copts = xnnpack_optional_dnnl_copts(), |
Marat Dukhan | 8d3c693 | 2020-03-06 20:27:27 -0800 | [diff] [blame] | 3729 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(), |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 3730 | ) |
| 3731 | |
| 3732 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 3733 | name = "f32_vscaleexpminusmax_bench", |
| 3734 | srcs = [ |
| 3735 | "bench/f32-vscaleexpminusmax.cc", |
| 3736 | "src/xnnpack/AlignedAllocator.h", |
| 3737 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3738 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3739 | ) |
| 3740 | |
| 3741 | xnnpack_benchmark( |
| 3742 | name = "f32_vscaleextexp_bench", |
| 3743 | srcs = [ |
| 3744 | "bench/f32-vscaleextexp.cc", |
| 3745 | "src/xnnpack/AlignedAllocator.h", |
| 3746 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3747 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3748 | ) |
| 3749 | |
| 3750 | xnnpack_benchmark( |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 3751 | name = "f32_vsqrt_bench", |
| 3752 | srcs = [ |
| 3753 | "bench/f32-vsqrt.cc", |
| 3754 | "src/xnnpack/AlignedAllocator.h", |
| 3755 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3756 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3757 | ) |
| 3758 | |
| 3759 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3760 | name = "f32_im2col_gemm_bench", |
| 3761 | srcs = [ |
| 3762 | "bench/f32-im2col-gemm.cc", |
| 3763 | "bench/conv.h", |
| 3764 | "src/xnnpack/AlignedAllocator.h", |
| 3765 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3766 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3767 | ":im2col", |
| 3768 | ":packing", |
| 3769 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3770 | ) |
| 3771 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3772 | xnnpack_benchmark( |
| 3773 | name = "requantization_bench", |
| 3774 | srcs = [ |
| 3775 | "bench/requantization.cc", |
| 3776 | "src/xnnpack/requantization-stubs.h", |
| 3777 | "src/xnnpack/AlignedAllocator.h", |
| 3778 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3779 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3780 | ) |
| 3781 | |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 3782 | xnnpack_benchmark( |
| 3783 | name = "rounding_bench", |
| 3784 | srcs = [ |
| 3785 | "bench/rounding.cc", |
| 3786 | "src/xnnpack/math-stubs.h", |
| 3787 | "src/xnnpack/AlignedAllocator.h", |
| 3788 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 3789 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 3790 | ) |
| 3791 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3792 | ########################### Benchmarks for operators ########################### |
| 3793 | |
| 3794 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3795 | name = "average_pooling_bench", |
| 3796 | srcs = ["bench/average-pooling.cc"], |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 3797 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3798 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3799 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3800 | ) |
| 3801 | |
| 3802 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 3803 | name = "bankers_rounding_bench", |
| 3804 | srcs = ["bench/bankers-rounding.cc"], |
| 3805 | copts = xnnpack_optional_tflite_copts(), |
| 3806 | tags = ["nowin32"], |
| 3807 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 3808 | ) |
| 3809 | |
| 3810 | xnnpack_benchmark( |
| 3811 | name = "ceiling_bench", |
| 3812 | srcs = ["bench/ceiling.cc"], |
| 3813 | copts = xnnpack_optional_tflite_copts(), |
| 3814 | tags = ["nowin32"], |
| 3815 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 3816 | ) |
| 3817 | |
| 3818 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3819 | name = "channel_shuffle_bench", |
| 3820 | srcs = ["bench/channel-shuffle.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3821 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3822 | ) |
| 3823 | |
| 3824 | xnnpack_benchmark( |
| 3825 | name = "convolution_bench", |
| 3826 | srcs = ["bench/convolution.cc"], |
| 3827 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3828 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3829 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3830 | ) |
| 3831 | |
| 3832 | xnnpack_benchmark( |
| 3833 | name = "deconvolution_bench", |
| 3834 | srcs = ["bench/deconvolution.cc"], |
| 3835 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3836 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3837 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3838 | ) |
| 3839 | |
| 3840 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 3841 | name = "floor_bench", |
| 3842 | srcs = ["bench/floor.cc"], |
| 3843 | copts = xnnpack_optional_tflite_copts(), |
| 3844 | tags = ["nowin32"], |
| 3845 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 3846 | ) |
| 3847 | |
| 3848 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3849 | name = "global_average_pooling_bench", |
| 3850 | srcs = ["bench/global-average-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3851 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3852 | ) |
| 3853 | |
| 3854 | xnnpack_benchmark( |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 3855 | name = "hardswish_bench", |
| 3856 | srcs = ["bench/hardswish.cc"], |
| 3857 | copts = xnnpack_optional_tflite_copts(), |
| 3858 | tags = ["nowin32"], |
| 3859 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 3860 | ) |
| 3861 | |
| 3862 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3863 | name = "max_pooling_bench", |
| 3864 | srcs = ["bench/max-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3865 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3866 | ) |
| 3867 | |
| 3868 | xnnpack_benchmark( |
| 3869 | name = "sigmoid_bench", |
| 3870 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 3871 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 3872 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3873 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3874 | ) |
| 3875 | |
| 3876 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 3877 | name = "prelu_bench", |
| 3878 | srcs = ["bench/prelu.cc"], |
| 3879 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 3880 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3881 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 3882 | ) |
| 3883 | |
| 3884 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 3885 | name = "softmax_bench", |
| 3886 | srcs = ["bench/softmax.cc"], |
Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame] | 3887 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 3888 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3889 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3890 | ) |
| 3891 | |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 3892 | xnnpack_benchmark( |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 3893 | name = "square_root_bench", |
| 3894 | srcs = ["bench/square-root.cc"], |
| 3895 | copts = xnnpack_optional_tflite_copts(), |
| 3896 | tags = ["nowin32"], |
| 3897 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 3898 | ) |
| 3899 | |
| 3900 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 3901 | name = "truncation_bench", |
| 3902 | srcs = ["bench/truncation.cc"], |
| 3903 | deps = OPERATOR_BENCHMARK_DEPS, |
| 3904 | ) |
| 3905 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3906 | ############################# End-to-end benchmarks ############################ |
| 3907 | |
| 3908 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3909 | name = "fp32_mobilenet_v1", |
| 3910 | srcs = ["models/fp32-mobilenet-v1.cc"], |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3911 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3912 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3913 | linkstatic = True, |
| 3914 | deps = [ |
| 3915 | ":XNNPACK", |
| 3916 | "@pthreadpool", |
| 3917 | ], |
| 3918 | ) |
| 3919 | |
| 3920 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3921 | name = "fp16_mobilenet_v1", |
| 3922 | srcs = ["models/fp16-mobilenet-v1.cc"], |
| 3923 | hdrs = ["models/models.h"], |
| 3924 | copts = xnnpack_std_cxxopts(), |
| 3925 | linkstatic = True, |
| 3926 | deps = [ |
| 3927 | ":XNNPACK", |
| 3928 | "@FP16", |
| 3929 | "@pthreadpool", |
| 3930 | ], |
| 3931 | ) |
| 3932 | |
| 3933 | cc_library( |
| 3934 | name = "fp32_mobilenet_v2", |
| 3935 | srcs = ["models/fp32-mobilenet-v2.cc"], |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3936 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3937 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3938 | linkstatic = True, |
| 3939 | deps = [ |
| 3940 | ":XNNPACK", |
| 3941 | "@pthreadpool", |
| 3942 | ], |
| 3943 | ) |
| 3944 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3945 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3946 | name = "fp16_mobilenet_v2", |
| 3947 | srcs = ["models/fp16-mobilenet-v2.cc"], |
| 3948 | hdrs = ["models/models.h"], |
| 3949 | copts = xnnpack_std_cxxopts(), |
| 3950 | linkstatic = True, |
| 3951 | deps = [ |
| 3952 | ":XNNPACK", |
| 3953 | "@FP16", |
| 3954 | "@pthreadpool", |
| 3955 | ], |
| 3956 | ) |
| 3957 | |
| 3958 | cc_library( |
| 3959 | name = "fp32_mobilenet_v3_large", |
| 3960 | srcs = ["models/fp32-mobilenet-v3-large.cc"], |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3961 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3962 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3963 | linkstatic = True, |
| 3964 | deps = [ |
| 3965 | ":XNNPACK", |
| 3966 | "@pthreadpool", |
| 3967 | ], |
| 3968 | ) |
| 3969 | |
| 3970 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3971 | name = "fp32_mobilenet_v3_small", |
| 3972 | srcs = ["models/fp32-mobilenet-v3-small.cc"], |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3973 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 3974 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3975 | linkstatic = True, |
| 3976 | deps = [ |
| 3977 | ":XNNPACK", |
| 3978 | "@pthreadpool", |
| 3979 | ], |
| 3980 | ) |
| 3981 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 3982 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 3983 | name = "f32_dwconv_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3984 | srcs = [ |
| 3985 | "bench/f32-dwconv-e2e.cc", |
| 3986 | "bench/end2end.h", |
| 3987 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 3988 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 3989 | ":XNNPACK", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3990 | ":fp32_mobilenet_v1", |
| 3991 | ":fp32_mobilenet_v2", |
| 3992 | ":fp32_mobilenet_v3_large", |
| 3993 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 3994 | ], |
| 3995 | ) |
| 3996 | |
| 3997 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 3998 | name = "f32_gemm_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3999 | srcs = [ |
| 4000 | "bench/f32-gemm-e2e.cc", |
| 4001 | "bench/end2end.h", |
| 4002 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4003 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4004 | ":XNNPACK", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4005 | ":fp32_mobilenet_v1", |
| 4006 | ":fp32_mobilenet_v2", |
| 4007 | ":fp32_mobilenet_v3_large", |
| 4008 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4009 | ], |
| 4010 | ) |
| 4011 | |
| 4012 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4013 | name = "end2end_bench", |
| 4014 | srcs = ["bench/end2end.cc"], |
| 4015 | deps = [ |
| 4016 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 4017 | ":bench_utils", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4018 | ":fp16_mobilenet_v1", |
| 4019 | ":fp16_mobilenet_v2", |
| 4020 | ":fp32_mobilenet_v1", |
| 4021 | ":fp32_mobilenet_v2", |
| 4022 | ":fp32_mobilenet_v3_large", |
| 4023 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4024 | "@pthreadpool", |
| 4025 | ], |
| 4026 | ) |
| 4027 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 4028 | #################### Accuracy evaluation for math functions #################### |
| 4029 | |
| 4030 | xnnpack_benchmark( |
| 4031 | name = "f32_exp_eval", |
| 4032 | srcs = [ |
| 4033 | "eval/f32-exp.cc", |
| 4034 | "src/xnnpack/AlignedAllocator.h", |
| 4035 | ] + ACCURACY_EVAL_HDRS, |
| 4036 | deps = ACCURACY_EVAL_DEPS, |
| 4037 | ) |
| 4038 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 4039 | xnnpack_benchmark( |
| 4040 | name = "f32_expminus_eval", |
| 4041 | srcs = [ |
| 4042 | "eval/f32-expminus.cc", |
| 4043 | "src/xnnpack/AlignedAllocator.h", |
| 4044 | ] + ACCURACY_EVAL_HDRS, |
| 4045 | deps = ACCURACY_EVAL_DEPS, |
| 4046 | ) |
| 4047 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 4048 | xnnpack_benchmark( |
| 4049 | name = "f32_extexp_eval", |
| 4050 | srcs = [ |
| 4051 | "eval/f32-extexp.cc", |
| 4052 | "src/xnnpack/AlignedAllocator.h", |
| 4053 | ] + ACCURACY_EVAL_HDRS, |
| 4054 | deps = ACCURACY_EVAL_DEPS, |
| 4055 | ) |
| 4056 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 4057 | xnnpack_unit_test( |
| 4058 | name = "f32_roundne_eval", |
| 4059 | srcs = [ |
| 4060 | "eval/f32-roundne.cc", |
| 4061 | "src/xnnpack/AlignedAllocator.h", |
| 4062 | "src/xnnpack/math-stubs.h", |
| 4063 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | 22eed3d | 2020-05-11 20:13:37 -0700 | [diff] [blame] | 4064 | automatic = False, |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 4065 | deps = MICROKERNEL_TEST_DEPS, |
| 4066 | ) |
| 4067 | |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 4068 | xnnpack_unit_test( |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 4069 | name = "f32_roundd_eval", |
| 4070 | srcs = [ |
| 4071 | "eval/f32-roundd.cc", |
| 4072 | "src/xnnpack/AlignedAllocator.h", |
| 4073 | "src/xnnpack/math-stubs.h", |
| 4074 | ] + MICROKERNEL_TEST_HDRS, |
| 4075 | automatic = False, |
| 4076 | deps = MICROKERNEL_TEST_DEPS, |
| 4077 | ) |
| 4078 | |
| 4079 | xnnpack_unit_test( |
| 4080 | name = "f32_roundu_eval", |
| 4081 | srcs = [ |
| 4082 | "eval/f32-roundu.cc", |
| 4083 | "src/xnnpack/AlignedAllocator.h", |
| 4084 | "src/xnnpack/math-stubs.h", |
| 4085 | ] + MICROKERNEL_TEST_HDRS, |
| 4086 | automatic = False, |
| 4087 | deps = MICROKERNEL_TEST_DEPS, |
| 4088 | ) |
| 4089 | |
| 4090 | xnnpack_unit_test( |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 4091 | name = "f32_roundz_eval", |
| 4092 | srcs = [ |
| 4093 | "eval/f32-roundz.cc", |
| 4094 | "src/xnnpack/AlignedAllocator.h", |
| 4095 | "src/xnnpack/math-stubs.h", |
| 4096 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 4097 | automatic = False, |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 4098 | deps = MICROKERNEL_TEST_DEPS, |
| 4099 | ) |
| 4100 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 4101 | xnnpack_benchmark( |
| 4102 | name = "f32_sigmoid_eval", |
| 4103 | srcs = [ |
| 4104 | "eval/f32-sigmoid.cc", |
| 4105 | "src/xnnpack/AlignedAllocator.h", |
| 4106 | ] + ACCURACY_EVAL_HDRS, |
| 4107 | deps = ACCURACY_EVAL_DEPS, |
| 4108 | ) |
| 4109 | |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 4110 | xnnpack_benchmark( |
| 4111 | name = "f32_sqrt_eval", |
| 4112 | srcs = [ |
| 4113 | "eval/f32-sqrt.cc", |
| 4114 | "src/xnnpack/AlignedAllocator.h", |
| 4115 | ] + ACCURACY_EVAL_HDRS, |
| 4116 | deps = ACCURACY_EVAL_DEPS + [ |
| 4117 | ":bench_utils", |
| 4118 | ], |
| 4119 | ) |
| 4120 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4121 | ######################### Unit tests for micro-kernels ######################### |
| 4122 | |
| 4123 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4124 | name = "f16_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4125 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4126 | "test/f16-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4127 | "test/gemm-microkernel-tester.h", |
| 4128 | "src/xnnpack/AlignedAllocator.h", |
| 4129 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4130 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4131 | ) |
| 4132 | |
| 4133 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4134 | name = "f16_spmm_minmax_test", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4135 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4136 | "test/f16-spmm-minmax.cc", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4137 | "test/spmm-microkernel-tester.h", |
| 4138 | "src/xnnpack/AlignedAllocator.h", |
| 4139 | ] + MICROKERNEL_TEST_HDRS, |
| 4140 | deps = MICROKERNEL_TEST_DEPS, |
| 4141 | ) |
| 4142 | |
| 4143 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4144 | name = "f32_argmaxpool_test", |
| 4145 | srcs = [ |
| 4146 | "test/f32-argmaxpool.cc", |
| 4147 | "test/argmaxpool-microkernel-tester.h", |
| 4148 | "src/xnnpack/AlignedAllocator.h", |
| 4149 | ] + MICROKERNEL_TEST_HDRS, |
| 4150 | deps = MICROKERNEL_TEST_DEPS, |
| 4151 | ) |
| 4152 | |
| 4153 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4154 | name = "f32_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4155 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4156 | "test/f32-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4157 | "test/avgpool-microkernel-tester.h", |
| 4158 | "src/xnnpack/AlignedAllocator.h", |
| 4159 | ] + MICROKERNEL_TEST_HDRS, |
| 4160 | deps = MICROKERNEL_TEST_DEPS, |
| 4161 | ) |
| 4162 | |
| 4163 | xnnpack_unit_test( |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 4164 | name = "f32_ibilinear_test", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 4165 | srcs = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 4166 | "test/f32-ibilinear.cc", |
| 4167 | "test/ibilinear-microkernel-tester.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 4168 | "src/xnnpack/AlignedAllocator.h", |
| 4169 | ] + MICROKERNEL_TEST_HDRS, |
| 4170 | deps = MICROKERNEL_TEST_DEPS, |
| 4171 | ) |
| 4172 | |
| 4173 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 4174 | name = "f16_clamp_test", |
| 4175 | srcs = [ |
| 4176 | "test/f16-clamp.cc", |
| 4177 | "test/clamp-microkernel-tester.h", |
| 4178 | ] + MICROKERNEL_TEST_HDRS, |
| 4179 | deps = MICROKERNEL_TEST_DEPS, |
| 4180 | ) |
| 4181 | |
| 4182 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4183 | name = "f32_clamp_test", |
| 4184 | srcs = [ |
| 4185 | "test/f32-clamp.cc", |
| 4186 | "test/clamp-microkernel-tester.h", |
| 4187 | ] + MICROKERNEL_TEST_HDRS, |
| 4188 | deps = MICROKERNEL_TEST_DEPS, |
| 4189 | ) |
| 4190 | |
| 4191 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 4192 | name = "f32_igemm_test", |
| 4193 | srcs = [ |
| 4194 | "test/f32-igemm.cc", |
| 4195 | "test/gemm-microkernel-tester.h", |
| 4196 | "src/xnnpack/AlignedAllocator.h", |
| 4197 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4198 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 4199 | ) |
| 4200 | |
| 4201 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 4202 | name = "f32_igemm_relu_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4203 | srcs = [ |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 4204 | "test/f32-igemm-relu.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4205 | "test/gemm-microkernel-tester.h", |
| 4206 | "src/xnnpack/AlignedAllocator.h", |
| 4207 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4208 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4209 | ) |
| 4210 | |
| 4211 | xnnpack_unit_test( |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 4212 | name = "f32_igemm_minmax_test", |
| 4213 | srcs = [ |
| 4214 | "test/f32-igemm-minmax.cc", |
| 4215 | "test/gemm-microkernel-tester.h", |
| 4216 | "src/xnnpack/AlignedAllocator.h", |
| 4217 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4218 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 4219 | ) |
| 4220 | |
| 4221 | xnnpack_unit_test( |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 4222 | name = "f16_igemm_minmax_test", |
| 4223 | srcs = [ |
| 4224 | "test/f16-igemm-minmax.cc", |
| 4225 | "test/gemm-microkernel-tester.h", |
| 4226 | "src/xnnpack/AlignedAllocator.h", |
| 4227 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4228 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 4229 | ) |
| 4230 | |
| 4231 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4232 | name = "f32_conv_hwc_test", |
| 4233 | srcs = [ |
| 4234 | "test/f32-conv-hwc.cc", |
| 4235 | "test/conv-hwc-microkernel-tester.h", |
| 4236 | "src/xnnpack/AlignedAllocator.h", |
| 4237 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4238 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4239 | ) |
| 4240 | |
| 4241 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4242 | name = "f32_conv_hwc2chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4243 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4244 | "test/f32-conv-hwc2chw.cc", |
| 4245 | "test/conv-hwc2chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4246 | "src/xnnpack/AlignedAllocator.h", |
| 4247 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4248 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4249 | ) |
| 4250 | |
| 4251 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 4252 | name = "f32_dwconv_test", |
| 4253 | srcs = [ |
| 4254 | "test/f32-dwconv.cc", |
| 4255 | "test/dwconv-microkernel-tester.h", |
| 4256 | "src/xnnpack/AlignedAllocator.h", |
| 4257 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4258 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 4259 | ) |
| 4260 | |
| 4261 | xnnpack_unit_test( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4262 | name = "f16_dwconv_minmax_test", |
| 4263 | srcs = [ |
| 4264 | "test/f16-dwconv-minmax.cc", |
| 4265 | "test/dwconv-microkernel-tester.h", |
| 4266 | "src/xnnpack/AlignedAllocator.h", |
| 4267 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4268 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4269 | ) |
| 4270 | |
| 4271 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4272 | name = "f32_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4273 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4274 | "test/f32-dwconv-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4275 | "test/dwconv-microkernel-tester.h", |
| 4276 | "src/xnnpack/AlignedAllocator.h", |
| 4277 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4278 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4279 | ) |
| 4280 | |
| 4281 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4282 | name = "f32_dwconv_chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4283 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4284 | "test/f32-dwconv-chw.cc", |
| 4285 | "test/dwconv-chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4286 | "src/xnnpack/AlignedAllocator.h", |
| 4287 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4288 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4289 | ) |
| 4290 | |
| 4291 | xnnpack_unit_test( |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 4292 | name = "f16_gavgpool_minmax_test", |
| 4293 | srcs = [ |
| 4294 | "test/f16-gavgpool-minmax.cc", |
| 4295 | "test/gavgpool-microkernel-tester.h", |
| 4296 | "src/xnnpack/AlignedAllocator.h", |
| 4297 | ] + MICROKERNEL_TEST_HDRS, |
| 4298 | deps = MICROKERNEL_TEST_DEPS, |
| 4299 | ) |
| 4300 | |
| 4301 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4302 | name = "f32_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4303 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4304 | "test/f32-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4305 | "test/gavgpool-microkernel-tester.h", |
| 4306 | "src/xnnpack/AlignedAllocator.h", |
| 4307 | ] + MICROKERNEL_TEST_HDRS, |
| 4308 | deps = MICROKERNEL_TEST_DEPS, |
| 4309 | ) |
| 4310 | |
| 4311 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4312 | name = "f32_gavgpool_cw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4313 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4314 | "test/f32-gavgpool-cw.cc", |
| 4315 | "test/gavgpool-cw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4316 | "src/xnnpack/AlignedAllocator.h", |
| 4317 | ] + MICROKERNEL_TEST_HDRS, |
| 4318 | deps = MICROKERNEL_TEST_DEPS, |
| 4319 | ) |
| 4320 | |
| 4321 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 4322 | name = "f32_gemm_test", |
| 4323 | srcs = [ |
| 4324 | "test/f32-gemm.cc", |
| 4325 | "test/gemm-microkernel-tester.h", |
| 4326 | "src/xnnpack/AlignedAllocator.h", |
| 4327 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4328 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 4329 | ) |
| 4330 | |
| 4331 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 4332 | name = "f32_gemm_relu_test", |
| 4333 | srcs = [ |
| 4334 | "test/f32-gemm-relu.cc", |
| 4335 | "test/gemm-microkernel-tester.h", |
| 4336 | "src/xnnpack/AlignedAllocator.h", |
| 4337 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4338 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 4339 | ) |
| 4340 | |
| 4341 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4342 | name = "f32_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4343 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4344 | "test/f32-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4345 | "test/gemm-microkernel-tester.h", |
| 4346 | "src/xnnpack/AlignedAllocator.h", |
| 4347 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4348 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4349 | ) |
| 4350 | |
| 4351 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4352 | name = "f32_gemminc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4353 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4354 | "test/f32-gemminc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4355 | "test/gemm-microkernel-tester.h", |
| 4356 | "src/xnnpack/AlignedAllocator.h", |
| 4357 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4358 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4359 | ) |
| 4360 | |
| 4361 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 4362 | name = "f16_hswish_test", |
| 4363 | srcs = [ |
| 4364 | "test/f16-hswish.cc", |
| 4365 | "test/hswish-microkernel-tester.h", |
| 4366 | ] + MICROKERNEL_TEST_HDRS, |
| 4367 | deps = MICROKERNEL_TEST_DEPS, |
| 4368 | ) |
| 4369 | |
| 4370 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4371 | name = "f32_hswish_test", |
| 4372 | srcs = [ |
| 4373 | "test/f32-hswish.cc", |
| 4374 | "test/hswish-microkernel-tester.h", |
| 4375 | ] + MICROKERNEL_TEST_HDRS, |
| 4376 | deps = MICROKERNEL_TEST_DEPS, |
| 4377 | ) |
| 4378 | |
| 4379 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4380 | name = "f32_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4381 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4382 | "test/f32-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4383 | "test/maxpool-microkernel-tester.h", |
| 4384 | ] + MICROKERNEL_TEST_HDRS, |
| 4385 | deps = MICROKERNEL_TEST_DEPS, |
| 4386 | ) |
| 4387 | |
| 4388 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4389 | name = "f32_pavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4390 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4391 | "test/f32-pavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4392 | "test/avgpool-microkernel-tester.h", |
| 4393 | "src/xnnpack/AlignedAllocator.h", |
| 4394 | ] + MICROKERNEL_TEST_HDRS, |
| 4395 | deps = MICROKERNEL_TEST_DEPS, |
| 4396 | ) |
| 4397 | |
| 4398 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4399 | name = "f32_ppmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4400 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 4401 | "test/f32-ppmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4402 | "test/gemm-microkernel-tester.h", |
| 4403 | "src/xnnpack/AlignedAllocator.h", |
| 4404 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4405 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4406 | ) |
| 4407 | |
| 4408 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 4409 | name = "f16_prelu_test", |
| 4410 | srcs = [ |
| 4411 | "test/f16-prelu.cc", |
| 4412 | "test/prelu-microkernel-tester.h", |
| 4413 | "src/xnnpack/AlignedAllocator.h", |
| 4414 | ] + MICROKERNEL_TEST_HDRS, |
| 4415 | deps = MICROKERNEL_TEST_DEPS, |
| 4416 | ) |
| 4417 | |
| 4418 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4419 | name = "f32_prelu_test", |
| 4420 | srcs = [ |
| 4421 | "test/f32-prelu.cc", |
| 4422 | "test/prelu-microkernel-tester.h", |
| 4423 | "src/xnnpack/AlignedAllocator.h", |
| 4424 | ] + MICROKERNEL_TEST_HDRS, |
| 4425 | deps = MICROKERNEL_TEST_DEPS, |
| 4426 | ) |
| 4427 | |
| 4428 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4429 | name = "f32_raddexpminusmax_test", |
| 4430 | srcs = [ |
| 4431 | "test/f32-raddexpminusmax.cc", |
| 4432 | "test/raddexpminusmax-microkernel-tester.h", |
| 4433 | ] + MICROKERNEL_TEST_HDRS, |
| 4434 | deps = MICROKERNEL_TEST_DEPS, |
| 4435 | ) |
| 4436 | |
| 4437 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 4438 | name = "f32_raddextexp_test", |
| 4439 | srcs = [ |
| 4440 | "test/f32-raddextexp.cc", |
| 4441 | "test/raddextexp-microkernel-tester.h", |
| 4442 | ] + MICROKERNEL_TEST_HDRS, |
| 4443 | deps = MICROKERNEL_TEST_DEPS, |
| 4444 | ) |
| 4445 | |
| 4446 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4447 | name = "f32_raddstoreexpminusmax_test", |
| 4448 | srcs = [ |
| 4449 | "test/f32-raddstoreexpminusmax.cc", |
| 4450 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 4451 | ] + MICROKERNEL_TEST_HDRS, |
| 4452 | deps = MICROKERNEL_TEST_DEPS, |
| 4453 | ) |
| 4454 | |
| 4455 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4456 | name = "f32_rmax_test", |
| 4457 | srcs = [ |
| 4458 | "test/f32-rmax.cc", |
| 4459 | "test/rmax-microkernel-tester.h", |
| 4460 | ] + MICROKERNEL_TEST_HDRS, |
| 4461 | deps = MICROKERNEL_TEST_DEPS, |
| 4462 | ) |
| 4463 | |
| 4464 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 4465 | name = "f32_sigmoid_test", |
| 4466 | srcs = [ |
| 4467 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4468 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 4469 | ] + MICROKERNEL_TEST_HDRS, |
| 4470 | deps = MICROKERNEL_TEST_DEPS, |
| 4471 | ) |
| 4472 | |
| 4473 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4474 | name = "f32_spmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4475 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4476 | "test/f32-spmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4477 | "test/spmm-microkernel-tester.h", |
| 4478 | "src/xnnpack/AlignedAllocator.h", |
| 4479 | ] + MICROKERNEL_TEST_HDRS, |
| 4480 | deps = MICROKERNEL_TEST_DEPS, |
| 4481 | ) |
| 4482 | |
| 4483 | xnnpack_unit_test( |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 4484 | name = "f16_vadd_minmax_test", |
| 4485 | srcs = [ |
| 4486 | "test/f16-vadd-minmax.cc", |
| 4487 | "test/vbinary-microkernel-tester.h", |
| 4488 | ] + MICROKERNEL_TEST_HDRS, |
| 4489 | deps = MICROKERNEL_TEST_DEPS, |
| 4490 | ) |
| 4491 | |
| 4492 | xnnpack_unit_test( |
| 4493 | name = "f16_vaddc_minmax_test", |
| 4494 | srcs = [ |
| 4495 | "test/f16-vaddc-minmax.cc", |
| 4496 | "test/vbinaryc-microkernel-tester.h", |
| 4497 | ] + MICROKERNEL_TEST_HDRS, |
| 4498 | deps = MICROKERNEL_TEST_DEPS, |
| 4499 | ) |
| 4500 | |
| 4501 | xnnpack_unit_test( |
| 4502 | name = "f16_vdiv_minmax_test", |
| 4503 | srcs = [ |
| 4504 | "test/f16-vdiv-minmax.cc", |
| 4505 | "test/vbinary-microkernel-tester.h", |
| 4506 | ] + MICROKERNEL_TEST_HDRS, |
| 4507 | deps = MICROKERNEL_TEST_DEPS, |
| 4508 | ) |
| 4509 | |
| 4510 | xnnpack_unit_test( |
| 4511 | name = "f16_vdivc_minmax_test", |
| 4512 | srcs = [ |
| 4513 | "test/f16-vdivc-minmax.cc", |
| 4514 | "test/vbinaryc-microkernel-tester.h", |
| 4515 | ] + MICROKERNEL_TEST_HDRS, |
| 4516 | deps = MICROKERNEL_TEST_DEPS, |
| 4517 | ) |
| 4518 | |
| 4519 | xnnpack_unit_test( |
| 4520 | name = "f16_vrdivc_minmax_test", |
| 4521 | srcs = [ |
| 4522 | "test/f16-vrdivc-minmax.cc", |
| 4523 | "test/vbinaryc-microkernel-tester.h", |
| 4524 | ] + MICROKERNEL_TEST_HDRS, |
| 4525 | deps = MICROKERNEL_TEST_DEPS, |
| 4526 | ) |
| 4527 | |
| 4528 | xnnpack_unit_test( |
| 4529 | name = "f16_vmax_test", |
| 4530 | srcs = [ |
| 4531 | "test/f16-vmax.cc", |
| 4532 | "test/vbinary-microkernel-tester.h", |
| 4533 | ] + MICROKERNEL_TEST_HDRS, |
| 4534 | deps = MICROKERNEL_TEST_DEPS, |
| 4535 | ) |
| 4536 | |
| 4537 | xnnpack_unit_test( |
| 4538 | name = "f16_vmaxc_test", |
| 4539 | srcs = [ |
| 4540 | "test/f16-vmaxc.cc", |
| 4541 | "test/vbinaryc-microkernel-tester.h", |
| 4542 | ] + MICROKERNEL_TEST_HDRS, |
| 4543 | deps = MICROKERNEL_TEST_DEPS, |
| 4544 | ) |
| 4545 | |
| 4546 | xnnpack_unit_test( |
| 4547 | name = "f16_vmin_test", |
| 4548 | srcs = [ |
| 4549 | "test/f16-vmin.cc", |
| 4550 | "test/vbinary-microkernel-tester.h", |
| 4551 | ] + MICROKERNEL_TEST_HDRS, |
| 4552 | deps = MICROKERNEL_TEST_DEPS, |
| 4553 | ) |
| 4554 | |
| 4555 | xnnpack_unit_test( |
| 4556 | name = "f16_vminc_test", |
| 4557 | srcs = [ |
| 4558 | "test/f16-vminc.cc", |
| 4559 | "test/vbinaryc-microkernel-tester.h", |
| 4560 | ] + MICROKERNEL_TEST_HDRS, |
| 4561 | deps = MICROKERNEL_TEST_DEPS, |
| 4562 | ) |
| 4563 | |
| 4564 | xnnpack_unit_test( |
| 4565 | name = "f16_vmul_minmax_test", |
| 4566 | srcs = [ |
| 4567 | "test/f16-vmul-minmax.cc", |
| 4568 | "test/vbinary-microkernel-tester.h", |
| 4569 | ] + MICROKERNEL_TEST_HDRS, |
| 4570 | deps = MICROKERNEL_TEST_DEPS, |
| 4571 | ) |
| 4572 | |
| 4573 | xnnpack_unit_test( |
| 4574 | name = "f16_vmulc_minmax_test", |
| 4575 | srcs = [ |
| 4576 | "test/f16-vmulc-minmax.cc", |
| 4577 | "test/vbinaryc-microkernel-tester.h", |
| 4578 | ] + MICROKERNEL_TEST_HDRS, |
| 4579 | deps = MICROKERNEL_TEST_DEPS, |
| 4580 | ) |
| 4581 | |
| 4582 | xnnpack_unit_test( |
| 4583 | name = "f16_vsub_minmax_test", |
| 4584 | srcs = [ |
| 4585 | "test/f16-vsub-minmax.cc", |
| 4586 | "test/vbinary-microkernel-tester.h", |
| 4587 | ] + MICROKERNEL_TEST_HDRS, |
| 4588 | deps = MICROKERNEL_TEST_DEPS, |
| 4589 | ) |
| 4590 | |
| 4591 | xnnpack_unit_test( |
| 4592 | name = "f16_vsubc_minmax_test", |
| 4593 | srcs = [ |
| 4594 | "test/f16-vsubc-minmax.cc", |
| 4595 | "test/vbinaryc-microkernel-tester.h", |
| 4596 | ] + MICROKERNEL_TEST_HDRS, |
| 4597 | deps = MICROKERNEL_TEST_DEPS, |
| 4598 | ) |
| 4599 | |
| 4600 | xnnpack_unit_test( |
| 4601 | name = "f16_vrsubc_minmax_test", |
| 4602 | srcs = [ |
| 4603 | "test/f16-vrsubc-minmax.cc", |
| 4604 | "test/vbinaryc-microkernel-tester.h", |
| 4605 | ] + MICROKERNEL_TEST_HDRS, |
| 4606 | deps = MICROKERNEL_TEST_DEPS, |
| 4607 | ) |
| 4608 | |
| 4609 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4610 | name = "f32_vabs_test", |
| 4611 | srcs = [ |
| 4612 | "test/f32-vabs.cc", |
| 4613 | "test/vunary-microkernel-tester.h", |
| 4614 | ] + MICROKERNEL_TEST_HDRS, |
| 4615 | deps = MICROKERNEL_TEST_DEPS, |
| 4616 | ) |
| 4617 | |
| 4618 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4619 | name = "f32_vadd_test", |
| 4620 | srcs = [ |
| 4621 | "test/f32-vadd.cc", |
| 4622 | "test/vbinary-microkernel-tester.h", |
| 4623 | ] + MICROKERNEL_TEST_HDRS, |
| 4624 | deps = MICROKERNEL_TEST_DEPS, |
| 4625 | ) |
| 4626 | |
| 4627 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4628 | name = "f32_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4629 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4630 | "test/f32-vadd-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4631 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4632 | ] + MICROKERNEL_TEST_HDRS, |
| 4633 | deps = MICROKERNEL_TEST_DEPS, |
| 4634 | ) |
| 4635 | |
| 4636 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4637 | name = "f32_vaddc_test", |
| 4638 | srcs = [ |
| 4639 | "test/f32-vaddc.cc", |
| 4640 | "test/vbinaryc-microkernel-tester.h", |
| 4641 | ] + MICROKERNEL_TEST_HDRS, |
| 4642 | deps = MICROKERNEL_TEST_DEPS, |
| 4643 | ) |
| 4644 | |
| 4645 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4646 | name = "f32_vaddc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4647 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4648 | "test/f32-vaddc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4649 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4650 | ] + MICROKERNEL_TEST_HDRS, |
| 4651 | deps = MICROKERNEL_TEST_DEPS, |
| 4652 | ) |
| 4653 | |
| 4654 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4655 | name = "f32_vdiv_test", |
| 4656 | srcs = [ |
| 4657 | "test/f32-vdiv.cc", |
| 4658 | "test/vbinary-microkernel-tester.h", |
| 4659 | ] + MICROKERNEL_TEST_HDRS, |
| 4660 | deps = MICROKERNEL_TEST_DEPS, |
| 4661 | ) |
| 4662 | |
| 4663 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4664 | name = "f32_vdiv_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4665 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4666 | "test/f32-vdiv-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4667 | "test/vbinary-microkernel-tester.h", |
| 4668 | ] + MICROKERNEL_TEST_HDRS, |
| 4669 | deps = MICROKERNEL_TEST_DEPS, |
| 4670 | ) |
| 4671 | |
| 4672 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4673 | name = "f32_vdivc_test", |
| 4674 | srcs = [ |
| 4675 | "test/f32-vdivc.cc", |
| 4676 | "test/vbinaryc-microkernel-tester.h", |
| 4677 | ] + MICROKERNEL_TEST_HDRS, |
| 4678 | deps = MICROKERNEL_TEST_DEPS, |
| 4679 | ) |
| 4680 | |
| 4681 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4682 | name = "f32_vdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4683 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4684 | "test/f32-vdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4685 | "test/vbinaryc-microkernel-tester.h", |
| 4686 | ] + MICROKERNEL_TEST_HDRS, |
| 4687 | deps = MICROKERNEL_TEST_DEPS, |
| 4688 | ) |
| 4689 | |
| 4690 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4691 | name = "f32_vrdivc_test", |
| 4692 | srcs = [ |
| 4693 | "test/f32-vrdivc.cc", |
| 4694 | "test/vbinaryc-microkernel-tester.h", |
| 4695 | ] + MICROKERNEL_TEST_HDRS, |
| 4696 | deps = MICROKERNEL_TEST_DEPS, |
| 4697 | ) |
| 4698 | |
| 4699 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4700 | name = "f32_vrdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4701 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4702 | "test/f32-vrdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 4703 | "test/vbinaryc-microkernel-tester.h", |
| 4704 | ] + MICROKERNEL_TEST_HDRS, |
| 4705 | deps = MICROKERNEL_TEST_DEPS, |
| 4706 | ) |
| 4707 | |
| 4708 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 4709 | name = "f32_vmax_test", |
| 4710 | srcs = [ |
| 4711 | "test/f32-vmax.cc", |
| 4712 | "test/vbinary-microkernel-tester.h", |
| 4713 | ] + MICROKERNEL_TEST_HDRS, |
| 4714 | deps = MICROKERNEL_TEST_DEPS, |
| 4715 | ) |
| 4716 | |
| 4717 | xnnpack_unit_test( |
| 4718 | name = "f32_vmaxc_test", |
| 4719 | srcs = [ |
| 4720 | "test/f32-vmaxc.cc", |
| 4721 | "test/vbinaryc-microkernel-tester.h", |
| 4722 | ] + MICROKERNEL_TEST_HDRS, |
| 4723 | deps = MICROKERNEL_TEST_DEPS, |
| 4724 | ) |
| 4725 | |
| 4726 | xnnpack_unit_test( |
| 4727 | name = "f32_vmin_test", |
| 4728 | srcs = [ |
| 4729 | "test/f32-vmin.cc", |
| 4730 | "test/vbinary-microkernel-tester.h", |
| 4731 | ] + MICROKERNEL_TEST_HDRS, |
| 4732 | deps = MICROKERNEL_TEST_DEPS, |
| 4733 | ) |
| 4734 | |
| 4735 | xnnpack_unit_test( |
| 4736 | name = "f32_vminc_test", |
| 4737 | srcs = [ |
| 4738 | "test/f32-vminc.cc", |
| 4739 | "test/vbinaryc-microkernel-tester.h", |
| 4740 | ] + MICROKERNEL_TEST_HDRS, |
| 4741 | deps = MICROKERNEL_TEST_DEPS, |
| 4742 | ) |
| 4743 | |
| 4744 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4745 | name = "f32_vmul_test", |
| 4746 | srcs = [ |
| 4747 | "test/f32-vmul.cc", |
| 4748 | "test/vbinary-microkernel-tester.h", |
| 4749 | ] + MICROKERNEL_TEST_HDRS, |
| 4750 | deps = MICROKERNEL_TEST_DEPS, |
| 4751 | ) |
| 4752 | |
| 4753 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4754 | name = "f32_vmul_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4755 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4756 | "test/f32-vmul-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4757 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4758 | ] + MICROKERNEL_TEST_HDRS, |
| 4759 | deps = MICROKERNEL_TEST_DEPS, |
| 4760 | ) |
| 4761 | |
| 4762 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4763 | name = "f32_vmulc_test", |
| 4764 | srcs = [ |
| 4765 | "test/f32-vmulc.cc", |
| 4766 | "test/vbinaryc-microkernel-tester.h", |
| 4767 | ] + MICROKERNEL_TEST_HDRS, |
| 4768 | deps = MICROKERNEL_TEST_DEPS, |
| 4769 | ) |
| 4770 | |
| 4771 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4772 | name = "f32_vmulc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4773 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4774 | "test/f32-vmulc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4775 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4776 | ] + MICROKERNEL_TEST_HDRS, |
| 4777 | deps = MICROKERNEL_TEST_DEPS, |
| 4778 | ) |
| 4779 | |
| 4780 | xnnpack_unit_test( |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 4781 | name = "f16_vmulcaddc_minmax_test", |
| 4782 | srcs = [ |
| 4783 | "test/f16-vmulcaddc-minmax.cc", |
| 4784 | "test/vmulcaddc-microkernel-tester.h", |
| 4785 | "src/xnnpack/AlignedAllocator.h", |
| 4786 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4787 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 4788 | ) |
| 4789 | |
| 4790 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4791 | name = "f32_vmulcaddc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4792 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4793 | "test/f32-vmulcaddc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4794 | "test/vmulcaddc-microkernel-tester.h", |
| 4795 | "src/xnnpack/AlignedAllocator.h", |
| 4796 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4797 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4798 | ) |
| 4799 | |
| 4800 | xnnpack_unit_test( |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 4801 | name = "f32_vlrelu_test", |
| 4802 | srcs = [ |
| 4803 | "test/f32-vlrelu.cc", |
| 4804 | "test/vunary-microkernel-tester.h", |
| 4805 | ] + MICROKERNEL_TEST_HDRS, |
| 4806 | deps = MICROKERNEL_TEST_DEPS, |
| 4807 | ) |
| 4808 | |
| 4809 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4810 | name = "f32_vneg_test", |
| 4811 | srcs = [ |
| 4812 | "test/f32-vneg.cc", |
| 4813 | "test/vunary-microkernel-tester.h", |
| 4814 | ] + MICROKERNEL_TEST_HDRS, |
| 4815 | deps = MICROKERNEL_TEST_DEPS, |
| 4816 | ) |
| 4817 | |
| 4818 | xnnpack_unit_test( |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 4819 | name = "f32_vrndne_test", |
| 4820 | srcs = [ |
| 4821 | "test/f32-vrndne.cc", |
| 4822 | "test/vunary-microkernel-tester.h", |
| 4823 | ] + MICROKERNEL_TEST_HDRS, |
| 4824 | deps = MICROKERNEL_TEST_DEPS, |
| 4825 | ) |
| 4826 | |
| 4827 | xnnpack_unit_test( |
| 4828 | name = "f32_vrndz_test", |
| 4829 | srcs = [ |
| 4830 | "test/f32-vrndz.cc", |
| 4831 | "test/vunary-microkernel-tester.h", |
| 4832 | ] + MICROKERNEL_TEST_HDRS, |
| 4833 | deps = MICROKERNEL_TEST_DEPS, |
| 4834 | ) |
| 4835 | |
| 4836 | xnnpack_unit_test( |
| 4837 | name = "f32_vrndu_test", |
| 4838 | srcs = [ |
| 4839 | "test/f32-vrndu.cc", |
| 4840 | "test/vunary-microkernel-tester.h", |
| 4841 | ] + MICROKERNEL_TEST_HDRS, |
| 4842 | deps = MICROKERNEL_TEST_DEPS, |
| 4843 | ) |
| 4844 | |
| 4845 | xnnpack_unit_test( |
| 4846 | name = "f32_vrndd_test", |
| 4847 | srcs = [ |
| 4848 | "test/f32-vrndd.cc", |
| 4849 | "test/vunary-microkernel-tester.h", |
| 4850 | ] + MICROKERNEL_TEST_HDRS, |
| 4851 | deps = MICROKERNEL_TEST_DEPS, |
| 4852 | ) |
| 4853 | |
| 4854 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 4855 | name = "f32_vscale_test", |
| 4856 | srcs = [ |
| 4857 | "test/f32-vscale.cc", |
| 4858 | "test/vscale-microkernel-tester.h", |
| 4859 | ] + MICROKERNEL_TEST_HDRS, |
| 4860 | deps = MICROKERNEL_TEST_DEPS, |
| 4861 | ) |
| 4862 | |
| 4863 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4864 | name = "f32_vscaleexpminusmax_test", |
| 4865 | srcs = [ |
| 4866 | "test/f32-vscaleexpminusmax.cc", |
| 4867 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 4868 | ] + MICROKERNEL_TEST_HDRS, |
| 4869 | deps = MICROKERNEL_TEST_DEPS, |
| 4870 | ) |
| 4871 | |
| 4872 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 4873 | name = "f32_vscaleextexp_test", |
| 4874 | srcs = [ |
| 4875 | "test/f32-vscaleextexp.cc", |
| 4876 | "test/vscaleextexp-microkernel-tester.h", |
| 4877 | ] + MICROKERNEL_TEST_HDRS, |
| 4878 | deps = MICROKERNEL_TEST_DEPS, |
| 4879 | ) |
| 4880 | |
| 4881 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 4882 | name = "f32_vsqr_test", |
| 4883 | srcs = [ |
| 4884 | "test/f32-vsqr.cc", |
| 4885 | "test/vunary-microkernel-tester.h", |
| 4886 | ] + MICROKERNEL_TEST_HDRS, |
| 4887 | deps = MICROKERNEL_TEST_DEPS, |
| 4888 | ) |
| 4889 | |
| 4890 | xnnpack_unit_test( |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 4891 | name = "f32_vsqrdiff_test", |
| 4892 | srcs = [ |
| 4893 | "test/f32-vsqrdiff.cc", |
| 4894 | "test/vbinary-microkernel-tester.h", |
| 4895 | ] + MICROKERNEL_TEST_HDRS, |
| 4896 | deps = MICROKERNEL_TEST_DEPS, |
| 4897 | ) |
| 4898 | |
| 4899 | xnnpack_unit_test( |
| 4900 | name = "f32_vsqrdiffc_test", |
| 4901 | srcs = [ |
| 4902 | "test/f32-vsqrdiffc.cc", |
| 4903 | "test/vbinaryc-microkernel-tester.h", |
| 4904 | ] + MICROKERNEL_TEST_HDRS, |
| 4905 | deps = MICROKERNEL_TEST_DEPS, |
| 4906 | ) |
| 4907 | |
| 4908 | xnnpack_unit_test( |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 4909 | name = "f32_vsqrt_test", |
| 4910 | srcs = [ |
| 4911 | "test/f32-vsqrt.cc", |
| 4912 | "test/vunary-microkernel-tester.h", |
| 4913 | ] + MICROKERNEL_TEST_HDRS, |
| 4914 | deps = MICROKERNEL_TEST_DEPS, |
| 4915 | ) |
| 4916 | |
| 4917 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4918 | name = "f32_vsub_test", |
| 4919 | srcs = [ |
| 4920 | "test/f32-vsub.cc", |
| 4921 | "test/vbinary-microkernel-tester.h", |
| 4922 | ] + MICROKERNEL_TEST_HDRS, |
| 4923 | deps = MICROKERNEL_TEST_DEPS, |
| 4924 | ) |
| 4925 | |
| 4926 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4927 | name = "f32_vsub_minmax_test", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4928 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4929 | "test/f32-vsub-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4930 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4931 | ] + MICROKERNEL_TEST_HDRS, |
| 4932 | deps = MICROKERNEL_TEST_DEPS, |
| 4933 | ) |
| 4934 | |
| 4935 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4936 | name = "f32_vsubc_test", |
| 4937 | srcs = [ |
| 4938 | "test/f32-vsubc.cc", |
| 4939 | "test/vbinaryc-microkernel-tester.h", |
| 4940 | ] + MICROKERNEL_TEST_HDRS, |
| 4941 | deps = MICROKERNEL_TEST_DEPS, |
| 4942 | ) |
| 4943 | |
| 4944 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4945 | name = "f32_vsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4946 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4947 | "test/f32-vsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4948 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4949 | ] + MICROKERNEL_TEST_HDRS, |
| 4950 | deps = MICROKERNEL_TEST_DEPS, |
| 4951 | ) |
| 4952 | |
| 4953 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 4954 | name = "f32_vrsubc_test", |
| 4955 | srcs = [ |
| 4956 | "test/f32-vrsubc.cc", |
| 4957 | "test/vbinaryc-microkernel-tester.h", |
| 4958 | ] + MICROKERNEL_TEST_HDRS, |
| 4959 | deps = MICROKERNEL_TEST_DEPS, |
| 4960 | ) |
| 4961 | |
| 4962 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4963 | name = "f32_vrsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 4964 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 4965 | "test/f32-vrsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 4966 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 4967 | ] + MICROKERNEL_TEST_HDRS, |
| 4968 | deps = MICROKERNEL_TEST_DEPS, |
| 4969 | ) |
| 4970 | |
| 4971 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4972 | name = "q8_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4973 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4974 | "test/q8-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4975 | "test/avgpool-microkernel-tester.h", |
| 4976 | "src/xnnpack/AlignedAllocator.h", |
| 4977 | ] + MICROKERNEL_TEST_HDRS, |
| 4978 | deps = MICROKERNEL_TEST_DEPS, |
| 4979 | ) |
| 4980 | |
| 4981 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4982 | name = "q8_igemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4983 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4984 | "test/q8-igemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4985 | "test/gemm-microkernel-tester.h", |
| 4986 | "src/xnnpack/AlignedAllocator.h", |
| 4987 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4988 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4989 | ) |
| 4990 | |
| 4991 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4992 | name = "q8_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4993 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4994 | "test/q8-dwconv-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4995 | "test/dwconv-microkernel-tester.h", |
| 4996 | "src/xnnpack/AlignedAllocator.h", |
| 4997 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4998 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4999 | ) |
| 5000 | |
| 5001 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5002 | name = "q8_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5003 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5004 | "test/q8-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5005 | "test/gavgpool-microkernel-tester.h", |
| 5006 | "src/xnnpack/AlignedAllocator.h", |
| 5007 | ] + MICROKERNEL_TEST_HDRS, |
| 5008 | deps = MICROKERNEL_TEST_DEPS, |
| 5009 | ) |
| 5010 | |
| 5011 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 5012 | name = "q8_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5013 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 5014 | "test/q8-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5015 | "test/gemm-microkernel-tester.h", |
| 5016 | "src/xnnpack/AlignedAllocator.h", |
| 5017 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5018 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5019 | ) |
| 5020 | |
| 5021 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5022 | name = "q8_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5023 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5024 | "test/q8-vadd-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5025 | "test/vadd-microkernel-tester.h", |
| 5026 | ] + MICROKERNEL_TEST_HDRS, |
| 5027 | deps = MICROKERNEL_TEST_DEPS, |
| 5028 | ) |
| 5029 | |
| 5030 | xnnpack_unit_test( |
| 5031 | name = "u8_clamp_test", |
| 5032 | srcs = [ |
| 5033 | "test/u8-clamp.cc", |
| 5034 | "test/clamp-microkernel-tester.h", |
| 5035 | ] + MICROKERNEL_TEST_HDRS, |
| 5036 | deps = MICROKERNEL_TEST_DEPS, |
| 5037 | ) |
| 5038 | |
| 5039 | xnnpack_unit_test( |
| 5040 | name = "u8_lut32norm_test", |
| 5041 | srcs = [ |
| 5042 | "test/u8-lut32norm.cc", |
| 5043 | "test/lut-norm-microkernel-tester.h", |
| 5044 | ] + MICROKERNEL_TEST_HDRS, |
| 5045 | deps = MICROKERNEL_TEST_DEPS, |
| 5046 | ) |
| 5047 | |
| 5048 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5049 | name = "u8_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5050 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5051 | "test/u8-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5052 | "test/maxpool-microkernel-tester.h", |
| 5053 | ] + MICROKERNEL_TEST_HDRS, |
| 5054 | deps = MICROKERNEL_TEST_DEPS, |
| 5055 | ) |
| 5056 | |
| 5057 | xnnpack_unit_test( |
| 5058 | name = "u8_rmax_test", |
| 5059 | srcs = [ |
| 5060 | "test/u8-rmax.cc", |
| 5061 | "test/rmax-microkernel-tester.h", |
| 5062 | ] + MICROKERNEL_TEST_HDRS, |
| 5063 | deps = MICROKERNEL_TEST_DEPS, |
| 5064 | ) |
| 5065 | |
| 5066 | xnnpack_unit_test( |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 5067 | name = "x32_fill_test", |
| 5068 | srcs = [ |
| 5069 | "test/x32-fill.cc", |
| 5070 | "test/fill-microkernel-tester.h", |
| 5071 | ] + MICROKERNEL_TEST_HDRS, |
| 5072 | deps = MICROKERNEL_TEST_DEPS, |
| 5073 | ) |
| 5074 | |
| 5075 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5076 | name = "x32_packx_test", |
| 5077 | srcs = [ |
| 5078 | "test/x32-packx.cc", |
| 5079 | "test/pack-microkernel-tester.h", |
| 5080 | "src/xnnpack/AlignedAllocator.h", |
| 5081 | ] + MICROKERNEL_TEST_HDRS, |
| 5082 | deps = MICROKERNEL_TEST_DEPS, |
| 5083 | ) |
| 5084 | |
| 5085 | xnnpack_unit_test( |
| 5086 | name = "x32_pad_test", |
| 5087 | srcs = [ |
| 5088 | "test/x32-pad.cc", |
| 5089 | "test/pad-microkernel-tester.h", |
| 5090 | ] + MICROKERNEL_TEST_HDRS, |
| 5091 | deps = MICROKERNEL_TEST_DEPS, |
| 5092 | ) |
| 5093 | |
| 5094 | xnnpack_unit_test( |
| 5095 | name = "x32_unpool_test", |
| 5096 | srcs = [ |
| 5097 | "test/x32-unpool.cc", |
| 5098 | "test/unpool-microkernel-tester.h", |
| 5099 | ] + MICROKERNEL_TEST_HDRS, |
| 5100 | deps = MICROKERNEL_TEST_DEPS, |
| 5101 | ) |
| 5102 | |
| 5103 | xnnpack_unit_test( |
| 5104 | name = "x32_zip_test", |
| 5105 | srcs = [ |
| 5106 | "test/x32-zip.cc", |
| 5107 | "test/zip-microkernel-tester.h", |
| 5108 | ] + MICROKERNEL_TEST_HDRS, |
| 5109 | deps = MICROKERNEL_TEST_DEPS, |
| 5110 | ) |
| 5111 | |
| 5112 | xnnpack_unit_test( |
| 5113 | name = "x8_lut_test", |
| 5114 | srcs = [ |
| 5115 | "test/x8-lut.cc", |
| 5116 | "test/lut-microkernel-tester.h", |
| 5117 | ] + MICROKERNEL_TEST_HDRS, |
| 5118 | deps = MICROKERNEL_TEST_DEPS, |
| 5119 | ) |
| 5120 | |
| 5121 | xnnpack_unit_test( |
| 5122 | name = "x8_zip_test", |
| 5123 | srcs = [ |
| 5124 | "test/x8-zip.cc", |
| 5125 | "test/zip-microkernel-tester.h", |
| 5126 | ] + MICROKERNEL_TEST_HDRS, |
| 5127 | deps = MICROKERNEL_TEST_DEPS, |
| 5128 | ) |
| 5129 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 5130 | xnnpack_unit_test( |
| 5131 | name = "requantization_test", |
| 5132 | srcs = [ |
| 5133 | "src/xnnpack/requantization-stubs.h", |
| 5134 | "test/requantization.cc", |
| 5135 | "test/requantization-tester.h", |
| 5136 | ] + MICROKERNEL_TEST_HDRS, |
| 5137 | deps = MICROKERNEL_TEST_DEPS, |
| 5138 | ) |
| 5139 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 5140 | ########################## Size tests for the library ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5141 | |
| 5142 | xnnpack_binary( |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 5143 | name = "operator_size_test", |
| 5144 | srcs = ["test/operator-size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 5145 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5146 | ) |
| 5147 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 5148 | xnnpack_binary( |
| 5149 | name = "subgraph_size_test", |
| 5150 | srcs = ["test/subgraph-size.c"], |
| 5151 | deps = [":XNNPACK"], |
| 5152 | ) |
| 5153 | |
| 5154 | ########################### Unit tests for operators ########################## |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5155 | |
| 5156 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5157 | name = "abs_nc_test", |
| 5158 | srcs = [ |
| 5159 | "test/abs-nc.cc", |
| 5160 | "test/abs-operator-tester.h", |
| 5161 | ], |
| 5162 | deps = OPERATOR_TEST_DEPS, |
| 5163 | ) |
| 5164 | |
| 5165 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 5166 | name = "add_nd_test", |
| 5167 | srcs = [ |
| 5168 | "test/add-nd.cc", |
| 5169 | "test/binary-elementwise-operator-tester.h", |
| 5170 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5171 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 5172 | ) |
| 5173 | |
| 5174 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5175 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5176 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5177 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5178 | "test/argmax-pooling-operator-tester.h", |
| 5179 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5180 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5181 | ) |
| 5182 | |
| 5183 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5184 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5185 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5186 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5187 | "test/average-pooling-operator-tester.h", |
| 5188 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5189 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5190 | ) |
| 5191 | |
| 5192 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 5193 | name = "bankers_rounding_nc_test", |
| 5194 | srcs = [ |
| 5195 | "test/bankers-rounding-nc.cc", |
| 5196 | "test/bankers-rounding-operator-tester.h", |
| 5197 | ], |
| 5198 | deps = OPERATOR_TEST_DEPS, |
| 5199 | ) |
| 5200 | |
| 5201 | xnnpack_unit_test( |
| 5202 | name = "ceiling_nc_test", |
| 5203 | srcs = [ |
| 5204 | "test/ceiling-nc.cc", |
| 5205 | "test/ceiling-operator-tester.h", |
| 5206 | ], |
| 5207 | deps = OPERATOR_TEST_DEPS, |
| 5208 | ) |
| 5209 | |
| 5210 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5211 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5212 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5213 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5214 | "test/channel-shuffle-operator-tester.h", |
| 5215 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5216 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5217 | ) |
| 5218 | |
| 5219 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5220 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5221 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5222 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5223 | "test/clamp-operator-tester.h", |
| 5224 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5225 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5226 | ) |
| 5227 | |
| 5228 | xnnpack_unit_test( |
Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 5229 | name = "constant_pad_nd_test", |
| 5230 | srcs = [ |
| 5231 | "test/constant-pad-nd.cc", |
| 5232 | "test/constant-pad-operator-tester.h", |
| 5233 | ], |
| 5234 | deps = OPERATOR_TEST_DEPS, |
| 5235 | ) |
| 5236 | |
| 5237 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5238 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5239 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5240 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5241 | "test/convolution-operator-tester.h", |
| 5242 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5243 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5244 | ) |
| 5245 | |
| 5246 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5247 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5248 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5249 | "test/convolution-nchw.cc", |
| 5250 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5251 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5252 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5253 | ) |
| 5254 | |
| 5255 | xnnpack_unit_test( |
Marat Dukhan | 4e21b27 | 2020-06-04 18:45:01 -0700 | [diff] [blame] | 5256 | name = "copy_nc_test", |
| 5257 | srcs = [ |
| 5258 | "test/copy-nc.cc", |
| 5259 | "test/copy-operator-tester.h", |
| 5260 | ], |
| 5261 | deps = OPERATOR_TEST_DEPS, |
| 5262 | ) |
| 5263 | |
| 5264 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5265 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5266 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5267 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5268 | "test/deconvolution-operator-tester.h", |
| 5269 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5270 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5271 | ) |
| 5272 | |
| 5273 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 5274 | name = "divide_nd_test", |
| 5275 | srcs = [ |
| 5276 | "test/binary-elementwise-operator-tester.h", |
| 5277 | "test/divide-nd.cc", |
| 5278 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5279 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 5280 | ) |
| 5281 | |
| 5282 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5283 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5284 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5285 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5286 | "test/fully-connected-operator-tester.h", |
| 5287 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5288 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5289 | ) |
| 5290 | |
| 5291 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 5292 | name = "floor_nc_test", |
| 5293 | srcs = [ |
| 5294 | "test/floor-nc.cc", |
| 5295 | "test/floor-operator-tester.h", |
| 5296 | ], |
| 5297 | deps = OPERATOR_TEST_DEPS, |
| 5298 | ) |
| 5299 | |
| 5300 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5301 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5302 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5303 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5304 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | ef61d02 | 2020-06-19 13:54:49 -0700 | [diff] [blame] | 5305 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5306 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5307 | ) |
| 5308 | |
| 5309 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5310 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5311 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5312 | "test/global-average-pooling-ncw.cc", |
| 5313 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5314 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5315 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5316 | ) |
| 5317 | |
| 5318 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5319 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5320 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5321 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5322 | "test/hardswish-operator-tester.h", |
| 5323 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5324 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5325 | ) |
| 5326 | |
| 5327 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5328 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5329 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5330 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5331 | "test/leaky-relu-operator-tester.h", |
| 5332 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5333 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5334 | ) |
| 5335 | |
| 5336 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5337 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5338 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5339 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5340 | "test/max-pooling-operator-tester.h", |
| 5341 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5342 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5343 | ) |
| 5344 | |
| 5345 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 5346 | name = "maximum_nd_test", |
| 5347 | srcs = [ |
| 5348 | "test/binary-elementwise-operator-tester.h", |
| 5349 | "test/maximum-nd.cc", |
| 5350 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5351 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 5352 | ) |
| 5353 | |
| 5354 | xnnpack_unit_test( |
| 5355 | name = "minimum_nd_test", |
| 5356 | srcs = [ |
| 5357 | "test/binary-elementwise-operator-tester.h", |
| 5358 | "test/minimum-nd.cc", |
| 5359 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5360 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 5361 | ) |
| 5362 | |
| 5363 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5364 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 5365 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 5366 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5367 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 5368 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5369 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 5370 | ) |
| 5371 | |
| 5372 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5373 | name = "negate_nc_test", |
| 5374 | srcs = [ |
| 5375 | "test/negate-nc.cc", |
| 5376 | "test/negate-operator-tester.h", |
| 5377 | ], |
| 5378 | deps = OPERATOR_TEST_DEPS, |
| 5379 | ) |
| 5380 | |
| 5381 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5382 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5383 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5384 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5385 | "test/prelu-operator-tester.h", |
| 5386 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5387 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5388 | ) |
| 5389 | |
| 5390 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5391 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 5392 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5393 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 5394 | "test/resize-bilinear-operator-tester.h", |
| 5395 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5396 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 5397 | ) |
| 5398 | |
| 5399 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5400 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5401 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5402 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5403 | "test/sigmoid-operator-tester.h", |
| 5404 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5405 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5406 | ) |
| 5407 | |
| 5408 | xnnpack_unit_test( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 5409 | name = "softmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5410 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 5411 | "test/softmax-nc.cc", |
| 5412 | "test/softmax-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5413 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5414 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5415 | ) |
| 5416 | |
| 5417 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5418 | name = "square_nc_test", |
| 5419 | srcs = [ |
| 5420 | "test/square-nc.cc", |
| 5421 | "test/square-operator-tester.h", |
| 5422 | ], |
| 5423 | deps = OPERATOR_TEST_DEPS, |
| 5424 | ) |
| 5425 | |
| 5426 | xnnpack_unit_test( |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 5427 | name = "square_root_nc_test", |
| 5428 | srcs = [ |
| 5429 | "test/square-root-nc.cc", |
| 5430 | "test/square-root-operator-tester.h", |
| 5431 | ], |
| 5432 | deps = OPERATOR_TEST_DEPS, |
| 5433 | ) |
| 5434 | |
| 5435 | xnnpack_unit_test( |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 5436 | name = "squared_difference_nd_test", |
| 5437 | srcs = [ |
| 5438 | "test/binary-elementwise-operator-tester.h", |
| 5439 | "test/squared-difference-nd.cc", |
| 5440 | ], |
| 5441 | deps = OPERATOR_TEST_DEPS, |
| 5442 | ) |
| 5443 | |
| 5444 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 5445 | name = "subtract_nd_test", |
| 5446 | srcs = [ |
| 5447 | "test/binary-elementwise-operator-tester.h", |
| 5448 | "test/subtract-nd.cc", |
| 5449 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5450 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 5451 | ) |
| 5452 | |
| 5453 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 5454 | name = "truncation_nc_test", |
| 5455 | srcs = [ |
| 5456 | "test/truncation-nc.cc", |
| 5457 | "test/truncation-operator-tester.h", |
| 5458 | ], |
| 5459 | deps = OPERATOR_TEST_DEPS, |
| 5460 | ) |
| 5461 | |
| 5462 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5463 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5464 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 5465 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5466 | "test/unpooling-operator-tester.h", |
| 5467 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 5468 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5469 | ) |
| 5470 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 5471 | ############################### Misc unit tests ############################### |
| 5472 | |
| 5473 | xnnpack_unit_test( |
| 5474 | name = "memory_planner_test", |
| 5475 | srcs = [ |
| 5476 | "test/memory-planner-test.cc", |
| 5477 | ], |
| 5478 | deps = [ |
| 5479 | ":XNNPACK", |
| 5480 | ":memory_planner", |
| 5481 | ], |
| 5482 | ) |
| 5483 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5484 | ############################# Build configurations ############################# |
| 5485 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5486 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5487 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5488 | name = "xnn_enable_assembly_explicit_true", |
| 5489 | define_values = {"xnn_enable_assembly": "true"}, |
| 5490 | ) |
| 5491 | |
| 5492 | # Disables usage of assembly kernels. |
| 5493 | config_setting( |
| 5494 | name = "xnn_enable_assembly_explicit_false", |
| 5495 | define_values = {"xnn_enable_assembly": "false"}, |
| 5496 | ) |
| 5497 | |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 5498 | # Enables usage of sparse inference. |
| 5499 | config_setting( |
| 5500 | name = "xnn_enable_sparse_explicit_true", |
| 5501 | define_values = {"xnn_enable_sparse": "true"}, |
| 5502 | ) |
| 5503 | |
| 5504 | # Disables usage of sparse inference. |
| 5505 | config_setting( |
| 5506 | name = "xnn_enable_sparse_explicit_false", |
| 5507 | define_values = {"xnn_enable_sparse": "false"}, |
| 5508 | ) |
| 5509 | |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 5510 | # Disables usage of HMP-aware optimizations. |
| 5511 | config_setting( |
| 5512 | name = "xnn_enable_hmp_explicit_false", |
| 5513 | define_values = {"xnn_enable_hmp": "false"}, |
| 5514 | ) |
| 5515 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 5516 | # Enable usage of optimized memory allocation |
| 5517 | config_setting( |
| 5518 | name = "xnn_enable_memopt_explicit_true", |
| 5519 | define_values = {"xnnpack_enable_memopt": "true"}, |
| 5520 | ) |
| 5521 | |
| 5522 | # Disable usage of optimized memory allocation |
| 5523 | config_setting( |
| 5524 | name = "xnn_enable_memopt_explicit_false", |
| 5525 | define_values = {"xnnpack_enable_memopt": "false"}, |
| 5526 | ) |
| 5527 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5528 | # Builds with -c dbg |
| 5529 | config_setting( |
| 5530 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5531 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5532 | "compilation_mode": "dbg", |
| 5533 | }, |
| 5534 | ) |
| 5535 | |
| 5536 | # Builds with -c opt |
| 5537 | config_setting( |
| 5538 | name = "optimized_build", |
| 5539 | values = { |
| 5540 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5541 | }, |
| 5542 | ) |
| 5543 | |
| 5544 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 5545 | name = "linux_k8", |
| 5546 | values = {"cpu": "k8"}, |
| 5547 | ) |
| 5548 | |
| 5549 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 5550 | name = "linux_arm", |
| 5551 | values = {"cpu": "arm"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 5552 | ) |
| 5553 | |
| 5554 | config_setting( |
Marat Dukhan | f0bd4de | 2020-06-15 15:53:19 -0700 | [diff] [blame] | 5555 | name = "linux_armeabi", |
| 5556 | values = {"cpu": "armeabi"}, |
| 5557 | ) |
| 5558 | |
| 5559 | config_setting( |
Terry Heo | 68eef3f | 2020-04-13 22:53:52 -0700 | [diff] [blame] | 5560 | name = "linux_armhf", |
| 5561 | values = {"cpu": "armhf"}, |
| 5562 | ) |
| 5563 | |
| 5564 | config_setting( |
Marat Dukhan | a720e93 | 2020-06-10 13:01:11 -0700 | [diff] [blame] | 5565 | name = "linux_armv7a", |
| 5566 | values = {"cpu": "armv7a"}, |
| 5567 | ) |
| 5568 | |
| 5569 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 5570 | name = "linux_aarch64", |
| 5571 | values = {"cpu": "aarch64"}, |
| 5572 | ) |
| 5573 | |
| 5574 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5575 | name = "android", |
| 5576 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 5577 | ) |
| 5578 | |
| 5579 | config_setting( |
| 5580 | name = "android_armv7", |
| 5581 | values = { |
| 5582 | "crosstool_top": "//external:android/crosstool", |
| 5583 | "cpu": "armeabi-v7a", |
| 5584 | }, |
| 5585 | ) |
| 5586 | |
| 5587 | config_setting( |
| 5588 | name = "android_arm64", |
| 5589 | values = { |
| 5590 | "crosstool_top": "//external:android/crosstool", |
| 5591 | "cpu": "arm64-v8a", |
| 5592 | }, |
| 5593 | ) |
| 5594 | |
| 5595 | config_setting( |
| 5596 | name = "android_x86", |
| 5597 | values = { |
| 5598 | "crosstool_top": "//external:android/crosstool", |
| 5599 | "cpu": "x86", |
| 5600 | }, |
| 5601 | ) |
| 5602 | |
| 5603 | config_setting( |
| 5604 | name = "android_x86_64", |
| 5605 | values = { |
| 5606 | "crosstool_top": "//external:android/crosstool", |
| 5607 | "cpu": "x86_64", |
| 5608 | }, |
| 5609 | ) |
| 5610 | |
| 5611 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 5612 | name = "windows_x86_64", |
| 5613 | values = {"cpu": "x64_windows"}, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 5614 | ) |
| 5615 | |
| 5616 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 5617 | name = "windows_x86_64_clang", |
| 5618 | values = { |
| 5619 | "compiler": "clang-cl", |
| 5620 | "cpu": "x64_windows", |
| 5621 | }, |
| 5622 | ) |
| 5623 | |
| 5624 | config_setting( |
| 5625 | name = "windows_x86_64_mingw", |
| 5626 | values = { |
| 5627 | "compiler": "mingw-gcc", |
| 5628 | "cpu": "x64_windows", |
| 5629 | }, |
| 5630 | ) |
| 5631 | |
| 5632 | config_setting( |
| 5633 | name = "windows_x86_64_msys", |
| 5634 | values = { |
| 5635 | "compiler": "msys-gcc", |
| 5636 | "cpu": "x64_windows", |
| 5637 | }, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 5638 | ) |
| 5639 | |
| 5640 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 5641 | name = "macos_x86_64", |
| 5642 | values = { |
| 5643 | "apple_platform_type": "macos", |
| 5644 | "cpu": "darwin", |
| 5645 | }, |
| 5646 | ) |
| 5647 | |
| 5648 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5649 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5650 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5651 | ) |
| 5652 | |
| 5653 | config_setting( |
| 5654 | name = "emscripten_wasm", |
| 5655 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5656 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5657 | "cpu": "wasm", |
| 5658 | }, |
| 5659 | ) |
| 5660 | |
| 5661 | config_setting( |
| 5662 | name = "emscripten_wasmsimd", |
| 5663 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5664 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5665 | "cpu": "wasm", |
Marat Dukhan | 81c6260 | 2020-05-29 13:22:49 -0700 | [diff] [blame] | 5666 | "copt": "-msimd128", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5667 | }, |
| 5668 | ) |
| 5669 | |
| 5670 | config_setting( |
| 5671 | name = "emscripten_asmjs", |
| 5672 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 5673 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5674 | "cpu": "asmjs", |
| 5675 | }, |
| 5676 | ) |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5677 | |
| 5678 | config_setting( |
| 5679 | name = "ios_armv7", |
| 5680 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5681 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5682 | "cpu": "ios_armv7", |
| 5683 | }, |
| 5684 | ) |
| 5685 | |
| 5686 | config_setting( |
| 5687 | name = "ios_arm64", |
| 5688 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5689 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5690 | "cpu": "ios_arm64", |
| 5691 | }, |
| 5692 | ) |
| 5693 | |
| 5694 | config_setting( |
| 5695 | name = "ios_arm64e", |
| 5696 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5697 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5698 | "cpu": "ios_arm64e", |
| 5699 | }, |
| 5700 | ) |
| 5701 | |
| 5702 | config_setting( |
| 5703 | name = "ios_x86", |
| 5704 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5705 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5706 | "cpu": "ios_i386", |
| 5707 | }, |
| 5708 | ) |
| 5709 | |
| 5710 | config_setting( |
| 5711 | name = "ios_x86_64", |
| 5712 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5713 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5714 | "cpu": "ios_x86_64", |
| 5715 | }, |
| 5716 | ) |
| 5717 | |
| 5718 | config_setting( |
| 5719 | name = "watchos_armv7k", |
| 5720 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5721 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5722 | "cpu": "watchos_armv7k", |
| 5723 | }, |
| 5724 | ) |
| 5725 | |
| 5726 | config_setting( |
| 5727 | name = "watchos_arm64_32", |
| 5728 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5729 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5730 | "cpu": "watchos_arm64_32", |
| 5731 | }, |
| 5732 | ) |
| 5733 | |
| 5734 | config_setting( |
| 5735 | name = "watchos_x86", |
| 5736 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5737 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5738 | "cpu": "watchos_i386", |
| 5739 | }, |
| 5740 | ) |
| 5741 | |
| 5742 | config_setting( |
| 5743 | name = "watchos_x86_64", |
| 5744 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5745 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5746 | "cpu": "watchos_x86_64", |
| 5747 | }, |
| 5748 | ) |
| 5749 | |
| 5750 | config_setting( |
| 5751 | name = "tvos_arm64", |
| 5752 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5753 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5754 | "cpu": "tvos_arm64", |
| 5755 | }, |
| 5756 | ) |
| 5757 | |
| 5758 | config_setting( |
| 5759 | name = "tvos_x86_64", |
| 5760 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 5761 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 5762 | "cpu": "tvos_x86_64", |
| 5763 | }, |
| 5764 | ) |