Frank Barchard | 1a95305 | 2020-11-16 18:44:58 -0800 | [diff] [blame] | 1 | # Copyright 2020 Google LLC |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 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", |
Marat Dukhan | 13b68f2 | 2020-11-12 11:55:19 -0800 | [diff] [blame] | 62 | "src/operators/depth-to-space-nchw2nhwc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 63 | "src/operators/fully-connected-nc.c", |
| 64 | "src/operators/global-average-pooling-ncw.c", |
| 65 | "src/operators/global-average-pooling-nwc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 66 | "src/operators/leaky-relu-nc.c", |
| 67 | "src/operators/max-pooling-nhwc.c", |
| 68 | "src/operators/prelu-nc.c", |
Artsiom Ablavatski | 9791810 | 2020-10-27 15:52:59 -0700 | [diff] [blame] | 69 | "src/operators/resize-bilinear-nchw.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 70 | "src/operators/resize-bilinear-nhwc.c", |
| 71 | "src/operators/sigmoid-nc.c", |
| 72 | "src/operators/softmax-nc.c", |
Marat Dukhan | c3065f5 | 2020-06-04 13:33:32 -0700 | [diff] [blame] | 73 | "src/operators/unary-elementwise-nc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 74 | "src/operators/unpooling-nhwc.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 75 | ] |
| 76 | |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 77 | SUBGRAPH_SRCS = [ |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 78 | "src/subgraph/abs.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 79 | "src/subgraph/add2.c", |
| 80 | "src/subgraph/argmax-pooling-2d.c", |
| 81 | "src/subgraph/average-pooling-2d.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 82 | "src/subgraph/bankers-rounding.c", |
| 83 | "src/subgraph/ceiling.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 84 | "src/subgraph/clamp.c", |
| 85 | "src/subgraph/convolution-2d.c", |
| 86 | "src/subgraph/deconvolution-2d.c", |
Artsiom Ablavatski | bbe8506 | 2020-11-05 14:07:37 -0800 | [diff] [blame] | 87 | "src/subgraph/depth-to-space.c", |
Frank Barchard | 9cef5ea | 2020-11-18 14:52:08 -0800 | [diff] [blame] | 88 | "src/subgraph/depthwise-convolution-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 89 | "src/subgraph/divide.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 90 | "src/subgraph/floor.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 91 | "src/subgraph/fully-connected.c", |
Marat Dukhan | a059b7d | 2020-06-11 11:41:27 -0700 | [diff] [blame] | 92 | "src/subgraph/global-average-pooling-2d.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 93 | "src/subgraph/hardswish.c", |
Marat Dukhan | 5bbebac | 2020-06-10 19:42:15 -0700 | [diff] [blame] | 94 | "src/subgraph/leaky-relu.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 95 | "src/subgraph/max-pooling-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 96 | "src/subgraph/maximum2.c", |
| 97 | "src/subgraph/minimum2.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 98 | "src/subgraph/multiply2.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 99 | "src/subgraph/negate.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 100 | "src/subgraph/prelu.c", |
| 101 | "src/subgraph/sigmoid.c", |
| 102 | "src/subgraph/softmax.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 103 | "src/subgraph/square-root.c", |
| 104 | "src/subgraph/square.c", |
| 105 | "src/subgraph/squared-difference.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 106 | "src/subgraph/static-constant-pad.c", |
Marat Dukhan | d27202d | 2020-07-09 23:43:40 -0700 | [diff] [blame] | 107 | "src/subgraph/static-reshape.c", |
Marat Dukhan | aff24e2 | 2020-07-23 01:43:58 -0700 | [diff] [blame] | 108 | "src/subgraph/static-resize-bilinear-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 109 | "src/subgraph/subtract.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 110 | "src/subgraph/unpooling-2d.c", |
| 111 | ] |
| 112 | |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 113 | TABLE_SRCS = [ |
| 114 | "src/tables/exp2-k-over-64.c", |
| 115 | "src/tables/exp2-k-over-2048.c", |
Marat Dukhan | c60742b | 2020-11-23 12:33:27 -0800 | [diff] [blame] | 116 | "src/tables/exp2minus-k-over-16.c", |
Marat Dukhan | 1f256fc | 2020-09-24 21:27:14 -0700 | [diff] [blame] | 117 | "src/tables/exp2minus-k-over-64.c", |
| 118 | "src/tables/exp2minus-k-over-2048.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 119 | ] |
| 120 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 121 | SCALAR_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 122 | "src/f32-argmaxpool/4x-scalar-c1.c", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 123 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 124 | "src/f32-argmaxpool/9x-scalar-c1.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 125 | "src/f32-avgpool/9p8x-minmax-scalar-c1.c", |
| 126 | "src/f32-avgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 127 | "src/f32-clamp/gen/scalar-x1.c", |
| 128 | "src/f32-clamp/gen/scalar-x2.c", |
| 129 | "src/f32-clamp/gen/scalar-x4.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 130 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 131 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 132 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 133 | "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 134 | "src/f32-dwconv/gen/up1x4-minmax-scalar.c", |
| 135 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 136 | "src/f32-dwconv/gen/up1x4-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 137 | "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 138 | "src/f32-dwconv/gen/up1x9-minmax-scalar.c", |
| 139 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 140 | "src/f32-dwconv/gen/up1x9-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 141 | "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 142 | "src/f32-dwconv/gen/up1x25-minmax-scalar.c", |
| 143 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 144 | "src/f32-dwconv/gen/up1x25-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 145 | "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 146 | "src/f32-dwconv/gen/up2x4-minmax-scalar.c", |
| 147 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 148 | "src/f32-dwconv/gen/up2x4-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 149 | "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 150 | "src/f32-dwconv/gen/up2x9-minmax-scalar.c", |
| 151 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 152 | "src/f32-dwconv/gen/up2x9-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 153 | "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 154 | "src/f32-dwconv/gen/up2x25-minmax-scalar.c", |
| 155 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 156 | "src/f32-dwconv/gen/up2x25-scalar.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 157 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c", |
| 158 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c", |
| 159 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c", |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 160 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 161 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c", |
Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 162 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c", |
| 163 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c", |
| 164 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c", |
| 165 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c", |
| 166 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c", |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 167 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c", |
| 168 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c", |
| 169 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 170 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c", |
Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 171 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 172 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c", |
| 173 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c", |
| 174 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c", |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 175 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c", |
| 176 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c", |
| 177 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c", |
| 178 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 179 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c", |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 180 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c", |
| 181 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 182 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c", |
Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame] | 183 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 184 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c", |
Marat Dukhan | 29c0c33 | 2020-10-28 22:11:00 -0700 | [diff] [blame] | 185 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c", |
| 186 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c", |
| 187 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c", |
| 188 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c", |
| 189 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c", |
| 190 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c", |
| 191 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c", |
| 192 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c", |
| 193 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c", |
| 194 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 195 | "src/f32-gavgpool-cw/scalar-x1.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 196 | "src/f32-gavgpool/7p7x-minmax-scalar-c1.c", |
| 197 | "src/f32-gavgpool/7x-minmax-scalar-c1.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 198 | "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c", |
| 199 | "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c", |
| 200 | "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 201 | "src/f32-gemm/gen/1x4-minmax-scalar.c", |
| 202 | "src/f32-gemm/gen/1x4-relu-scalar.c", |
| 203 | "src/f32-gemm/gen/1x4-scalar.c", |
| 204 | "src/f32-gemm/gen/2x4-minmax-scalar.c", |
| 205 | "src/f32-gemm/gen/2x4-relu-scalar.c", |
| 206 | "src/f32-gemm/gen/2x4-scalar.c", |
| 207 | "src/f32-gemm/gen/4x2-minmax-scalar.c", |
| 208 | "src/f32-gemm/gen/4x2-relu-scalar.c", |
| 209 | "src/f32-gemm/gen/4x2-scalar.c", |
| 210 | "src/f32-gemm/gen/4x4-minmax-scalar.c", |
| 211 | "src/f32-gemm/gen/4x4-relu-scalar.c", |
| 212 | "src/f32-gemm/gen/4x4-scalar.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 213 | "src/f32-hswish/gen/hswish-scalar-x1.c", |
| 214 | "src/f32-hswish/gen/hswish-scalar-x2.c", |
| 215 | "src/f32-hswish/gen/hswish-scalar-x4.c", |
XNNPACK Team | 2143267 | 2020-10-19 19:58:48 -0700 | [diff] [blame] | 216 | "src/f32-ibilinear-chw/gen/scalar-p1.c", |
| 217 | "src/f32-ibilinear-chw/gen/scalar-p2.c", |
| 218 | "src/f32-ibilinear-chw/gen/scalar-p4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 219 | "src/f32-ibilinear/gen/scalar-c1.c", |
| 220 | "src/f32-ibilinear/gen/scalar-c2.c", |
| 221 | "src/f32-ibilinear/gen/scalar-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 222 | "src/f32-igemm/gen/1x4-minmax-scalar.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 223 | "src/f32-igemm/gen/1x4-relu-scalar.c", |
| 224 | "src/f32-igemm/gen/1x4-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 225 | "src/f32-igemm/gen/2x4-minmax-scalar.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 226 | "src/f32-igemm/gen/2x4-relu-scalar.c", |
| 227 | "src/f32-igemm/gen/2x4-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 228 | "src/f32-igemm/gen/4x2-minmax-scalar.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 229 | "src/f32-igemm/gen/4x2-relu-scalar.c", |
| 230 | "src/f32-igemm/gen/4x2-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 231 | "src/f32-igemm/gen/4x4-minmax-scalar.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 232 | "src/f32-igemm/gen/4x4-relu-scalar.c", |
| 233 | "src/f32-igemm/gen/4x4-scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 234 | "src/f32-maxpool/9p8x-minmax-scalar-c1.c", |
| 235 | "src/f32-pavgpool/9p8x-minmax-scalar-c1.c", |
| 236 | "src/f32-pavgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 237 | "src/f32-ppmm/gen/2x4-minmax-scalar.c", |
| 238 | "src/f32-ppmm/gen/3x3-minmax-scalar.c", |
| 239 | "src/f32-ppmm/gen/4x2-minmax-scalar.c", |
| 240 | "src/f32-ppmm/gen/4x4-minmax-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 241 | "src/f32-prelu/gen/scalar-2x1.c", |
| 242 | "src/f32-prelu/gen/scalar-2x4.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 243 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 244 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 245 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 246 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 247 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 248 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 249 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 250 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 251 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 252 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 253 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 254 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 255 | "src/f32-relu/gen/scalar-x1.c", |
| 256 | "src/f32-relu/gen/scalar-x2.c", |
| 257 | "src/f32-relu/gen/scalar-x4.c", |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 258 | "src/f32-relu/gen/scalar-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 259 | "src/f32-rmax/scalar.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 260 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 261 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 262 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 263 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 264 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 265 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 266 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 267 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 268 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 269 | "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c", |
| 270 | "src/f32-spmm/gen/1x1-minmax-scalar.c", |
| 271 | "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c", |
| 272 | "src/f32-spmm/gen/2x1-minmax-scalar.c", |
| 273 | "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c", |
| 274 | "src/f32-spmm/gen/4x1-minmax-scalar.c", |
| 275 | "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c", |
| 276 | "src/f32-spmm/gen/8x1-minmax-scalar.c", |
| 277 | "src/f32-spmm/gen/8x2-minmax-scalar.c", |
| 278 | "src/f32-spmm/gen/8x4-minmax-scalar.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 279 | "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c", |
| 280 | "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c", |
| 281 | "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 282 | "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 283 | "src/f32-vbinary/gen/vadd-relu-scalar-x1.c", |
| 284 | "src/f32-vbinary/gen/vadd-relu-scalar-x2.c", |
| 285 | "src/f32-vbinary/gen/vadd-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 286 | "src/f32-vbinary/gen/vadd-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 287 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 288 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 289 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 290 | "src/f32-vbinary/gen/vadd-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 291 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c", |
| 292 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c", |
| 293 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 294 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 295 | "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c", |
| 296 | "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c", |
| 297 | "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 298 | "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 299 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 300 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 301 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 302 | "src/f32-vbinary/gen/vaddc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 303 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c", |
| 304 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c", |
| 305 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 306 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 307 | "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c", |
| 308 | "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c", |
| 309 | "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 310 | "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 311 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 312 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 313 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 314 | "src/f32-vbinary/gen/vdiv-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 315 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c", |
| 316 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c", |
| 317 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 318 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 319 | "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c", |
| 320 | "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c", |
| 321 | "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 322 | "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 323 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 324 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 325 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 326 | "src/f32-vbinary/gen/vdivc-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 327 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 328 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 329 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 330 | "src/f32-vbinary/gen/vmax-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 331 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 332 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 333 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 334 | "src/f32-vbinary/gen/vmaxc-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 335 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 336 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 337 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 338 | "src/f32-vbinary/gen/vmin-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 339 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 340 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 341 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 342 | "src/f32-vbinary/gen/vminc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 343 | "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c", |
| 344 | "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c", |
| 345 | "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 346 | "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 347 | "src/f32-vbinary/gen/vmul-relu-scalar-x1.c", |
| 348 | "src/f32-vbinary/gen/vmul-relu-scalar-x2.c", |
| 349 | "src/f32-vbinary/gen/vmul-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 350 | "src/f32-vbinary/gen/vmul-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 351 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 352 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 353 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 354 | "src/f32-vbinary/gen/vmul-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 355 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c", |
| 356 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c", |
| 357 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 358 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 359 | "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c", |
| 360 | "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c", |
| 361 | "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 362 | "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 363 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 364 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 365 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 366 | "src/f32-vbinary/gen/vmulc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 367 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c", |
| 368 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c", |
| 369 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 370 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 371 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c", |
| 372 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c", |
| 373 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 374 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 375 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 376 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 377 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 378 | "src/f32-vbinary/gen/vrdivc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 379 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c", |
| 380 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c", |
| 381 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 382 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 383 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c", |
| 384 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c", |
| 385 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 386 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 387 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 388 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 389 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 390 | "src/f32-vbinary/gen/vrsubc-scalar-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 391 | "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c", |
| 392 | "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c", |
| 393 | "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 394 | "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 395 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c", |
| 396 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c", |
| 397 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 398 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 399 | "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c", |
| 400 | "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c", |
| 401 | "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 402 | "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 403 | "src/f32-vbinary/gen/vsub-relu-scalar-x1.c", |
| 404 | "src/f32-vbinary/gen/vsub-relu-scalar-x2.c", |
| 405 | "src/f32-vbinary/gen/vsub-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 406 | "src/f32-vbinary/gen/vsub-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 407 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 408 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 409 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 410 | "src/f32-vbinary/gen/vsub-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 411 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c", |
| 412 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c", |
| 413 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 414 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 415 | "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c", |
| 416 | "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c", |
| 417 | "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 418 | "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 419 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 420 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 421 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 422 | "src/f32-vbinary/gen/vsubc-scalar-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 423 | "src/f32-vlrelu/gen/vlrelu-scalar-x1.c", |
| 424 | "src/f32-vlrelu/gen/vlrelu-scalar-x2.c", |
| 425 | "src/f32-vlrelu/gen/vlrelu-scalar-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 426 | "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c", |
| 427 | "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c", |
| 428 | "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 429 | "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c", |
| 430 | "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c", |
| 431 | "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 432 | "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c", |
| 433 | "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c", |
| 434 | "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c", |
| 435 | "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c", |
| 436 | "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c", |
| 437 | "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c", |
| 438 | "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c", |
| 439 | "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c", |
| 440 | "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 441 | "src/f32-vsqrt/gen/scalar-sqrt-x1.c", |
| 442 | "src/f32-vsqrt/gen/scalar-sqrt-x2.c", |
| 443 | "src/f32-vsqrt/gen/scalar-sqrt-x4.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 444 | "src/f32-vunary/gen/vabs-scalar-x1.c", |
| 445 | "src/f32-vunary/gen/vabs-scalar-x2.c", |
| 446 | "src/f32-vunary/gen/vabs-scalar-x4.c", |
| 447 | "src/f32-vunary/gen/vneg-scalar-x1.c", |
| 448 | "src/f32-vunary/gen/vneg-scalar-x2.c", |
| 449 | "src/f32-vunary/gen/vneg-scalar-x4.c", |
| 450 | "src/f32-vunary/gen/vsqr-scalar-x1.c", |
| 451 | "src/f32-vunary/gen/vsqr-scalar-x2.c", |
| 452 | "src/f32-vunary/gen/vsqr-scalar-x4.c", |
Marat Dukhan | c60742b | 2020-11-23 12:33:27 -0800 | [diff] [blame] | 453 | "src/math/expm1minus-scalar-rr2-lut16-p3.c", |
| 454 | "src/math/expm1minus-scalar-rr2-lut16-p4.c", |
| 455 | "src/math/expm1minus-scalar-rr2-p5.c", |
| 456 | "src/math/expm1minus-scalar-rr2-p6.c", |
Frank Barchard | 2213606 | 2020-11-24 18:44:46 -0800 | [diff] [blame^] | 457 | "src/math/expminus-scalar-rr2-lut64-p2.c", |
| 458 | "src/math/expminus-scalar-rr2-lut2048-p1.c", |
| 459 | "src/math/expminus-scalar-rr2-p5.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 460 | "src/math/roundd-scalar-addsub.c", |
| 461 | "src/math/roundd-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 462 | "src/math/roundd-scalar-floor.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 463 | "src/math/roundne-scalar-addsub.c", |
| 464 | "src/math/roundne-scalar-nearbyint.c", |
| 465 | "src/math/roundne-scalar-rint.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 466 | "src/math/roundu-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 467 | "src/math/roundu-scalar-ceil.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 468 | "src/math/roundu-scalar-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 469 | "src/math/roundz-scalar-addsub.c", |
| 470 | "src/math/roundz-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 471 | "src/math/roundz-scalar-trunc.c", |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 472 | "src/math/sigmoid-scalar-rr2-lut64-p2-div.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 473 | "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c", |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 474 | "src/math/sigmoid-scalar-rr2-p5-div.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 475 | "src/qs8-gemm/gen/8x8c4-minmax-scalar.c", |
| 476 | "src/qs8-gemm/gen/12x4c4-minmax-scalar.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 477 | "src/qs8-requantization/fp32-scalar-lrintf.c", |
| 478 | "src/qs8-requantization/fp32-scalar-magic.c", |
| 479 | "src/qs8-requantization/precise-scalar-signed64.c", |
| 480 | "src/qs8-requantization/precise-scalar-unsigned32.c", |
| 481 | "src/qs8-requantization/precise-scalar-unsigned64.c", |
| 482 | "src/qs8-requantization/q31-scalar.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 483 | "src/qu8-avgpool/9p8x-minmax-scalar-c1.c", |
| 484 | "src/qu8-avgpool/9x-minmax-scalar-c1.c", |
| 485 | "src/qu8-dwconv/up1x9-minmax-scalar.c", |
| 486 | "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c", |
| 487 | "src/qu8-gavgpool/7x-minmax-scalar-c1.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 488 | "src/qu8-gemm/2x2-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 489 | "src/qu8-gemm/gen/8x8c4-minmax-scalar.c", |
| 490 | "src/qu8-gemm/gen/12x4c4-minmax-scalar.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 491 | "src/qu8-igemm/2x2-minmax-scalar.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 492 | "src/qu8-requantization/fp32-scalar-lrintf.c", |
| 493 | "src/qu8-requantization/fp32-scalar-magic.c", |
| 494 | "src/qu8-requantization/precise-scalar-signed64.c", |
| 495 | "src/qu8-requantization/precise-scalar-unsigned32.c", |
| 496 | "src/qu8-requantization/precise-scalar-unsigned64.c", |
| 497 | "src/qu8-requantization/q31-scalar.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 498 | "src/qu8-vadd/minmax-scalar.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 499 | "src/u8-clamp/scalar-x4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 500 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 501 | "src/u8-maxpool/9p8x-minmax-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 502 | "src/u8-rmax/scalar.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 503 | "src/x8-lut/scalar.c", |
| 504 | "src/x8-zip/x2-scalar.c", |
| 505 | "src/x8-zip/x3-scalar.c", |
| 506 | "src/x8-zip/x4-scalar.c", |
| 507 | "src/x8-zip/xm-scalar.c", |
Marat Dukhan | ad71b9a | 2020-11-20 00:01:51 -0800 | [diff] [blame] | 508 | "src/x32-depthtospace2d-chw2hwc/scalar.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 509 | "src/x32-fill/scalar-float.c", |
| 510 | "src/x32-fill/scalar-int.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 511 | "src/x32-packx/x2-scalar.c", |
| 512 | "src/x32-packx/x3-scalar.c", |
| 513 | "src/x32-packx/x4-scalar.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 514 | "src/x32-pad/scalar-float.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 515 | "src/x32-pad/scalar-int.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 516 | "src/x32-unpool/scalar.c", |
| 517 | "src/x32-zip/x2-scalar.c", |
| 518 | "src/x32-zip/x3-scalar.c", |
| 519 | "src/x32-zip/x4-scalar.c", |
| 520 | "src/x32-zip/xm-scalar.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 521 | ] |
| 522 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 523 | WASM_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 524 | "src/f32-avgpool/9p8x-minmax-wasm-c1.c", |
| 525 | "src/f32-avgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 526 | "src/f32-clamp/gen/wasm-x1.c", |
| 527 | "src/f32-clamp/gen/wasm-x2.c", |
| 528 | "src/f32-clamp/gen/wasm-x4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 529 | "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c", |
| 530 | "src/f32-dwconv/gen/up1x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 531 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 532 | "src/f32-dwconv/gen/up1x4-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 533 | "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c", |
| 534 | "src/f32-dwconv/gen/up1x9-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 535 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 536 | "src/f32-dwconv/gen/up1x9-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 537 | "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c", |
| 538 | "src/f32-dwconv/gen/up1x25-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 539 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 540 | "src/f32-dwconv/gen/up1x25-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 541 | "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c", |
| 542 | "src/f32-dwconv/gen/up2x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 543 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 544 | "src/f32-dwconv/gen/up2x4-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 545 | "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c", |
| 546 | "src/f32-dwconv/gen/up2x9-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 547 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 548 | "src/f32-dwconv/gen/up2x9-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 549 | "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c", |
| 550 | "src/f32-dwconv/gen/up2x25-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 551 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 552 | "src/f32-dwconv/gen/up2x25-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 553 | "src/f32-gavgpool/7p7x-minmax-wasm-c1.c", |
| 554 | "src/f32-gavgpool/7x-minmax-wasm-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 555 | "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c", |
| 556 | "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c", |
| 557 | "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c", |
| 558 | "src/f32-gemm/gen/1x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 559 | "src/f32-gemm/gen/1x4-relu-wasm.c", |
| 560 | "src/f32-gemm/gen/1x4-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 561 | "src/f32-gemm/gen/2x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 562 | "src/f32-gemm/gen/2x4-relu-wasm.c", |
| 563 | "src/f32-gemm/gen/2x4-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 564 | "src/f32-gemm/gen/4x2-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 565 | "src/f32-gemm/gen/4x2-relu-wasm.c", |
| 566 | "src/f32-gemm/gen/4x2-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 567 | "src/f32-gemm/gen/4x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 568 | "src/f32-gemm/gen/4x4-relu-wasm.c", |
| 569 | "src/f32-gemm/gen/4x4-wasm.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 570 | "src/f32-hswish/gen/hswish-wasm-x1.c", |
| 571 | "src/f32-hswish/gen/hswish-wasm-x2.c", |
| 572 | "src/f32-hswish/gen/hswish-wasm-x4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 573 | "src/f32-igemm/gen/1x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 574 | "src/f32-igemm/gen/1x4-relu-wasm.c", |
| 575 | "src/f32-igemm/gen/1x4-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 576 | "src/f32-igemm/gen/2x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 577 | "src/f32-igemm/gen/2x4-relu-wasm.c", |
| 578 | "src/f32-igemm/gen/2x4-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 579 | "src/f32-igemm/gen/4x2-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 580 | "src/f32-igemm/gen/4x2-relu-wasm.c", |
| 581 | "src/f32-igemm/gen/4x2-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 582 | "src/f32-igemm/gen/4x4-minmax-wasm.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 583 | "src/f32-igemm/gen/4x4-relu-wasm.c", |
| 584 | "src/f32-igemm/gen/4x4-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 585 | "src/f32-maxpool/9p8x-minmax-wasm-c1.c", |
| 586 | "src/f32-pavgpool/9p8x-minmax-wasm-c1.c", |
| 587 | "src/f32-pavgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 7c1f808 | 2020-06-25 13:26:20 -0700 | [diff] [blame] | 588 | "src/f32-prelu/gen/wasm-2x1.c", |
| 589 | "src/f32-prelu/gen/wasm-2x4.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 590 | "src/f32-relu/gen/wasm-x1.c", |
| 591 | "src/f32-relu/gen/wasm-x2.c", |
| 592 | "src/f32-relu/gen/wasm-x4.c", |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 593 | "src/f32-relu/gen/wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 594 | "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c", |
| 595 | "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c", |
| 596 | "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 597 | "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 598 | "src/f32-vbinary/gen/vadd-relu-wasm-x1.c", |
| 599 | "src/f32-vbinary/gen/vadd-relu-wasm-x2.c", |
| 600 | "src/f32-vbinary/gen/vadd-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 601 | "src/f32-vbinary/gen/vadd-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 602 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c", |
| 603 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c", |
| 604 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c", |
| 605 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 606 | "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c", |
| 607 | "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c", |
| 608 | "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 609 | "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 610 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c", |
| 611 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c", |
| 612 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c", |
| 613 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 614 | "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c", |
| 615 | "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c", |
| 616 | "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 617 | "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 618 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c", |
| 619 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c", |
| 620 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c", |
| 621 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 622 | "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c", |
| 623 | "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c", |
| 624 | "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 625 | "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 626 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 627 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 628 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 629 | "src/f32-vbinary/gen/vmax-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 630 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 631 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 632 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 633 | "src/f32-vbinary/gen/vmaxc-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 634 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 635 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 636 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 637 | "src/f32-vbinary/gen/vmin-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 638 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 639 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 640 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 641 | "src/f32-vbinary/gen/vminc-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 642 | "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c", |
| 643 | "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c", |
| 644 | "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 645 | "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 646 | "src/f32-vbinary/gen/vmul-relu-wasm-x1.c", |
| 647 | "src/f32-vbinary/gen/vmul-relu-wasm-x2.c", |
| 648 | "src/f32-vbinary/gen/vmul-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 649 | "src/f32-vbinary/gen/vmul-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 650 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c", |
| 651 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c", |
| 652 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c", |
| 653 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 654 | "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c", |
| 655 | "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c", |
| 656 | "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 657 | "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 658 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c", |
| 659 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c", |
| 660 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c", |
| 661 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 662 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c", |
| 663 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c", |
| 664 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 665 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 666 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c", |
| 667 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c", |
| 668 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c", |
| 669 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 670 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c", |
| 671 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c", |
| 672 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 673 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 674 | "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c", |
| 675 | "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c", |
| 676 | "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c", |
| 677 | "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 678 | "src/f32-vbinary/gen/vsub-relu-wasm-x1.c", |
| 679 | "src/f32-vbinary/gen/vsub-relu-wasm-x2.c", |
| 680 | "src/f32-vbinary/gen/vsub-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 681 | "src/f32-vbinary/gen/vsub-relu-wasm-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 682 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c", |
| 683 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c", |
| 684 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c", |
| 685 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 686 | "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c", |
| 687 | "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c", |
| 688 | "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 689 | "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 690 | "src/f32-vlrelu/gen/vlrelu-wasm-x1.c", |
| 691 | "src/f32-vlrelu/gen/vlrelu-wasm-x2.c", |
| 692 | "src/f32-vlrelu/gen/vlrelu-wasm-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 693 | "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c", |
| 694 | "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c", |
| 695 | "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 696 | ] |
| 697 | |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 698 | WASMSIMD_UKERNELS = [ |
Marat Dukhan | 40f0552 | 2020-07-16 22:33:12 -0700 | [diff] [blame] | 699 | "src/f32-argmaxpool/4x-wasmsimd-c4.c", |
| 700 | "src/f32-argmaxpool/9p8x-wasmsimd-c4.c", |
| 701 | "src/f32-argmaxpool/9x-wasmsimd-c4.c", |
Marat Dukhan | 3b7432d | 2020-07-16 17:46:32 -0700 | [diff] [blame] | 702 | "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 703 | "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| 704 | "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c", |
| 705 | "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 706 | "src/f32-clamp/gen/wasmsimd-arm-x4.c", |
| 707 | "src/f32-clamp/gen/wasmsimd-arm-x8.c", |
| 708 | "src/f32-clamp/gen/wasmsimd-x86-x4.c", |
| 709 | "src/f32-clamp/gen/wasmsimd-x86-x8.c", |
Frank Barchard | 2213606 | 2020-11-24 18:44:46 -0800 | [diff] [blame^] | 710 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 711 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 712 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 713 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 714 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 715 | "src/f32-dwconv/gen/up4x4-wasmsimd.c", |
| 716 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 717 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 718 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 719 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 720 | "src/f32-dwconv/gen/up4x9-wasmsimd.c", |
| 721 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c", |
| 722 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c", |
| 723 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c", |
| 724 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c", |
| 725 | "src/f32-dwconv/gen/up4x25-wasmsimd.c", |
| 726 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 727 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 728 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 729 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 730 | "src/f32-dwconv/gen/up8x4-wasmsimd.c", |
| 731 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 732 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 733 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 734 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 735 | "src/f32-dwconv/gen/up8x9-wasmsimd.c", |
| 736 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c", |
| 737 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c", |
| 738 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c", |
| 739 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c", |
| 740 | "src/f32-dwconv/gen/up8x25-wasmsimd.c", |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 741 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c", |
| 742 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c", |
| 743 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c", |
| 744 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c", |
| 745 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c", |
| 746 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c", |
| 747 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c", |
| 748 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c", |
| 749 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c", |
| 750 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c", |
| 751 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c", |
| 752 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c", |
| 753 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c", |
| 754 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c", |
| 755 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c", |
| 756 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c", |
| 757 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c", |
| 758 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c", |
| 759 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c", |
| 760 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c", |
Frank Barchard | 9cef5ea | 2020-11-18 14:52:08 -0800 | [diff] [blame] | 761 | "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c", |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 762 | "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c", |
Frank Barchard | 9cef5ea | 2020-11-18 14:52:08 -0800 | [diff] [blame] | 763 | "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c", |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 764 | "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c", |
Frank Barchard | 9cef5ea | 2020-11-18 14:52:08 -0800 | [diff] [blame] | 765 | "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c", |
Frank Barchard | db5c32d | 2020-11-16 23:58:42 -0800 | [diff] [blame] | 766 | "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 767 | "src/f32-gavgpool-cw/wasmsimd-arm-x4.c", |
| 768 | "src/f32-gavgpool-cw/wasmsimd-x86-x4.c", |
Marat Dukhan | c601680 | 2020-07-16 18:51:28 -0700 | [diff] [blame] | 769 | "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c", |
| 770 | "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c", |
| 771 | "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c", |
| 772 | "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 773 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 774 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 775 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c", |
| 776 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c", |
| 777 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c", |
| 778 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c", |
| 779 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 780 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 781 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c", |
| 782 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c", |
| 783 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c", |
| 784 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c", |
| 785 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 786 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 787 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c", |
| 788 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c", |
| 789 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c", |
| 790 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c", |
| 791 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 792 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 793 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c", |
| 794 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c", |
| 795 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c", |
| 796 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c", |
| 797 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 798 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 799 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c", |
| 800 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c", |
| 801 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c", |
| 802 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c", |
| 803 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 804 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 805 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 806 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 807 | "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c", |
| 808 | "src/f32-gemm/gen/1x8-wasmsimd-splat.c", |
| 809 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 810 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 811 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 812 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 813 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 814 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 815 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 816 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 817 | "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c", |
| 818 | "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c", |
| 819 | "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c", |
| 820 | "src/f32-gemm/gen/4x2c4-wasmsimd.c", |
| 821 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 822 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 823 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 824 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 825 | "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c", |
| 826 | "src/f32-gemm/gen/4x8-wasmsimd-splat.c", |
| 827 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 828 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 829 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 830 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 831 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 832 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 833 | "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c", |
| 834 | "src/f32-gemm/gen/5x8-wasmsimd-splat.c", |
| 835 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 836 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 837 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 838 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 839 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 840 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 841 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 842 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 843 | "src/f32-hswish/gen/hswish-wasmsimd-x4.c", |
| 844 | "src/f32-hswish/gen/hswish-wasmsimd-x8.c", |
| 845 | "src/f32-hswish/gen/hswish-wasmsimd-x16.c", |
XNNPACK Team | 965272b | 2020-10-23 21:10:15 -0700 | [diff] [blame] | 846 | "src/f32-ibilinear-chw/gen/wasmsimd-p4.c", |
| 847 | "src/f32-ibilinear-chw/gen/wasmsimd-p8.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 848 | "src/f32-ibilinear/gen/wasmsimd-c4.c", |
| 849 | "src/f32-ibilinear/gen/wasmsimd-c8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 850 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 851 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 852 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 853 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 854 | "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c", |
| 855 | "src/f32-igemm/gen/1x8-wasmsimd-splat.c", |
| 856 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 857 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 858 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 859 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 860 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 861 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 862 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 863 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 864 | "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c", |
| 865 | "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c", |
| 866 | "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c", |
| 867 | "src/f32-igemm/gen/4x2c4-wasmsimd.c", |
| 868 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 869 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 870 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 871 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 872 | "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c", |
| 873 | "src/f32-igemm/gen/4x8-wasmsimd-splat.c", |
| 874 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 875 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 876 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 877 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 878 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 879 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 880 | "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c", |
| 881 | "src/f32-igemm/gen/5x8-wasmsimd-splat.c", |
| 882 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 883 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 884 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 885 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 886 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 887 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 888 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 889 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
Marat Dukhan | f6e2480 | 2020-07-08 22:20:40 -0700 | [diff] [blame] | 890 | "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 891 | "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 1483c53 | 2020-07-16 18:08:19 -0700 | [diff] [blame] | 892 | "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 893 | "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| 894 | "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c", |
| 895 | "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 0d0d882 | 2020-07-23 23:37:56 -0700 | [diff] [blame] | 896 | "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 897 | "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 898 | "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c", |
| 899 | "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c", |
| 900 | "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c", |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 901 | "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c", |
| 902 | "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 903 | "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c", |
| 904 | "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c", |
| 905 | "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c", |
| 906 | "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c", |
| 907 | "src/f32-prelu/gen/wasmsimd-minmax-1x4.c", |
| 908 | "src/f32-prelu/gen/wasmsimd-minmax-1x8.c", |
| 909 | "src/f32-prelu/gen/wasmsimd-minmax-1x16.c", |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 910 | "src/f32-prelu/gen/wasmsimd-minmax-2x4.c", |
| 911 | "src/f32-prelu/gen/wasmsimd-minmax-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 912 | "src/f32-prelu/gen/wasmsimd-minmax-2x16.c", |
| 913 | "src/f32-prelu/gen/wasmsimd-minmax-4x4.c", |
| 914 | "src/f32-prelu/gen/wasmsimd-minmax-4x8.c", |
| 915 | "src/f32-prelu/gen/wasmsimd-minmax-4x16.c", |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 916 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c", |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 917 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 918 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c", |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 919 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c", |
| 920 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 921 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c", |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 922 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c", |
| 923 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 924 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c", |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 925 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c", |
| 926 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 927 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 928 | "src/f32-relu/gen/wasmsimd-x4.c", |
| 929 | "src/f32-relu/gen/wasmsimd-x8.c", |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 930 | "src/f32-relu/gen/wasmsimd-x16.c", |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 931 | "src/f32-rmax/wasmsimd-arm.c", |
| 932 | "src/f32-rmax/wasmsimd-x86.c", |
Marat Dukhan | d187a5b | 2020-07-20 01:07:17 -0700 | [diff] [blame] | 933 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c", |
| 934 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c", |
| 935 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c", |
| 936 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c", |
| 937 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c", |
| 938 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 939 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c", |
| 940 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c", |
| 941 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c", |
| 942 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c", |
| 943 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c", |
| 944 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 945 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c", |
| 946 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 947 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c", |
| 948 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 949 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 950 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c", |
| 951 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 952 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c", |
| 953 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 954 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 955 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c", |
| 956 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 957 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c", |
| 958 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 959 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 960 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c", |
| 961 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 962 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c", |
| 963 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 964 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 965 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c", |
| 966 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 967 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c", |
| 968 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 969 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 970 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c", |
| 971 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 972 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c", |
| 973 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 974 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 975 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c", |
| 976 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 977 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c", |
| 978 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c", |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 979 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 980 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c", |
| 981 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 982 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c", |
| 983 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c", |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 984 | "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 985 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c", |
| 986 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 987 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 988 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c", |
| 989 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 990 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 991 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c", |
| 992 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 993 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 994 | "src/f32-vbinary/gen/vadd-wasmsimd-x4.c", |
| 995 | "src/f32-vbinary/gen/vadd-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 996 | "src/f32-vbinary/gen/vadd-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 997 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c", |
| 998 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 999 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1000 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c", |
| 1001 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1002 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1003 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c", |
| 1004 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1005 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1006 | "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c", |
| 1007 | "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1008 | "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1009 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c", |
| 1010 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1011 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1012 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c", |
| 1013 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1014 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1015 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c", |
| 1016 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1017 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1018 | "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c", |
| 1019 | "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1020 | "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1021 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c", |
| 1022 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1023 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1024 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c", |
| 1025 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1026 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1027 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c", |
| 1028 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1029 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1030 | "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c", |
| 1031 | "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1032 | "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1033 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c", |
| 1034 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1035 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1036 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c", |
| 1037 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1038 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1039 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c", |
| 1040 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1041 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1042 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c", |
| 1043 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1044 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1045 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c", |
| 1046 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1047 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1048 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c", |
| 1049 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1050 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1051 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c", |
| 1052 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1053 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1054 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c", |
| 1055 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1056 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1057 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c", |
| 1058 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1059 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1060 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c", |
| 1061 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1062 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1063 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c", |
| 1064 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1065 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1066 | "src/f32-vbinary/gen/vmul-wasmsimd-x4.c", |
| 1067 | "src/f32-vbinary/gen/vmul-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1068 | "src/f32-vbinary/gen/vmul-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1069 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c", |
| 1070 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1071 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1072 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c", |
| 1073 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1074 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1075 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c", |
| 1076 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1077 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1078 | "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c", |
| 1079 | "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1080 | "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1081 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c", |
| 1082 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1083 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1084 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c", |
| 1085 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1086 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1087 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c", |
| 1088 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1089 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1090 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c", |
| 1091 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1092 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1093 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c", |
| 1094 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1095 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1096 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c", |
| 1097 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1098 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1099 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c", |
| 1100 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1101 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1102 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c", |
| 1103 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1104 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1105 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c", |
| 1106 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1107 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1108 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c", |
| 1109 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1110 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1111 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c", |
| 1112 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1113 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1114 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c", |
| 1115 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1116 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1117 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c", |
| 1118 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1119 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1120 | "src/f32-vbinary/gen/vsub-wasmsimd-x4.c", |
| 1121 | "src/f32-vbinary/gen/vsub-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1122 | "src/f32-vbinary/gen/vsub-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1123 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c", |
| 1124 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1125 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1126 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c", |
| 1127 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1128 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1129 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c", |
| 1130 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1131 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1132 | "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c", |
| 1133 | "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1134 | "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1135 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c", |
| 1136 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c", |
| 1137 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c", |
| 1138 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c", |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1139 | "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c", |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1140 | "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1141 | "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c", |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1142 | "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c", |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 1143 | "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c", |
| 1144 | "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c", |
| 1145 | "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c", |
| 1146 | "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1147 | "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c", |
| 1148 | "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c", |
| 1149 | "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c", |
| 1150 | "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c", |
| 1151 | "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c", |
| 1152 | "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c", |
| 1153 | "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c", |
| 1154 | "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c", |
| 1155 | "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c", |
| 1156 | "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1157 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c", |
| 1158 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c", |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 1159 | "src/f32-vunary/gen/vabs-wasmsimd-x4.c", |
| 1160 | "src/f32-vunary/gen/vabs-wasmsimd-x8.c", |
| 1161 | "src/f32-vunary/gen/vneg-wasmsimd-x4.c", |
| 1162 | "src/f32-vunary/gen/vneg-wasmsimd-x8.c", |
| 1163 | "src/f32-vunary/gen/vsqr-wasmsimd-x4.c", |
| 1164 | "src/f32-vunary/gen/vsqr-wasmsimd-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1165 | "src/math/roundd-wasmsimd-addsub.c", |
| 1166 | "src/math/roundd-wasmsimd-cvt.c", |
| 1167 | "src/math/roundne-wasmsimd-addsub.c", |
| 1168 | "src/math/roundu-wasmsimd-addsub.c", |
| 1169 | "src/math/roundu-wasmsimd-cvt.c", |
| 1170 | "src/math/roundz-wasmsimd-addsub.c", |
| 1171 | "src/math/roundz-wasmsimd-cvt.c", |
| 1172 | "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c", |
| 1173 | "src/math/sigmoid-wasmsimd-rr2-p5-div.c", |
Marat Dukhan | cc8f34c | 2020-08-05 16:36:38 -0700 | [diff] [blame] | 1174 | "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c", |
| 1175 | "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c", |
| 1176 | "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c", |
Marat Dukhan | b5e3d17 | 2020-08-06 13:29:53 -0700 | [diff] [blame] | 1177 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c", |
| 1178 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c", |
| 1179 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1180 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c", |
| 1181 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c", |
| 1182 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1183 | "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1184 | "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1185 | "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1186 | "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c", |
| 1187 | "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1188 | "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1189 | "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c", |
| 1190 | "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1191 | "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c", |
| 1192 | "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1193 | "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1194 | "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1195 | "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1196 | "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1197 | "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1198 | "src/qs8-requantization/fp32-wasmsimd.c", |
Marat Dukhan | 138560c | 2020-08-03 18:57:34 -0700 | [diff] [blame] | 1199 | "src/qs8-requantization/q31-wasmsimd.c", |
Marat Dukhan | 5df27f8 | 2020-09-02 23:59:21 -0700 | [diff] [blame] | 1200 | "src/qs8-vadd/gen/minmax-wasmsimd-x8.c", |
| 1201 | "src/qs8-vadd/gen/minmax-wasmsimd-x16.c", |
| 1202 | "src/qs8-vadd/gen/minmax-wasmsimd-x24.c", |
| 1203 | "src/qs8-vadd/gen/minmax-wasmsimd-x32.c", |
| 1204 | "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c", |
| 1205 | "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c", |
| 1206 | "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c", |
| 1207 | "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1208 | "src/qu8-requantization/fp32-wasmsimd.c", |
Marat Dukhan | 22de5e7 | 2020-08-03 21:53:18 -0700 | [diff] [blame] | 1209 | "src/qu8-requantization/q31-wasmsimd.c", |
Marat Dukhan | 8ee3701 | 2020-07-16 13:17:13 -0700 | [diff] [blame] | 1210 | "src/x32-fill/wasmsimd.c", |
Marat Dukhan | 66d99e9 | 2020-07-16 12:56:21 -0700 | [diff] [blame] | 1211 | "src/x32-packx/x4-wasmsimd.c", |
Marat Dukhan | 9306ae0 | 2020-07-16 15:51:13 -0700 | [diff] [blame] | 1212 | "src/x32-pad/wasmsimd.c", |
Marat Dukhan | 9d4bfa2 | 2020-07-16 19:07:04 -0700 | [diff] [blame] | 1213 | "src/x32-unpool/wasmsimd.c", |
Marat Dukhan | e3b7876 | 2020-07-16 20:02:58 -0700 | [diff] [blame] | 1214 | "src/x32-zip/x2-wasmsimd.c", |
| 1215 | "src/x32-zip/x3-wasmsimd.c", |
| 1216 | "src/x32-zip/x4-wasmsimd.c", |
| 1217 | "src/x32-zip/xm-wasmsimd.c", |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 1218 | ] |
| 1219 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1220 | PSIMD_ACCMATH_UKERNELS = [ |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1221 | "src/qs8-requantization/fp32-psimd.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1222 | "src/qs8-requantization/precise-psimd.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1223 | "src/qu8-requantization/fp32-psimd.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1224 | "src/qu8-requantization/precise-psimd.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1225 | ] |
| 1226 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1227 | # ISA-specific micro-kernels |
| 1228 | NEON_UKERNELS = [ |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 1229 | "src/f32-argmaxpool/4x-neon-c4.c", |
| 1230 | "src/f32-argmaxpool/9p8x-neon-c4.c", |
| 1231 | "src/f32-argmaxpool/9x-neon-c4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1232 | "src/f32-avgpool/9p8x-minmax-neon-c4.c", |
| 1233 | "src/f32-avgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1234 | "src/f32-clamp/gen/neon-x4.c", |
| 1235 | "src/f32-clamp/gen/neon-x8.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1236 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1237 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1238 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1239 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1240 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1241 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1242 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1243 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1244 | "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1245 | "src/f32-dwconv/gen/up4x4-minmax-neon.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1246 | "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1247 | "src/f32-dwconv/gen/up4x9-minmax-neon.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1248 | "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1249 | "src/f32-dwconv/gen/up4x25-minmax-neon.c", |
| 1250 | "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", |
| 1251 | "src/f32-dwconv/gen/up8x4-minmax-neon.c", |
| 1252 | "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", |
| 1253 | "src/f32-dwconv/gen/up8x9-minmax-neon.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1254 | "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1255 | "src/f32-dwconv/gen/up8x25-minmax-neon.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1256 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c", |
| 1257 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c", |
| 1258 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c", |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 1259 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1260 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c", |
Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 1261 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c", |
| 1262 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c", |
| 1263 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c", |
| 1264 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c", |
| 1265 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c", |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 1266 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c", |
| 1267 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c", |
| 1268 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1269 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c", |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 1270 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1271 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c", |
| 1272 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c", |
| 1273 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1274 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c", |
| 1275 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c", |
| 1276 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c", |
| 1277 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1278 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1279 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c", |
| 1280 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1281 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1282 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1283 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1284 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1285 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c", |
| 1286 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c", |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 1287 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c", |
| 1288 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c", |
| 1289 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c", |
| 1290 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c", |
| 1291 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c", |
| 1292 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c", |
| 1293 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c", |
| 1294 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c", |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 1295 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1296 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1297 | "src/f32-gavgpool-cw/neon-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1298 | "src/f32-gavgpool/7p7x-minmax-neon-c4.c", |
| 1299 | "src/f32-gavgpool/7x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1300 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1301 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c", |
| 1302 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1303 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1304 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c", |
| 1305 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c", |
| 1306 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c", |
| 1307 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c", |
| 1308 | "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1309 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c", |
| 1310 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1311 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c", |
| 1312 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1313 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c", |
| 1314 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1315 | "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 1316 | "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 1317 | "src/f32-gemm/gen/1x8s4-minmax-neon.c", |
| 1318 | "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 1319 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 1320 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 1321 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 1322 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 1323 | "src/f32-gemm/gen/4x8s4-minmax-neon.c", |
| 1324 | "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c", |
| 1325 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 1326 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 1327 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 1328 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 1329 | "src/f32-gemm/gen/6x8s4-minmax-neon.c", |
| 1330 | "src/f32-gemm/gen/8x8s4-minmax-neon.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1331 | "src/f32-hswish/gen/hswish-neon-x4.c", |
| 1332 | "src/f32-hswish/gen/hswish-neon-x8.c", |
| 1333 | "src/f32-hswish/gen/hswish-neon-x16.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1334 | "src/f32-ibilinear/gen/neon-c4.c", |
| 1335 | "src/f32-ibilinear/gen/neon-c8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1336 | "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1337 | "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1338 | "src/f32-igemm/gen/1x8s4-minmax-neon.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1339 | "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 1340 | "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1341 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1342 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 1343 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 1344 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 1345 | "src/f32-igemm/gen/4x8s4-minmax-neon.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1346 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 1347 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1348 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 1349 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1350 | "src/f32-igemm/gen/6x8s4-minmax-neon.c", |
| 1351 | "src/f32-igemm/gen/8x8s4-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1352 | "src/f32-maxpool/9p8x-minmax-neon-c4.c", |
| 1353 | "src/f32-pavgpool/9p8x-minmax-neon-c4.c", |
| 1354 | "src/f32-pavgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1355 | "src/f32-ppmm/gen/4x8-minmax-neon.c", |
| 1356 | "src/f32-ppmm/gen/8x8-minmax-neon.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1357 | "src/f32-prelu/gen/neon-1x4.c", |
| 1358 | "src/f32-prelu/gen/neon-1x8.c", |
| 1359 | "src/f32-prelu/gen/neon-1x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1360 | "src/f32-prelu/gen/neon-2x4.c", |
| 1361 | "src/f32-prelu/gen/neon-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1362 | "src/f32-prelu/gen/neon-2x16.c", |
| 1363 | "src/f32-prelu/gen/neon-4x4.c", |
| 1364 | "src/f32-prelu/gen/neon-4x8.c", |
| 1365 | "src/f32-prelu/gen/neon-4x16.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1366 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1367 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1368 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1369 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 1370 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1371 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1372 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 1373 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1374 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1375 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 1376 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1377 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 1378 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 1379 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 1380 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 1381 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 1382 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 1383 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 1384 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 1385 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 1386 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 1387 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 1388 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 1389 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1390 | "src/f32-relu/gen/neon-x4.c", |
| 1391 | "src/f32-relu/gen/neon-x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1392 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1393 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1394 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 1395 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 1396 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 1397 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 1398 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 1399 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 1400 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 1401 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 1402 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 1403 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 1404 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 1405 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1406 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 1407 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 1408 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 1409 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 1410 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 1411 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1412 | "src/f32-vbinary/gen/vadd-minmax-neon-x4.c", |
| 1413 | "src/f32-vbinary/gen/vadd-minmax-neon-x8.c", |
| 1414 | "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c", |
| 1415 | "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1416 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 1417 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 1418 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 1419 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 1420 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 1421 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 1422 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 1423 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1424 | "src/f32-vbinary/gen/vmul-minmax-neon-x4.c", |
| 1425 | "src/f32-vbinary/gen/vmul-minmax-neon-x8.c", |
| 1426 | "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c", |
| 1427 | "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c", |
| 1428 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c", |
| 1429 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1430 | "src/f32-vbinary/gen/vsqrdiff-neon-x4.c", |
| 1431 | "src/f32-vbinary/gen/vsqrdiff-neon-x8.c", |
| 1432 | "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c", |
| 1433 | "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1434 | "src/f32-vbinary/gen/vsub-minmax-neon-x4.c", |
| 1435 | "src/f32-vbinary/gen/vsub-minmax-neon-x8.c", |
| 1436 | "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c", |
| 1437 | "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1438 | "src/f32-vlrelu/gen/vlrelu-neon-x4.c", |
| 1439 | "src/f32-vlrelu/gen/vlrelu-neon-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1440 | "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c", |
| 1441 | "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1442 | "src/f32-vrnd/gen/vrndd-neon-x4.c", |
| 1443 | "src/f32-vrnd/gen/vrndd-neon-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1444 | "src/f32-vrnd/gen/vrndne-neon-x4.c", |
| 1445 | "src/f32-vrnd/gen/vrndne-neon-x8.c", |
| 1446 | "src/f32-vrnd/gen/vrndu-neon-x4.c", |
| 1447 | "src/f32-vrnd/gen/vrndu-neon-x8.c", |
| 1448 | "src/f32-vrnd/gen/vrndz-neon-x4.c", |
| 1449 | "src/f32-vrnd/gen/vrndz-neon-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1450 | "src/f32-vunary/gen/vabs-neon-x4.c", |
| 1451 | "src/f32-vunary/gen/vabs-neon-x8.c", |
| 1452 | "src/f32-vunary/gen/vneg-neon-x4.c", |
| 1453 | "src/f32-vunary/gen/vneg-neon-x8.c", |
| 1454 | "src/f32-vunary/gen/vsqr-neon-x4.c", |
| 1455 | "src/f32-vunary/gen/vsqr-neon-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1456 | "src/math/roundd-neon-addsub.c", |
| 1457 | "src/math/roundd-neon-cvt.c", |
| 1458 | "src/math/roundne-neon-addsub.c", |
| 1459 | "src/math/roundu-neon-addsub.c", |
| 1460 | "src/math/roundu-neon-cvt.c", |
| 1461 | "src/math/roundz-neon-addsub.c", |
| 1462 | "src/math/roundz-neon-cvt.c", |
| 1463 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
| 1464 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 1465 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 1466 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 1467 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 1468 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 1469 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
| 1470 | "src/math/sqrt-neon-nr1rsqrts.c", |
| 1471 | "src/math/sqrt-neon-nr2rsqrts.c", |
| 1472 | "src/math/sqrt-neon-nr3rsqrts.c", |
Marat Dukhan | 023bcf9 | 2020-08-10 12:40:50 -0700 | [diff] [blame] | 1473 | "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c", |
| 1474 | "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c", |
| 1475 | "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c", |
| 1476 | "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c", |
Marat Dukhan | 281262d | 2020-08-10 13:23:21 -0700 | [diff] [blame] | 1477 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c", |
| 1478 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c", |
| 1479 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c", |
| 1480 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1481 | "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c", |
| 1482 | "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c", |
| 1483 | "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c", |
| 1484 | "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c", |
Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1485 | "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c", |
Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1486 | "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1487 | "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c", |
Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1488 | "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c", |
| 1489 | "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c", |
Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1490 | "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1491 | "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c", |
Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1492 | "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1493 | "src/qs8-requantization/fp32-neon.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1494 | "src/qs8-requantization/precise-neon.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1495 | "src/qs8-requantization/q31-neon.c", |
Marat Dukhan | ba7b279 | 2020-09-02 14:26:45 -0700 | [diff] [blame] | 1496 | "src/qs8-vadd/gen/minmax-neon-ld64-x8.c", |
| 1497 | "src/qs8-vadd/gen/minmax-neon-ld64-x16.c", |
| 1498 | "src/qs8-vadd/gen/minmax-neon-ld64-x24.c", |
| 1499 | "src/qs8-vadd/gen/minmax-neon-ld64-x32.c", |
| 1500 | "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c", |
| 1501 | "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c", |
| 1502 | "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c", |
| 1503 | "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1504 | "src/qu8-avgpool/9p8x-minmax-neon-c8.c", |
| 1505 | "src/qu8-avgpool/9x-minmax-neon-c8.c", |
| 1506 | "src/qu8-dwconv/up8x9-minmax-neon.c", |
| 1507 | "src/qu8-gavgpool/7p7x-minmax-neon-c8.c", |
| 1508 | "src/qu8-gavgpool/7x-minmax-neon-c8.c", |
| 1509 | "src/qu8-gemm/4x8-minmax-neon.c", |
| 1510 | "src/qu8-gemm/8x8-minmax-neon.c", |
| 1511 | "src/qu8-igemm/4x8-minmax-neon.c", |
| 1512 | "src/qu8-igemm/8x8-minmax-neon.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1513 | "src/qu8-requantization/fp32-neon.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1514 | "src/qu8-requantization/precise-neon.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1515 | "src/qu8-requantization/q31-neon.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1516 | "src/qu8-vadd/minmax-neon.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1517 | "src/u8-clamp/neon-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1518 | "src/u8-maxpool/9p8x-minmax-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1519 | "src/u8-rmax/neon.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1520 | "src/x8-zip/x2-neon.c", |
| 1521 | "src/x8-zip/x3-neon.c", |
| 1522 | "src/x8-zip/x4-neon.c", |
| 1523 | "src/x8-zip/xm-neon.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1524 | "src/x32-fill/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1525 | "src/x32-packx/x4-neon-st4.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1526 | "src/x32-pad/neon.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1527 | "src/x32-unpool/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1528 | "src/x32-zip/x2-neon.c", |
| 1529 | "src/x32-zip/x3-neon.c", |
| 1530 | "src/x32-zip/x4-neon.c", |
| 1531 | "src/x32-zip/xm-neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1532 | ] |
| 1533 | |
| 1534 | NEONFMA_UKERNELS = [ |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1535 | "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", |
| 1536 | "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", |
| 1537 | "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", |
| 1538 | "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", |
| 1539 | "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", |
| 1540 | "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", |
| 1541 | "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", |
| 1542 | "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", |
| 1543 | "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", |
| 1544 | "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", |
| 1545 | "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", |
| 1546 | "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", |
| 1547 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c", |
| 1548 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c", |
| 1549 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c", |
| 1550 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c", |
| 1551 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c", |
| 1552 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c", |
| 1553 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c", |
| 1554 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c", |
| 1555 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c", |
| 1556 | "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1557 | "src/f32-gemm/gen/1x8s4-minmax-neonfma.c", |
| 1558 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1559 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1560 | "src/f32-gemm/gen/4x8s4-minmax-neonfma.c", |
| 1561 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1562 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1563 | "src/f32-gemm/gen/6x8s4-minmax-neonfma.c", |
| 1564 | "src/f32-gemm/gen/8x8s4-minmax-neonfma.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1565 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 1566 | "src/f32-ibilinear/gen/neonfma-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1567 | "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1568 | "src/f32-igemm/gen/1x8s4-minmax-neonfma.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1569 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1570 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1571 | "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1572 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1573 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1574 | "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", |
| 1575 | "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1576 | "src/f32-ppmm/gen/4x8-minmax-neonfma.c", |
| 1577 | "src/f32-ppmm/gen/8x8-minmax-neonfma.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1578 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1579 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1580 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1581 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 1582 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1583 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1584 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 1585 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1586 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1587 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 1588 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1589 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 1590 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 1591 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 1592 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 1593 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 1594 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 1595 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 1596 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 1597 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 1598 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 1599 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 1600 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 1601 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1602 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 1603 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 1604 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 1605 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 1606 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 1607 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1608 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 1609 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 1610 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 1611 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 1612 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 1613 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1614 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 1615 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 1616 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 1617 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 1618 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 1619 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1620 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 1621 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 1622 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 1623 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 1624 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 1625 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1626 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 1627 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 1628 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 1629 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 1630 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 1631 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1632 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 1633 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 1634 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 1635 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 1636 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 1637 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1638 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 1639 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 1640 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 1641 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 1642 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 1643 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 1644 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 1645 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 1646 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 1647 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 1648 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 1649 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 1650 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 1651 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 1652 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 1653 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 1654 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 1655 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1656 | "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c", |
| 1657 | "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1658 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c", |
| 1659 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c", |
| 1660 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c", |
| 1661 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c", |
| 1662 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c", |
| 1663 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c", |
| 1664 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c", |
| 1665 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c", |
| 1666 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c", |
| 1667 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c", |
| 1668 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c", |
| 1669 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c", |
| 1670 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c", |
| 1671 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c", |
| 1672 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c", |
| 1673 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c", |
| 1674 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c", |
| 1675 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c", |
| 1676 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c", |
| 1677 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c", |
Marat Dukhan | b7633f2 | 2020-11-20 16:34:56 -0800 | [diff] [blame] | 1678 | "src/math/exp-neonfma-rr2-lut64-p2.c", |
| 1679 | "src/math/exp-neonfma-rr2-p5.c", |
Marat Dukhan | 9dd119a | 2020-11-20 18:20:04 -0800 | [diff] [blame] | 1680 | "src/math/expminus-neonfma-rr2-lut64-p2.c", |
| 1681 | "src/math/expminus-neonfma-rr2-lut2048-p1.c", |
| 1682 | "src/math/expminus-neonfma-rr2-p5.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1683 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 1684 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 1685 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1686 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 1687 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 1688 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1689 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 1690 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 1691 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1692 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 1693 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 1694 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1695 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 1696 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 1697 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1698 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 1699 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 1700 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1701 | "src/math/sqrt-neonfma-nr1fma.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1702 | "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1703 | "src/math/sqrt-neonfma-nr2fma.c", |
| 1704 | "src/math/sqrt-neonfma-nr2fma1adj.c", |
| 1705 | "src/math/sqrt-neonfma-nr3fma.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1706 | ] |
| 1707 | |
| 1708 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1709 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1710 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1711 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1712 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1713 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1714 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1715 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1716 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1717 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1718 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c", |
| 1719 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c", |
| 1720 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c", |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 1721 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1722 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c", |
Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 1723 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c", |
| 1724 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c", |
| 1725 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c", |
| 1726 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c", |
| 1727 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c", |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 1728 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c", |
| 1729 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c", |
| 1730 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1731 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c", |
Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 1732 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1733 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c", |
| 1734 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c", |
| 1735 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1736 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c", |
| 1737 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c", |
| 1738 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c", |
| 1739 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1740 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1741 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c", |
| 1742 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1743 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1744 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1745 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c", |
Marat Dukhan | 149f0ea | 2020-10-26 12:50:33 -0700 | [diff] [blame] | 1746 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1747 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c", |
| 1748 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c", |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 1749 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c", |
| 1750 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c", |
| 1751 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c", |
| 1752 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c", |
| 1753 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c", |
| 1754 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c", |
| 1755 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c", |
| 1756 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c", |
Marat Dukhan | 30d4b25 | 2020-10-29 16:33:22 -0700 | [diff] [blame] | 1757 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1758 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1759 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c", |
| 1760 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c", |
| 1761 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c", |
| 1762 | "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c", |
| 1763 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c", |
| 1764 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c", |
| 1765 | "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1766 | "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1767 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1768 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1769 | "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c", |
| 1770 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1771 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
| 1772 | "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1773 | "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1774 | "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c", |
| 1775 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1776 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1777 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1778 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1779 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 1780 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 1781 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 1782 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 1783 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 1784 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 1785 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 1786 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 1787 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 1788 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 1789 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 1790 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1791 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 1792 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 1793 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 1794 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 1795 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 1796 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1797 | "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1798 | "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1799 | "src/f32-spmm/gen/4x1-minmax-neonfma.c", |
| 1800 | "src/f32-spmm/gen/4x2-minmax-neonfma.c", |
| 1801 | "src/f32-spmm/gen/4x4-minmax-neonfma.c", |
| 1802 | "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1803 | "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1804 | "src/f32-spmm/gen/8x1-minmax-neonfma.c", |
| 1805 | "src/f32-spmm/gen/8x2-minmax-neonfma.c", |
| 1806 | "src/f32-spmm/gen/8x4-minmax-neonfma.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1807 | "src/f32-spmm/gen/12x1-minmax-neonfma.c", |
| 1808 | "src/f32-spmm/gen/12x2-minmax-neonfma.c", |
| 1809 | "src/f32-spmm/gen/12x4-minmax-neonfma.c", |
| 1810 | "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1811 | "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1812 | "src/f32-spmm/gen/16x1-minmax-neonfma.c", |
| 1813 | "src/f32-spmm/gen/16x2-minmax-neonfma.c", |
| 1814 | "src/f32-spmm/gen/16x4-minmax-neonfma.c", |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 1815 | "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1816 | "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c", |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 1817 | "src/f32-spmm/gen/32x1-minmax-neonfma.c", |
| 1818 | "src/f32-spmm/gen/32x2-minmax-neonfma.c", |
| 1819 | "src/f32-spmm/gen/32x4-minmax-neonfma.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1820 | "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c", |
| 1821 | "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c", |
| 1822 | "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c", |
| 1823 | "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c", |
| 1824 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c", |
| 1825 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1826 | "src/f32-vsqrt/gen/neon-sqrt-x4.c", |
| 1827 | "src/f32-vsqrt/gen/neon-sqrt-x8.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1828 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1829 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1830 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1831 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1832 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1833 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1834 | ] |
| 1835 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1836 | NEONV8_UKERNELS = [ |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1837 | "src/f32-vrnd/gen/vrndd-neonv8-x4.c", |
| 1838 | "src/f32-vrnd/gen/vrndd-neonv8-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1839 | "src/f32-vrnd/gen/vrndne-neonv8-x4.c", |
| 1840 | "src/f32-vrnd/gen/vrndne-neonv8-x8.c", |
| 1841 | "src/f32-vrnd/gen/vrndu-neonv8-x4.c", |
| 1842 | "src/f32-vrnd/gen/vrndu-neonv8-x8.c", |
| 1843 | "src/f32-vrnd/gen/vrndz-neonv8-x4.c", |
| 1844 | "src/f32-vrnd/gen/vrndz-neonv8-x8.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1845 | "src/math/roundd-neonv8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1846 | "src/math/roundne-neonv8.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1847 | "src/math/roundu-neonv8.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1848 | "src/math/roundz-neonv8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1849 | ] |
| 1850 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1851 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1852 | "src/f16-clamp/gen/neonfp16arith-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1853 | "src/f16-clamp/gen/neonfp16arith-x16.c", |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 1854 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c", |
| 1855 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c", |
| 1856 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c", |
| 1857 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1858 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c", |
| 1859 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c", |
| 1860 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c", |
| 1861 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c", |
| 1862 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c", |
| 1863 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c", |
| 1864 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c", |
| 1865 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c", |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 1866 | "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c", |
| 1867 | "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1868 | "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c", |
| 1869 | "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c", |
| 1870 | "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c", |
| 1871 | "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c", |
| 1872 | "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c", |
| 1873 | "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c", |
| 1874 | "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c", |
| 1875 | "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c", |
| 1876 | "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1877 | "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1878 | "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1879 | "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1880 | "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1881 | "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1882 | "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1883 | "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1884 | "src/f16-hswish/gen/hswish-neonfp16arith-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1885 | "src/f16-hswish/gen/hswish-neonfp16arith-x16.c", |
| 1886 | "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1887 | "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1888 | "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1889 | "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1890 | "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1891 | "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1892 | "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1893 | "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1894 | "src/f16-prelu/gen/neonfp16arith-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1895 | "src/f16-prelu/gen/neonfp16arith-2x16.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1896 | "src/f16-relu/gen/neonfp16arith-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1897 | "src/f16-relu/gen/neonfp16arith-x16.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1898 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1899 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1900 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1901 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1902 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1903 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 1904 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1905 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c", |
| 1906 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c", |
| 1907 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c", |
| 1908 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c", |
| 1909 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c", |
| 1910 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c", |
| 1911 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c", |
| 1912 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c", |
| 1913 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c", |
| 1914 | "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c", |
| 1915 | "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c", |
| 1916 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c", |
| 1917 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c", |
| 1918 | "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c", |
| 1919 | "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c", |
| 1920 | "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c", |
| 1921 | "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c", |
| 1922 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c", |
| 1923 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c", |
| 1924 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c", |
| 1925 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c", |
| 1926 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c", |
| 1927 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c", |
| 1928 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c", |
| 1929 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c", |
| 1930 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c", |
| 1931 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c", |
| 1932 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c", |
| 1933 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c", |
| 1934 | "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c", |
| 1935 | "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1936 | ] |
| 1937 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1938 | NEONDOT_UKERNELS = [ |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1939 | "src/qs8-gemm/gen/1x8c4-minmax-neondot.c", |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1940 | "src/qs8-gemm/gen/1x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1941 | "src/qs8-gemm/gen/4x8c4-minmax-neondot.c", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1942 | "src/qs8-gemm/gen/4x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1943 | "src/qs8-gemm/gen/6x8c4-minmax-neondot.c", |
Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1944 | "src/qs8-gemm/gen/6x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1945 | "src/qs8-gemm/gen/8x8c4-minmax-neondot.c", |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1946 | "src/qs8-gemm/gen/8x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1947 | "src/qs8-gemm/gen/12x8c4-minmax-neondot.c", |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1948 | "src/qs8-igemm/gen/1x8c4-minmax-neondot.c", |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1949 | "src/qs8-igemm/gen/1x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1950 | "src/qs8-igemm/gen/4x8c4-minmax-neondot.c", |
Benoit Jacob | a05487f | 2020-08-14 13:37:21 -0700 | [diff] [blame] | 1951 | "src/qs8-igemm/gen/4x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1952 | "src/qs8-igemm/gen/6x8c4-minmax-neondot.c", |
Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1953 | "src/qs8-igemm/gen/6x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1954 | "src/qs8-igemm/gen/8x8c4-minmax-neondot.c", |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1955 | "src/qs8-igemm/gen/8x16c4-minmax-neondot.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1956 | "src/qs8-igemm/gen/12x8c4-minmax-neondot.c", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1957 | ] |
| 1958 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1959 | SSE_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1960 | "src/f32-avgpool/9p8x-minmax-sse-c4.c", |
| 1961 | "src/f32-avgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1962 | "src/f32-clamp/gen/sse-x4.c", |
| 1963 | "src/f32-clamp/gen/sse-x8.c", |
Erich Elsen | b123340 | 2020-06-08 15:53:15 -0700 | [diff] [blame] | 1964 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c", |
| 1965 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1966 | "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c", |
| 1967 | "src/f32-dwconv/gen/up4x4-minmax-sse.c", |
| 1968 | "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c", |
| 1969 | "src/f32-dwconv/gen/up4x9-minmax-sse.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1970 | "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c", |
| 1971 | "src/f32-dwconv/gen/up4x25-minmax-sse.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1972 | "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c", |
| 1973 | "src/f32-dwconv/gen/up8x4-minmax-sse.c", |
| 1974 | "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c", |
| 1975 | "src/f32-dwconv/gen/up8x9-minmax-sse.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1976 | "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c", |
| 1977 | "src/f32-dwconv/gen/up8x25-minmax-sse.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1978 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c", |
| 1979 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c", |
| 1980 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c", |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 1981 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1982 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c", |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 1983 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c", |
| 1984 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c", |
| 1985 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c", |
| 1986 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c", |
| 1987 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c", |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 1988 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c", |
| 1989 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c", |
| 1990 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1991 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c", |
Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 1992 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 1993 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c", |
| 1994 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c", |
| 1995 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c", |
Marat Dukhan | d050389 | 2020-10-30 08:22:04 -0700 | [diff] [blame] | 1996 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c", |
| 1997 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c", |
| 1998 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c", |
| 1999 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c", |
| 2000 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c", |
| 2001 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c", |
| 2002 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c", |
| 2003 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c", |
| 2004 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c", |
| 2005 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c", |
| 2006 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c", |
| 2007 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c", |
| 2008 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c", |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 2009 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c", |
| 2010 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c", |
| 2011 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c", |
| 2012 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c", |
| 2013 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c", |
| 2014 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c", |
| 2015 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c", |
| 2016 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c", |
Marat Dukhan | ccca214 | 2020-10-30 17:32:45 -0700 | [diff] [blame] | 2017 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c", |
Frank Barchard | 8ef44cd | 2020-11-03 12:30:23 -0800 | [diff] [blame] | 2018 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2019 | "src/f32-gavgpool-cw/sse-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2020 | "src/f32-gavgpool/7p7x-minmax-sse-c4.c", |
| 2021 | "src/f32-gavgpool/7x-minmax-sse-c4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2022 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c", |
| 2023 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c", |
| 2024 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c", |
| 2025 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c", |
| 2026 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c", |
| 2027 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2028 | "src/f32-gemm/gen/1x8-minmax-sse-dup.c", |
| 2029 | "src/f32-gemm/gen/1x8-minmax-sse-load1.c", |
| 2030 | "src/f32-gemm/gen/1x8s4-minmax-sse.c", |
| 2031 | "src/f32-gemm/gen/4x2c4-minmax-sse.c", |
| 2032 | "src/f32-gemm/gen/4x8-minmax-sse-dup.c", |
| 2033 | "src/f32-gemm/gen/4x8-minmax-sse-load1.c", |
| 2034 | "src/f32-gemm/gen/4x8s4-minmax-sse.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2035 | "src/f32-hswish/gen/hswish-sse-x4.c", |
| 2036 | "src/f32-hswish/gen/hswish-sse-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2037 | "src/f32-ibilinear/gen/sse-c4.c", |
| 2038 | "src/f32-ibilinear/gen/sse-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2039 | "src/f32-igemm/gen/1x8-minmax-sse-dup.c", |
| 2040 | "src/f32-igemm/gen/1x8-minmax-sse-load1.c", |
| 2041 | "src/f32-igemm/gen/1x8s4-minmax-sse.c", |
| 2042 | "src/f32-igemm/gen/4x2c4-minmax-sse.c", |
| 2043 | "src/f32-igemm/gen/4x8-minmax-sse-dup.c", |
| 2044 | "src/f32-igemm/gen/4x8-minmax-sse-load1.c", |
| 2045 | "src/f32-igemm/gen/4x8s4-minmax-sse.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2046 | "src/f32-maxpool/9p8x-minmax-sse-c4.c", |
| 2047 | "src/f32-pavgpool/9p8x-minmax-sse-c4.c", |
| 2048 | "src/f32-pavgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2049 | "src/f32-ppmm/gen/4x8-minmax-sse.c", |
Marat Dukhan | 39b5e94 | 2020-06-24 15:03:48 -0700 | [diff] [blame] | 2050 | "src/f32-prelu/gen/sse-2x4.c", |
| 2051 | "src/f32-prelu/gen/sse-2x8.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2052 | "src/f32-relu/gen/sse-x4.c", |
| 2053 | "src/f32-relu/gen/sse-x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2054 | "src/f32-rmax/sse.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 2055 | "src/f32-spmm/gen/4x1-minmax-sse.c", |
| 2056 | "src/f32-spmm/gen/8x1-minmax-sse.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 2057 | "src/f32-spmm/gen/16x1-minmax-sse.c", |
Frank Barchard | 846c0c6 | 2020-10-26 15:01:39 -0700 | [diff] [blame] | 2058 | "src/f32-spmm/gen/32x1-minmax-sse.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2059 | "src/f32-vbinary/gen/vadd-minmax-sse-x4.c", |
| 2060 | "src/f32-vbinary/gen/vadd-minmax-sse-x8.c", |
| 2061 | "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c", |
| 2062 | "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c", |
| 2063 | "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c", |
| 2064 | "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c", |
| 2065 | "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c", |
| 2066 | "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 2067 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 2068 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 2069 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 2070 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 2071 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 2072 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 2073 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 2074 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2075 | "src/f32-vbinary/gen/vmul-minmax-sse-x4.c", |
| 2076 | "src/f32-vbinary/gen/vmul-minmax-sse-x8.c", |
| 2077 | "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c", |
| 2078 | "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c", |
| 2079 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c", |
| 2080 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c", |
| 2081 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c", |
| 2082 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2083 | "src/f32-vbinary/gen/vsqrdiff-sse-x4.c", |
| 2084 | "src/f32-vbinary/gen/vsqrdiff-sse-x8.c", |
| 2085 | "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c", |
| 2086 | "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2087 | "src/f32-vbinary/gen/vsub-minmax-sse-x4.c", |
| 2088 | "src/f32-vbinary/gen/vsub-minmax-sse-x8.c", |
| 2089 | "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c", |
| 2090 | "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2091 | "src/f32-vlrelu/gen/vlrelu-sse-x4.c", |
| 2092 | "src/f32-vlrelu/gen/vlrelu-sse-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2093 | "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c", |
| 2094 | "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2095 | "src/f32-vsqrt/gen/sse-sqrt-x4.c", |
| 2096 | "src/f32-vsqrt/gen/sse-sqrt-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2097 | "src/f32-vunary/gen/vabs-sse-x4.c", |
| 2098 | "src/f32-vunary/gen/vabs-sse-x8.c", |
| 2099 | "src/f32-vunary/gen/vneg-sse-x4.c", |
| 2100 | "src/f32-vunary/gen/vneg-sse-x8.c", |
| 2101 | "src/f32-vunary/gen/vsqr-sse-x4.c", |
| 2102 | "src/f32-vunary/gen/vsqr-sse-x8.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 2103 | "src/math/roundd-sse-addsub.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2104 | "src/math/roundne-sse-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 2105 | "src/math/roundu-sse-addsub.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 2106 | "src/math/roundz-sse-addsub.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2107 | "src/math/sqrt-sse-hh1mac.c", |
| 2108 | "src/math/sqrt-sse-nr1mac.c", |
| 2109 | "src/math/sqrt-sse-nr2mac.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2110 | "src/x32-fill/sse.c", |
| 2111 | "src/x32-packx/x4-sse.c", |
| 2112 | "src/x32-pad/sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2113 | ] |
| 2114 | |
| 2115 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 2116 | "src/f32-argmaxpool/4x-sse2-c4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2117 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 2118 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2119 | "src/f32-prelu/gen/sse2-2x4.c", |
| 2120 | "src/f32-prelu/gen/sse2-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 2121 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 2122 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2123 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 2124 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 2125 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2126 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 2127 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 2128 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2129 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 2130 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 2131 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2132 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 2133 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c", |
| 2134 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c", |
| 2135 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c", |
| 2136 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c", |
| 2137 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c", |
| 2138 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2139 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2140 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2141 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2142 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2143 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 2144 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2145 | "src/f32-vlrelu/gen/vlrelu-sse2-x4.c", |
| 2146 | "src/f32-vlrelu/gen/vlrelu-sse2-x8.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2147 | "src/f32-vrnd/gen/vrndd-sse2-x4.c", |
| 2148 | "src/f32-vrnd/gen/vrndd-sse2-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2149 | "src/f32-vrnd/gen/vrndne-sse2-x4.c", |
| 2150 | "src/f32-vrnd/gen/vrndne-sse2-x8.c", |
| 2151 | "src/f32-vrnd/gen/vrndu-sse2-x4.c", |
| 2152 | "src/f32-vrnd/gen/vrndu-sse2-x8.c", |
| 2153 | "src/f32-vrnd/gen/vrndz-sse2-x4.c", |
| 2154 | "src/f32-vrnd/gen/vrndz-sse2-x8.c", |
Marat Dukhan | b7633f2 | 2020-11-20 16:34:56 -0800 | [diff] [blame] | 2155 | "src/math/exp-sse2-rr2-lut64-p2.c", |
| 2156 | "src/math/exp-sse2-rr2-p5.c", |
Marat Dukhan | d6d321e | 2020-11-20 18:35:55 -0800 | [diff] [blame] | 2157 | "src/math/expm1minus-sse2-rr2-p5.c", |
Marat Dukhan | a438aca | 2020-11-20 15:45:01 -0800 | [diff] [blame] | 2158 | "src/math/expm1minus-sse2-rr2-p6.c", |
Frank Barchard | 3b80045 | 2020-11-22 12:12:35 -0800 | [diff] [blame] | 2159 | "src/math/expminus-sse2-rr2-p5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2160 | "src/math/roundd-sse2-cvt.c", |
| 2161 | "src/math/roundne-sse2-cvt.c", |
| 2162 | "src/math/roundu-sse2-cvt.c", |
| 2163 | "src/math/roundz-sse2-cvt.c", |
| 2164 | "src/math/sigmoid-sse2-rr2-lut64-p2-div.c", |
| 2165 | "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c", |
| 2166 | "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c", |
| 2167 | "src/math/sigmoid-sse2-rr2-p5-div.c", |
| 2168 | "src/math/sigmoid-sse2-rr2-p5-nr1.c", |
| 2169 | "src/math/sigmoid-sse2-rr2-p5-nr2.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2170 | "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c", |
| 2171 | "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c", |
| 2172 | "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c", |
Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2173 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c", |
| 2174 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c", |
| 2175 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2176 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c", |
| 2177 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c", |
| 2178 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2179 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2180 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2181 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2182 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2183 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2184 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2185 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c", |
| 2186 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2187 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2188 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c", |
| 2189 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2190 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2191 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c", |
| 2192 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c", |
| 2193 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2194 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2195 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2196 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2197 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2198 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2199 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2200 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2201 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2202 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c", |
| 2203 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2204 | "src/qs8-requantization/fp32-sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2205 | "src/qs8-requantization/precise-sse2.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2206 | "src/qs8-requantization/q31-sse2.c", |
Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 2207 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c", |
| 2208 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c", |
| 2209 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c", |
| 2210 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c", |
Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 2211 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c", |
| 2212 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c", |
| 2213 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c", |
| 2214 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2215 | "src/qu8-avgpool/9p8x-minmax-sse2-c8.c", |
| 2216 | "src/qu8-avgpool/9x-minmax-sse2-c8.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2217 | "src/qu8-dwconv/up8x9-minmax-sse2.c", |
| 2218 | "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c", |
| 2219 | "src/qu8-gavgpool/7x-minmax-sse2-c8.c", |
| 2220 | "src/qu8-gemm/2x4c8-minmax-sse2.c", |
| 2221 | "src/qu8-gemm/4x4c2-minmax-sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2222 | "src/qu8-igemm/4x4c2-minmax-sse2.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2223 | "src/qu8-requantization/fp32-sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2224 | "src/qu8-requantization/precise-sse2.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2225 | "src/qu8-requantization/q31-sse2.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2226 | "src/qu8-vadd/minmax-sse2.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2227 | "src/u8-clamp/sse2-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2228 | "src/u8-maxpool/9p8x-minmax-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2229 | "src/u8-rmax/sse2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2230 | "src/x8-zip/x2-sse2.c", |
| 2231 | "src/x8-zip/x3-sse2.c", |
| 2232 | "src/x8-zip/x4-sse2.c", |
| 2233 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 2234 | "src/x32-unpool/sse2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2235 | "src/x32-zip/x2-sse2.c", |
| 2236 | "src/x32-zip/x3-sse2.c", |
| 2237 | "src/x32-zip/x4-sse2.c", |
| 2238 | "src/x32-zip/xm-sse2.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2239 | ] |
| 2240 | |
| 2241 | SSSE3_UKERNELS = [ |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 2242 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c", |
| 2243 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c", |
| 2244 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c", |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 2245 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 2246 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c", |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 2247 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c", |
| 2248 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c", |
| 2249 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c", |
| 2250 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c", |
| 2251 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2252 | "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c", |
| 2253 | "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c", |
| 2254 | "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c", |
Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2255 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c", |
| 2256 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c", |
| 2257 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2258 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c", |
| 2259 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c", |
| 2260 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2261 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2262 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2263 | "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2264 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2265 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2266 | "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2267 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c", |
| 2268 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2269 | "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2270 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c", |
| 2271 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2272 | "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2273 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c", |
| 2274 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c", |
| 2275 | "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2276 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2277 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2278 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2279 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2280 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2281 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2282 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2283 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2284 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c", |
| 2285 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 2286 | "src/qs8-requantization/precise-ssse3.c", |
| 2287 | "src/qs8-requantization/q31-ssse3.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2288 | "src/qu8-requantization/precise-ssse3.c", |
| 2289 | "src/qu8-requantization/q31-ssse3.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2290 | ] |
| 2291 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2292 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2293 | "src/f32-prelu/gen/sse41-2x4.c", |
| 2294 | "src/f32-prelu/gen/sse41-2x8.c", |
Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 2295 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c", |
| 2296 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c", |
| 2297 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c", |
| 2298 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c", |
| 2299 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c", |
| 2300 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2301 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2302 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2303 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2304 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2305 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 2306 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2307 | "src/f32-vlrelu/gen/vlrelu-sse41-x4.c", |
| 2308 | "src/f32-vlrelu/gen/vlrelu-sse41-x8.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2309 | "src/f32-vrnd/gen/vrndd-sse41-x4.c", |
| 2310 | "src/f32-vrnd/gen/vrndd-sse41-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2311 | "src/f32-vrnd/gen/vrndne-sse41-x4.c", |
| 2312 | "src/f32-vrnd/gen/vrndne-sse41-x8.c", |
| 2313 | "src/f32-vrnd/gen/vrndu-sse41-x4.c", |
| 2314 | "src/f32-vrnd/gen/vrndu-sse41-x8.c", |
| 2315 | "src/f32-vrnd/gen/vrndz-sse41-x4.c", |
| 2316 | "src/f32-vrnd/gen/vrndz-sse41-x8.c", |
| 2317 | "src/math/roundd-sse41.c", |
| 2318 | "src/math/roundne-sse41.c", |
| 2319 | "src/math/roundu-sse41.c", |
| 2320 | "src/math/roundz-sse41.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2321 | "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c", |
| 2322 | "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c", |
| 2323 | "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c", |
Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2324 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c", |
| 2325 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c", |
| 2326 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2327 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c", |
| 2328 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c", |
| 2329 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2330 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2331 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2332 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2333 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2334 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2335 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2336 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c", |
| 2337 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2338 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2339 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c", |
| 2340 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2341 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2342 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c", |
| 2343 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c", |
| 2344 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2345 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2346 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2347 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2348 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2349 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2350 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2351 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2352 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2353 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c", |
| 2354 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 2355 | "src/qs8-requantization/fp32-sse4.c", |
| 2356 | "src/qs8-requantization/precise-sse4.c", |
| 2357 | "src/qs8-requantization/q31-sse4.c", |
Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 2358 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c", |
| 2359 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c", |
| 2360 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c", |
| 2361 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c", |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2362 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c", |
| 2363 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c", |
| 2364 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c", |
| 2365 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c", |
Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 2366 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c", |
| 2367 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c", |
| 2368 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c", |
| 2369 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c", |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2370 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c", |
| 2371 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c", |
| 2372 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c", |
| 2373 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2374 | "src/qu8-requantization/precise-sse4.c", |
| 2375 | "src/qu8-requantization/q31-sse4.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2376 | ] |
| 2377 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2378 | AVX_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2379 | "src/f32-clamp/gen/avx-x8.c", |
| 2380 | "src/f32-clamp/gen/avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2381 | "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c", |
| 2382 | "src/f32-dwconv/gen/up8x4-minmax-avx.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2383 | "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c", |
| 2384 | "src/f32-dwconv/gen/up8x9-minmax-avx.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2385 | "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c", |
| 2386 | "src/f32-dwconv/gen/up8x25-minmax-avx.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2387 | "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c", |
| 2388 | "src/f32-dwconv/gen/up16x4-minmax-avx.c", |
| 2389 | "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c", |
| 2390 | "src/f32-dwconv/gen/up16x9-minmax-avx.c", |
| 2391 | "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c", |
| 2392 | "src/f32-dwconv/gen/up16x25-minmax-avx.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2393 | "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2394 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c", |
| 2395 | "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2396 | "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2397 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2398 | "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2399 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2400 | "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c", |
| 2401 | "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c", |
| 2402 | "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c", |
| 2403 | "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c", |
| 2404 | "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c", |
| 2405 | "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c", |
| 2406 | "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c", |
| 2407 | "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c", |
| 2408 | "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c", |
| 2409 | "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c", |
| 2410 | "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2411 | "src/f32-hswish/gen/hswish-avx-x8.c", |
| 2412 | "src/f32-hswish/gen/hswish-avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2413 | "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2414 | "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c", |
| 2415 | "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2416 | "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2417 | "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2418 | "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2419 | "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2420 | "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c", |
| 2421 | "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2422 | "src/f32-prelu/gen/avx-2x8.c", |
| 2423 | "src/f32-prelu/gen/avx-2x16.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2424 | "src/f32-relu/gen/avx-x8.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2425 | "src/f32-relu/gen/avx-x16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2426 | "src/f32-rmax/avx.c", |
T.J. Alumbaugh | dc2b29c | 2020-10-14 13:56:08 -0700 | [diff] [blame] | 2427 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c", |
| 2428 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c", |
| 2429 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c", |
| 2430 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c", |
| 2431 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c", |
| 2432 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c", |
| 2433 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c", |
| 2434 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c", |
| 2435 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c", |
| 2436 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c", |
| 2437 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c", |
| 2438 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c", |
| 2439 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c", |
| 2440 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c", |
| 2441 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c", |
| 2442 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c", |
| 2443 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c", |
| 2444 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c", |
| 2445 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c", |
| 2446 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2447 | "src/f32-vbinary/gen/vadd-minmax-avx-x8.c", |
| 2448 | "src/f32-vbinary/gen/vadd-minmax-avx-x16.c", |
| 2449 | "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c", |
| 2450 | "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c", |
| 2451 | "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c", |
| 2452 | "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c", |
| 2453 | "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c", |
| 2454 | "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2455 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 2456 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 2457 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 2458 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 2459 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 2460 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 2461 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 2462 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2463 | "src/f32-vbinary/gen/vmul-minmax-avx-x8.c", |
| 2464 | "src/f32-vbinary/gen/vmul-minmax-avx-x16.c", |
| 2465 | "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c", |
| 2466 | "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c", |
| 2467 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c", |
| 2468 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c", |
| 2469 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c", |
| 2470 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2471 | "src/f32-vbinary/gen/vsqrdiff-avx-x8.c", |
| 2472 | "src/f32-vbinary/gen/vsqrdiff-avx-x16.c", |
| 2473 | "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c", |
| 2474 | "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2475 | "src/f32-vbinary/gen/vsub-minmax-avx-x8.c", |
| 2476 | "src/f32-vbinary/gen/vsub-minmax-avx-x16.c", |
| 2477 | "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c", |
| 2478 | "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2479 | "src/f32-vlrelu/gen/vlrelu-avx-x8.c", |
| 2480 | "src/f32-vlrelu/gen/vlrelu-avx-x16.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2481 | "src/f32-vrnd/gen/vrndd-avx-x8.c", |
| 2482 | "src/f32-vrnd/gen/vrndd-avx-x16.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2483 | "src/f32-vrnd/gen/vrndne-avx-x8.c", |
| 2484 | "src/f32-vrnd/gen/vrndne-avx-x16.c", |
| 2485 | "src/f32-vrnd/gen/vrndu-avx-x8.c", |
| 2486 | "src/f32-vrnd/gen/vrndu-avx-x16.c", |
| 2487 | "src/f32-vrnd/gen/vrndz-avx-x8.c", |
| 2488 | "src/f32-vrnd/gen/vrndz-avx-x16.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 2489 | "src/f32-vscale/avx-x32.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2490 | "src/f32-vsqrt/gen/avx-sqrt-x8.c", |
| 2491 | "src/f32-vsqrt/gen/avx-sqrt-x16.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2492 | "src/f32-vunary/gen/vabs-avx-x8.c", |
| 2493 | "src/f32-vunary/gen/vabs-avx-x16.c", |
| 2494 | "src/f32-vunary/gen/vneg-avx-x8.c", |
| 2495 | "src/f32-vunary/gen/vneg-avx-x16.c", |
| 2496 | "src/f32-vunary/gen/vsqr-avx-x8.c", |
| 2497 | "src/f32-vunary/gen/vsqr-avx-x16.c", |
Marat Dukhan | b7633f2 | 2020-11-20 16:34:56 -0800 | [diff] [blame] | 2498 | "src/math/exp-avx-rr2-p5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2499 | "src/math/sigmoid-avx-rr2-lut64-p2-div.c", |
Marat Dukhan | 2e4727b | 2020-09-24 00:48:39 -0700 | [diff] [blame] | 2500 | "src/math/sigmoid-avx-rr2-p5-div.c", |
Marat Dukhan | 39ab16f | 2020-10-16 14:30:35 -0700 | [diff] [blame] | 2501 | "src/math/sigmoid-avx-rr2-p5-nr1.c", |
| 2502 | "src/math/sigmoid-avx-rr2-p5-nr2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2503 | ] |
| 2504 | |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2505 | XOP_UKERNELS = [ |
| 2506 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2507 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2508 | "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2509 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2510 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2511 | "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2512 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c", |
| 2513 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2514 | "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2515 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c", |
| 2516 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2517 | "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2518 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c", |
| 2519 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c", |
| 2520 | "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2521 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2522 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2523 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2524 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2525 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2526 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2527 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2528 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2529 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c", |
| 2530 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c", |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2531 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c", |
| 2532 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c", |
| 2533 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c", |
| 2534 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c", |
| 2535 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c", |
| 2536 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c", |
| 2537 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c", |
| 2538 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2539 | ] |
| 2540 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2541 | FMA3_UKERNELS = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2542 | "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c", |
| 2543 | "src/f32-dwconv/gen/up8x4-minmax-fma3.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2544 | "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c", |
| 2545 | "src/f32-dwconv/gen/up8x9-minmax-fma3.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2546 | "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c", |
| 2547 | "src/f32-dwconv/gen/up8x25-minmax-fma3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2548 | "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c", |
| 2549 | "src/f32-dwconv/gen/up16x4-minmax-fma3.c", |
| 2550 | "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c", |
| 2551 | "src/f32-dwconv/gen/up16x9-minmax-fma3.c", |
| 2552 | "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c", |
| 2553 | "src/f32-dwconv/gen/up16x25-minmax-fma3.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2554 | "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2555 | "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c", |
| 2556 | "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c", |
| 2557 | "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c", |
| 2558 | "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2559 | "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2560 | "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c", |
| 2561 | "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2562 | "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2563 | "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c", |
| 2564 | "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2565 | "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c", |
| 2566 | "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c", |
| 2567 | "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2568 | "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c", |
| 2569 | "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c", |
| 2570 | "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 2571 | "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c", |
| 2572 | "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 2573 | "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c", |
| 2574 | "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c", |
| 2575 | "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 2576 | "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c", |
| 2577 | "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c", |
| 2578 | "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| 2579 | "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c", |
| 2580 | "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c", |
| 2581 | "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2582 | "src/f32-hswish/gen/hswish-fma3-x8.c", |
| 2583 | "src/f32-hswish/gen/hswish-fma3-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2584 | "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2585 | "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c", |
| 2586 | "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 2587 | "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c", |
| 2588 | "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2589 | "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2590 | "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c", |
| 2591 | "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2592 | "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2593 | "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c", |
| 2594 | "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2595 | "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c", |
| 2596 | "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c", |
| 2597 | "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2598 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c", |
| 2599 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c", |
| 2600 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c", |
| 2601 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c", |
| 2602 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c", |
| 2603 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c", |
| 2604 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c", |
| 2605 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2606 | "src/math/sqrt-fma3-nr1fma.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2607 | "src/math/sqrt-fma3-nr1fma1adj.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2608 | "src/math/sqrt-fma3-nr2fma.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2609 | ] |
| 2610 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2611 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2612 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 2613 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2614 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2615 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2616 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2617 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 2618 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2619 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2620 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 2621 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 2622 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2623 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2624 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 2625 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2626 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2627 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2628 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2629 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 2630 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2631 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2632 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 2633 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 2634 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2635 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2636 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 2637 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2638 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2639 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2640 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2641 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 2642 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2643 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2644 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 2645 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 2646 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2647 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 2648 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 2649 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 2650 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 2651 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 2652 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 2653 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 2654 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 2655 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 2656 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 2657 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 2658 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 2659 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 2660 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 2661 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 2662 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 2663 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 2664 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 2665 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 2666 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 2667 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 2668 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 2669 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 2670 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 2671 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 2672 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 2673 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 2674 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 2675 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 2676 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 2677 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2678 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 2679 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 2680 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 2681 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 2682 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 2683 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 2684 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 2685 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 2686 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 2687 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 2688 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 2689 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 2690 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 2691 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 2692 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 2693 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 2694 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 2695 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 2696 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 2697 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 2698 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 2699 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 2700 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 2701 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | b7633f2 | 2020-11-20 16:34:56 -0800 | [diff] [blame] | 2702 | "src/math/exp-avx2-rr2-lut8-p3-perm.c", |
| 2703 | "src/math/exp-avx2-rr2-lut8-p4-perm.c", |
| 2704 | "src/math/exp-avx2-rr2-p5.c", |
Marat Dukhan | 9dd119a | 2020-11-20 18:20:04 -0800 | [diff] [blame] | 2705 | "src/math/expminus-avx2-rr2-p5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2706 | "src/math/extexp-avx2-p5.c", |
| 2707 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c", |
| 2708 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c", |
| 2709 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c", |
| 2710 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c", |
| 2711 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 2712 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
| 2713 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 2714 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c", |
| 2715 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c", |
| 2716 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c", |
| 2717 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c", |
| 2718 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 2719 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 2720 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2721 | "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2722 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c", |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2723 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c", |
| 2724 | "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2725 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c", |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2726 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c", |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 2727 | "src/qs8-gemm/gen/1x8c8-minmax-avx2.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2728 | "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2729 | "src/qs8-gemm/gen/2x8c8-minmax-avx2.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2730 | "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2731 | "src/qs8-gemm/gen/3x8c8-minmax-avx2.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2732 | "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c", |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 2733 | "src/qs8-igemm/gen/1x8c8-minmax-avx2.c", |
| 2734 | "src/qs8-igemm/gen/2x8c8-minmax-avx2.c", |
| 2735 | "src/qs8-igemm/gen/3x8c8-minmax-avx2.c", |
Marat Dukhan | e6dc0b6 | 2020-09-08 23:57:14 -0700 | [diff] [blame] | 2736 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2737 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2738 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2739 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c", |
| 2740 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2741 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2742 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2743 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2744 | ] |
| 2745 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2746 | AVX512F_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2747 | "src/f32-clamp/gen/avx512f-x16.c", |
| 2748 | "src/f32-clamp/gen/avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2749 | "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c", |
| 2750 | "src/f32-dwconv/gen/up16x4-minmax-avx512f.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2751 | "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c", |
| 2752 | "src/f32-dwconv/gen/up16x9-minmax-avx512f.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2753 | "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c", |
| 2754 | "src/f32-dwconv/gen/up16x25-minmax-avx512f.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2755 | "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c", |
| 2756 | "src/f32-dwconv/gen/up32x4-minmax-avx512f.c", |
| 2757 | "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c", |
| 2758 | "src/f32-dwconv/gen/up32x9-minmax-avx512f.c", |
| 2759 | "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c", |
| 2760 | "src/f32-dwconv/gen/up32x25-minmax-avx512f.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2761 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c", |
| 2762 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c", |
| 2763 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c", |
| 2764 | "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c", |
| 2765 | "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c", |
| 2766 | "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2767 | "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2768 | "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2769 | "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2770 | "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2771 | "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2772 | "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2773 | "src/f32-hswish/gen/hswish-avx512f-x16.c", |
| 2774 | "src/f32-hswish/gen/hswish-avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2775 | "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2776 | "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2777 | "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2778 | "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2779 | "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2780 | "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2781 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 2782 | "src/f32-prelu/gen/avx512f-2x32.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2783 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2784 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2785 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2786 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2787 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2788 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2789 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2790 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2791 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2792 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2793 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2794 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2795 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2796 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2797 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2798 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2799 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2800 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2801 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2802 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2803 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2804 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2805 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2806 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2807 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2808 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2809 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2810 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2811 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2812 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2813 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2814 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2815 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2816 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2817 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2818 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2819 | "src/f32-relu/gen/avx512f-x16.c", |
| 2820 | "src/f32-relu/gen/avx512f-x32.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2821 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 2822 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c", |
| 2823 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c", |
| 2824 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c", |
| 2825 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c", |
| 2826 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c", |
| 2827 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c", |
| 2828 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c", |
| 2829 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c", |
| 2830 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c", |
| 2831 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c", |
| 2832 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c", |
| 2833 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c", |
| 2834 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c", |
| 2835 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c", |
| 2836 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c", |
| 2837 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2838 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c", |
| 2839 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c", |
| 2840 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c", |
| 2841 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c", |
| 2842 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c", |
| 2843 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c", |
| 2844 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c", |
| 2845 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c", |
| 2846 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c", |
| 2847 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c", |
| 2848 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c", |
| 2849 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c", |
| 2850 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c", |
| 2851 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c", |
| 2852 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c", |
| 2853 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c", |
Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 2854 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c", |
| 2855 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c", |
| 2856 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c", |
| 2857 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c", |
| 2858 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c", |
| 2859 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c", |
| 2860 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c", |
| 2861 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c", |
| 2862 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c", |
| 2863 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c", |
| 2864 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c", |
| 2865 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c", |
| 2866 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c", |
| 2867 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c", |
| 2868 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c", |
| 2869 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2870 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c", |
| 2871 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c", |
| 2872 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c", |
| 2873 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c", |
| 2874 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c", |
| 2875 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c", |
| 2876 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c", |
| 2877 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2878 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 2879 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 2880 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 2881 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 2882 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 2883 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 2884 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 2885 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2886 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c", |
| 2887 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c", |
| 2888 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c", |
| 2889 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c", |
| 2890 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c", |
| 2891 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c", |
| 2892 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c", |
| 2893 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2894 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c", |
| 2895 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c", |
| 2896 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c", |
| 2897 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2898 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c", |
| 2899 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c", |
| 2900 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c", |
| 2901 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2902 | "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c", |
| 2903 | "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2904 | "src/f32-vrnd/gen/vrndd-avx512f-x16.c", |
| 2905 | "src/f32-vrnd/gen/vrndd-avx512f-x32.c", |
| 2906 | "src/f32-vrnd/gen/vrndne-avx512f-x16.c", |
| 2907 | "src/f32-vrnd/gen/vrndne-avx512f-x32.c", |
| 2908 | "src/f32-vrnd/gen/vrndu-avx512f-x16.c", |
| 2909 | "src/f32-vrnd/gen/vrndu-avx512f-x32.c", |
| 2910 | "src/f32-vrnd/gen/vrndz-avx512f-x16.c", |
| 2911 | "src/f32-vrnd/gen/vrndz-avx512f-x32.c", |
Frank Barchard | beca652 | 2020-10-30 22:34:35 -0700 | [diff] [blame] | 2912 | "src/f32-vscale/avx512f-x64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2913 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 2914 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 2915 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 2916 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 2917 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 2918 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 2919 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 2920 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2921 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2922 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2923 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 2924 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2925 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 2926 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 2927 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 2928 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 2929 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 2930 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 2931 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 2932 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 2933 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 2934 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 2935 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 2936 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2937 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c", |
| 2938 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c", |
| 2939 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c", |
| 2940 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c", |
| 2941 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c", |
| 2942 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c", |
| 2943 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c", |
| 2944 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2945 | "src/f32-vunary/gen/vabs-avx512f-x16.c", |
| 2946 | "src/f32-vunary/gen/vabs-avx512f-x32.c", |
| 2947 | "src/f32-vunary/gen/vneg-avx512f-x16.c", |
| 2948 | "src/f32-vunary/gen/vneg-avx512f-x32.c", |
| 2949 | "src/f32-vunary/gen/vsqr-avx512f-x16.c", |
| 2950 | "src/f32-vunary/gen/vsqr-avx512f-x32.c", |
Marat Dukhan | b7633f2 | 2020-11-20 16:34:56 -0800 | [diff] [blame] | 2951 | "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c", |
| 2952 | "src/math/exp-avx512f-rr2-lut16-p3-perm.c", |
| 2953 | "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c", |
| 2954 | "src/math/exp-avx512f-rr2-lut32-p2-perm2.c", |
| 2955 | "src/math/exp-avx512f-rr2-p5-scalef.c", |
| 2956 | "src/math/exp-avx512f-rr2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2957 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2958 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2959 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2960 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2961 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2962 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2963 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c", |
Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2964 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c", |
Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2965 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2966 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c", |
| 2967 | "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c", |
| 2968 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c", |
| 2969 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c", |
| 2970 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c", |
| 2971 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c", |
| 2972 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c", |
| 2973 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c", |
| 2974 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c", |
| 2975 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c", |
Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2976 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c", |
Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2977 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2978 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c", |
| 2979 | "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c", |
| 2980 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c", |
| 2981 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2982 | "src/math/sqrt-avx512f-nr1fma.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2983 | "src/math/sqrt-avx512f-nr1fma1adj.c", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2984 | "src/math/sqrt-avx512f-nr2fma.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2985 | ] |
| 2986 | |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 2987 | AVX512SKX_UKERNELS = [ |
Marat Dukhan | 2ffc5e6 | 2020-09-06 22:33:38 -0700 | [diff] [blame] | 2988 | "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c", |
| 2989 | "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 2990 | "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c", |
| 2991 | "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c", |
| 2992 | "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c", |
| 2993 | "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c", |
| 2994 | "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c", |
| 2995 | "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c", |
| 2996 | "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c", |
| 2997 | "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c", |
| 2998 | ] |
| 2999 | |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3000 | WASM32_ASM_UKERNELS = [ |
| 3001 | "src/f32-relu/wasm_shr_x1.S", |
| 3002 | "src/f32-relu/wasm_shr_x2.S", |
| 3003 | "src/f32-relu/wasm_shr_x4.S", |
| 3004 | ] |
| 3005 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3006 | AARCH32_ASM_UKERNELS = [ |
Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 3007 | "src/f32-gemm/4x4-aarch32-vfp-ld64.S", |
Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 3008 | "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3009 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 3010 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 3011 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3012 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 3013 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 3014 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3015 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 3016 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 3017 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
| 3018 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 3019 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| 3020 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3021 | ] |
| 3022 | |
| 3023 | AARCH64_ASM_UKERNELS = [ |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 3024 | "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3025 | "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 3026 | "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3027 | "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 3028 | "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3029 | "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 3030 | "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3031 | "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 3032 | "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 3033 | "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 3034 | "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 3035 | "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 3036 | "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 3037 | "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3038 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S", |
| 3039 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3040 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 3041 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 3042 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3043 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S", |
| 3044 | "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3045 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 3046 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 3047 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 3048 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3049 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3050 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S", |
| 3051 | "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3052 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 3053 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 3054 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 3055 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3056 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3057 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3058 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3059 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3060 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S", |
| 3061 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 3062 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3063 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 3064 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S", |
| 3065 | "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 3066 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 3067 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 3068 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3069 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 3070 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S", |
| 3071 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S", |
| 3072 | "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 3073 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3074 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 3075 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 3076 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 3077 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3078 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 3079 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 3080 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S", |
| 3081 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3082 | "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3083 | "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3084 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 3085 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3086 | "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 3087 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 3088 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 3089 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 3090 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3091 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3092 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3093 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 3094 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3095 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 3096 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 3097 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | f1fd89e | 2020-10-06 23:27:11 -0700 | [diff] [blame] | 3098 | "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S", |
Frank Barchard | f2742c4 | 2020-10-12 21:16:46 -0700 | [diff] [blame] | 3099 | "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S", |
Frank Barchard | a5237a5 | 2020-10-06 16:18:46 -0700 | [diff] [blame] | 3100 | "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3101 | ] |
| 3102 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3103 | INTERNAL_MICROKERNEL_HDRS = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3104 | "src/xnnpack/argmaxpool.h", |
| 3105 | "src/xnnpack/avgpool.h", |
| 3106 | "src/xnnpack/clamp.h", |
| 3107 | "src/xnnpack/common.h", |
| 3108 | "src/xnnpack/conv.h", |
Yury Kartynnik | e784186 | 2020-11-04 18:22:18 -0800 | [diff] [blame] | 3109 | "src/xnnpack/depthtospace.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3110 | "src/xnnpack/dwconv.h", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3111 | "src/xnnpack/fill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3112 | "src/xnnpack/gavgpool.h", |
| 3113 | "src/xnnpack/gemm.h", |
| 3114 | "src/xnnpack/hswish.h", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 3115 | "src/xnnpack/ibilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3116 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 3117 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3118 | "src/xnnpack/lut.h", |
| 3119 | "src/xnnpack/math.h", |
| 3120 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3121 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3122 | "src/xnnpack/packx.h", |
| 3123 | "src/xnnpack/pad.h", |
| 3124 | "src/xnnpack/params.h", |
| 3125 | "src/xnnpack/pavgpool.h", |
| 3126 | "src/xnnpack/ppmm.h", |
| 3127 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 3128 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 3129 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 3130 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3131 | "src/xnnpack/rmax.h", |
| 3132 | "src/xnnpack/scalar-utils.h", |
| 3133 | "src/xnnpack/spmm.h", |
| 3134 | "src/xnnpack/unpool.h", |
| 3135 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 3136 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3137 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 3138 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 3139 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 3140 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 3141 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3142 | "src/xnnpack/zip.h", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3143 | ] |
| 3144 | |
| 3145 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3146 | "include/xnnpack.h", |
| 3147 | "src/xnnpack/allocator.h", |
| 3148 | "src/xnnpack/compute.h", |
| 3149 | "src/xnnpack/im2col.h", |
| 3150 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3151 | "src/xnnpack/math-stubs.h", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3152 | "src/xnnpack/memory-planner.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3153 | "src/xnnpack/operator.h", |
| 3154 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 3155 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3156 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3157 | "src/xnnpack/requantization.h", |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 3158 | "src/xnnpack/subgraph.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3159 | ] |
| 3160 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3161 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3162 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3163 | ] |
| 3164 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3165 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3166 | "include/xnnpack.h", |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 3167 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3168 | ] |
| 3169 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3170 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Frank Barchard | 35db7d0 | 2020-10-26 13:37:34 -0700 | [diff] [blame] | 3171 | "include/xnnpack.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3172 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 3173 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3174 | "src/xnnpack/requantization.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3175 | ] |
| 3176 | |
| 3177 | OPERATOR_TEST_PARAMS_HDRS = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3178 | "src/xnnpack/common.h", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3179 | "src/xnnpack/params.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3180 | ] |
| 3181 | |
| 3182 | WEIGHTS_PACK_HDRS = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3183 | "src/xnnpack/compute.h", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3184 | "src/xnnpack/operator.h", |
| 3185 | "src/xnnpack/pack.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3186 | ] |
| 3187 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 3188 | LOGGING_COPTS = select({ |
| 3189 | # No logging in optimized mode |
| 3190 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 3191 | # Full logging in debug mode |
| 3192 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 3193 | # Error-only logging in default (fastbuild) mode |
| 3194 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 3195 | }) |
| 3196 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3197 | LOGGING_SRCS = select({ |
| 3198 | # No logging in optimized mode |
| 3199 | ":optimized_build": [], |
| 3200 | "//conditions:default": [ |
| 3201 | "src/operator-strings.c", |
| 3202 | "src/subgraph-strings.c", |
| 3203 | ], |
| 3204 | }) |
| 3205 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 3206 | LOGGING_HDRS = [ |
| 3207 | "src/xnnpack/log.h", |
| 3208 | ] |
| 3209 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3210 | xnnpack_cc_library( |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3211 | name = "tables", |
| 3212 | srcs = TABLE_SRCS, |
| 3213 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3214 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3215 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3216 | ) |
| 3217 | |
| 3218 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3219 | name = "scalar_ukernels", |
| 3220 | srcs = SCALAR_UKERNELS, |
| 3221 | hdrs = INTERNAL_HDRS, |
| 3222 | aarch32_copts = ["-marm"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3223 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3224 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3225 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3226 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3227 | "@FP16", |
| 3228 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3229 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3230 | ], |
| 3231 | ) |
| 3232 | |
| 3233 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3234 | name = "scalar_ukernels_test_mode", |
| 3235 | srcs = SCALAR_UKERNELS, |
| 3236 | hdrs = INTERNAL_HDRS, |
| 3237 | aarch32_copts = ["-marm"], |
| 3238 | copts = [ |
| 3239 | "-UNDEBUG", |
| 3240 | "-DXNN_TEST_MODE=1", |
| 3241 | ], |
| 3242 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3243 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3244 | deps = [ |
| 3245 | ":tables", |
| 3246 | "@FP16", |
| 3247 | "@FXdiv", |
| 3248 | "@pthreadpool", |
| 3249 | ], |
| 3250 | ) |
| 3251 | |
| 3252 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3253 | name = "wasm_ukernels", |
| 3254 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3255 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3256 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3257 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 1c6e389 | 2020-06-25 23:56:10 -0700 | [diff] [blame] | 3258 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3259 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3260 | ":tables", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3261 | "@FP16", |
| 3262 | "@FXdiv", |
| 3263 | "@pthreadpool", |
| 3264 | ], |
| 3265 | ) |
| 3266 | |
| 3267 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3268 | name = "wasm_ukernels_test_mode", |
| 3269 | hdrs = INTERNAL_HDRS, |
| 3270 | copts = [ |
| 3271 | "-UNDEBUG", |
| 3272 | "-DXNN_TEST_MODE=1", |
| 3273 | ], |
| 3274 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3275 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3276 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 3277 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3278 | deps = [ |
| 3279 | ":tables", |
| 3280 | "@FP16", |
| 3281 | "@FXdiv", |
| 3282 | "@pthreadpool", |
| 3283 | ], |
| 3284 | ) |
| 3285 | |
| 3286 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3287 | name = "psimd_accmath_ukernels", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3288 | hdrs = INTERNAL_HDRS, |
| 3289 | aarch32_copts = [ |
| 3290 | "-marm", |
| 3291 | "-mfpu=neon", |
| 3292 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3293 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3294 | gcc_x86_copts = ["-msse2"], |
| 3295 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3296 | optimized_copts = [ |
| 3297 | "-O3", |
| 3298 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3299 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3300 | deps = [ |
| 3301 | ":tables", |
| 3302 | "@FP16", |
| 3303 | "@psimd", |
| 3304 | "@pthreadpool", |
| 3305 | ], |
| 3306 | ) |
| 3307 | |
| 3308 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3309 | name = "psimd_accmath_ukernels_test_mode", |
| 3310 | hdrs = INTERNAL_HDRS, |
| 3311 | aarch32_copts = [ |
| 3312 | "-marm", |
| 3313 | "-mfpu=neon", |
| 3314 | ], |
| 3315 | copts = [ |
| 3316 | "-UNDEBUG", |
| 3317 | "-DXNN_TEST_MODE=1", |
| 3318 | ], |
| 3319 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3320 | gcc_x86_copts = ["-msse2"], |
| 3321 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3322 | optimized_copts = [ |
| 3323 | "-O3", |
| 3324 | ], |
| 3325 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
| 3326 | deps = [ |
| 3327 | ":tables", |
| 3328 | "@FP16", |
| 3329 | "@psimd", |
| 3330 | "@pthreadpool", |
| 3331 | ], |
| 3332 | ) |
| 3333 | |
| 3334 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3335 | name = "neon_ukernels", |
| 3336 | hdrs = INTERNAL_HDRS, |
| 3337 | aarch32_copts = [ |
| 3338 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3339 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3340 | "-mfpu=neon", |
| 3341 | ], |
| 3342 | aarch32_srcs = NEON_UKERNELS, |
| 3343 | aarch64_srcs = NEON_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3344 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3345 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3346 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3347 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3348 | "@FP16", |
| 3349 | "@pthreadpool", |
| 3350 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3351 | ) |
| 3352 | |
| 3353 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3354 | name = "neon_ukernels_test_mode", |
| 3355 | hdrs = INTERNAL_HDRS, |
| 3356 | aarch32_copts = [ |
| 3357 | "-marm", |
| 3358 | "-march=armv7-a", |
| 3359 | "-mfpu=neon", |
| 3360 | ], |
| 3361 | aarch32_srcs = NEON_UKERNELS, |
| 3362 | aarch64_srcs = NEON_UKERNELS, |
| 3363 | copts = [ |
| 3364 | "-UNDEBUG", |
| 3365 | "-DXNN_TEST_MODE=1", |
| 3366 | ], |
| 3367 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3368 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3369 | deps = [ |
| 3370 | ":tables", |
| 3371 | "@FP16", |
| 3372 | "@pthreadpool", |
| 3373 | ], |
| 3374 | ) |
| 3375 | |
| 3376 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3377 | name = "neonfma_ukernels", |
| 3378 | hdrs = INTERNAL_HDRS, |
| 3379 | aarch32_copts = [ |
| 3380 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3381 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3382 | "-mfpu=neon-vfpv4", |
| 3383 | ], |
| 3384 | aarch32_srcs = NEONFMA_UKERNELS, |
| 3385 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3386 | apple_aarch32_copts = [ |
| 3387 | "-mcpu=swift", |
| 3388 | "-mtune=generic", |
| 3389 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3390 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3391 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3392 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3393 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3394 | "@FP16", |
| 3395 | "@pthreadpool", |
| 3396 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3397 | ) |
| 3398 | |
| 3399 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3400 | name = "neonfma_ukernels_test_mode", |
| 3401 | hdrs = INTERNAL_HDRS, |
| 3402 | aarch32_copts = [ |
| 3403 | "-marm", |
| 3404 | "-march=armv7-a", |
| 3405 | "-mfpu=neon-vfpv4", |
| 3406 | ], |
| 3407 | aarch32_srcs = NEONFMA_UKERNELS, |
| 3408 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3409 | apple_aarch32_copts = [ |
| 3410 | "-mcpu=swift", |
| 3411 | "-mtune=generic", |
| 3412 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3413 | copts = [ |
| 3414 | "-UNDEBUG", |
| 3415 | "-DXNN_TEST_MODE=1", |
| 3416 | ], |
| 3417 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3418 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3419 | deps = [ |
| 3420 | ":tables", |
| 3421 | "@FP16", |
| 3422 | "@pthreadpool", |
| 3423 | ], |
| 3424 | ) |
| 3425 | |
| 3426 | xnnpack_cc_library( |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3427 | name = "neonv8_ukernels", |
| 3428 | hdrs = INTERNAL_HDRS, |
| 3429 | aarch32_copts = [ |
| 3430 | "-marm", |
| 3431 | "-march=armv8-a", |
| 3432 | "-mfpu=neon-fp-armv8", |
| 3433 | ], |
| 3434 | aarch32_srcs = NEONV8_UKERNELS, |
| 3435 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3436 | apple_aarch32_copts = [ |
| 3437 | "-mcpu=cyclone", |
| 3438 | "-mtune=generic", |
| 3439 | ], |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3440 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3441 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3442 | deps = [ |
| 3443 | ":tables", |
| 3444 | "@FP16", |
| 3445 | "@pthreadpool", |
| 3446 | ], |
| 3447 | ) |
| 3448 | |
| 3449 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3450 | name = "neonv8_ukernels_test_mode", |
| 3451 | hdrs = INTERNAL_HDRS, |
| 3452 | aarch32_copts = [ |
| 3453 | "-marm", |
| 3454 | "-march=armv8-a", |
| 3455 | "-mfpu=neon-fp-armv8", |
| 3456 | ], |
| 3457 | aarch32_srcs = NEONV8_UKERNELS, |
| 3458 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3459 | apple_aarch32_copts = [ |
| 3460 | "-mcpu=cyclone", |
| 3461 | "-mtune=generic", |
| 3462 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3463 | copts = [ |
| 3464 | "-UNDEBUG", |
| 3465 | "-DXNN_TEST_MODE=1", |
| 3466 | ], |
| 3467 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3468 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3469 | deps = [ |
| 3470 | ":tables", |
| 3471 | "@FP16", |
| 3472 | "@pthreadpool", |
| 3473 | ], |
| 3474 | ) |
| 3475 | |
| 3476 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3477 | name = "neonfp16arith_ukernels", |
| 3478 | hdrs = INTERNAL_HDRS, |
| 3479 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 3480 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3481 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3482 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3483 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3484 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3485 | "@FP16", |
| 3486 | "@pthreadpool", |
| 3487 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3488 | ) |
| 3489 | |
| 3490 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3491 | name = "neonfp16arith_ukernels_test_mode", |
| 3492 | hdrs = INTERNAL_HDRS, |
| 3493 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 3494 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 3495 | copts = [ |
| 3496 | "-UNDEBUG", |
| 3497 | "-DXNN_TEST_MODE=1", |
| 3498 | ], |
| 3499 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3500 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3501 | deps = [ |
| 3502 | ":tables", |
| 3503 | "@FP16", |
| 3504 | "@pthreadpool", |
| 3505 | ], |
| 3506 | ) |
| 3507 | |
| 3508 | xnnpack_cc_library( |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3509 | name = "neondot_ukernels", |
| 3510 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 799ac75 | 2020-08-13 16:08:03 -0700 | [diff] [blame] | 3511 | aarch32_copts = [ |
| 3512 | "-marm", |
| 3513 | "-march=armv8.2-a+dotprod", |
| 3514 | "-mfpu=neon-fp-armv8", |
| 3515 | ], |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3516 | aarch32_srcs = NEONDOT_UKERNELS, |
| 3517 | aarch64_copts = ["-march=armv8.2-a+dotprod"], |
| 3518 | aarch64_srcs = NEONDOT_UKERNELS, |
| 3519 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3520 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3521 | deps = [ |
| 3522 | ":tables", |
| 3523 | "@FP16", |
| 3524 | "@pthreadpool", |
| 3525 | ], |
| 3526 | ) |
| 3527 | |
| 3528 | xnnpack_cc_library( |
| 3529 | name = "neondot_ukernels_test_mode", |
| 3530 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 799ac75 | 2020-08-13 16:08:03 -0700 | [diff] [blame] | 3531 | aarch32_copts = [ |
| 3532 | "-marm", |
| 3533 | "-march=armv8.2-a+dotprod", |
| 3534 | "-mfpu=neon-fp-armv8", |
| 3535 | ], |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3536 | aarch32_srcs = NEONDOT_UKERNELS, |
| 3537 | aarch64_copts = ["-march=armv8.2-a+dotprod"], |
| 3538 | aarch64_srcs = NEONDOT_UKERNELS, |
| 3539 | copts = [ |
| 3540 | "-UNDEBUG", |
| 3541 | "-DXNN_TEST_MODE=1", |
| 3542 | ], |
| 3543 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3544 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3545 | deps = [ |
| 3546 | ":tables", |
| 3547 | "@FP16", |
| 3548 | "@pthreadpool", |
| 3549 | ], |
| 3550 | ) |
| 3551 | |
| 3552 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3553 | name = "sse2_ukernels", |
| 3554 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3555 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3556 | gcc_x86_copts = ["-msse2"], |
| 3557 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3558 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3559 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3560 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3561 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3562 | "@FP16", |
| 3563 | "@pthreadpool", |
| 3564 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3565 | ) |
| 3566 | |
| 3567 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3568 | name = "sse2_ukernels_test_mode", |
| 3569 | hdrs = INTERNAL_HDRS, |
| 3570 | copts = [ |
| 3571 | "-UNDEBUG", |
| 3572 | "-DXNN_TEST_MODE=1", |
| 3573 | ], |
| 3574 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3575 | gcc_x86_copts = ["-msse2"], |
| 3576 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3577 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3578 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| 3579 | deps = [ |
| 3580 | ":tables", |
| 3581 | "@FP16", |
| 3582 | "@pthreadpool", |
| 3583 | ], |
| 3584 | ) |
| 3585 | |
| 3586 | xnnpack_cc_library( |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3587 | name = "ssse3_ukernels", |
| 3588 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3589 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3590 | gcc_x86_copts = ["-mssse3"], |
| 3591 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3592 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3593 | x86_srcs = SSSE3_UKERNELS, |
| 3594 | deps = [ |
| 3595 | ":tables", |
| 3596 | "@FP16", |
| 3597 | "@pthreadpool", |
| 3598 | ], |
| 3599 | ) |
| 3600 | |
| 3601 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3602 | name = "ssse3_ukernels_test_mode", |
| 3603 | hdrs = INTERNAL_HDRS, |
| 3604 | copts = [ |
| 3605 | "-UNDEBUG", |
| 3606 | "-DXNN_TEST_MODE=1", |
| 3607 | ], |
| 3608 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3609 | gcc_x86_copts = ["-mssse3"], |
| 3610 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3611 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3612 | x86_srcs = SSSE3_UKERNELS, |
| 3613 | deps = [ |
| 3614 | ":tables", |
| 3615 | "@FP16", |
| 3616 | "@pthreadpool", |
| 3617 | ], |
| 3618 | ) |
| 3619 | |
| 3620 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3621 | name = "sse41_ukernels", |
| 3622 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3623 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3624 | gcc_x86_copts = ["-msse4.1"], |
| 3625 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3626 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3627 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3628 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3629 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3630 | "@FP16", |
| 3631 | "@pthreadpool", |
| 3632 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3633 | ) |
| 3634 | |
| 3635 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3636 | name = "sse41_ukernels_test_mode", |
| 3637 | hdrs = INTERNAL_HDRS, |
| 3638 | copts = [ |
| 3639 | "-UNDEBUG", |
| 3640 | "-DXNN_TEST_MODE=1", |
| 3641 | ], |
| 3642 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3643 | gcc_x86_copts = ["-msse4.1"], |
| 3644 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3645 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3646 | x86_srcs = SSE41_UKERNELS, |
| 3647 | deps = [ |
| 3648 | ":tables", |
| 3649 | "@FP16", |
| 3650 | "@pthreadpool", |
| 3651 | ], |
| 3652 | ) |
| 3653 | |
| 3654 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3655 | name = "avx_ukernels", |
| 3656 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3657 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3658 | gcc_x86_copts = ["-mavx"], |
| 3659 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3660 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3661 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3662 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3663 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3664 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3665 | "@FP16", |
| 3666 | "@pthreadpool", |
| 3667 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3668 | ) |
| 3669 | |
| 3670 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3671 | name = "avx_ukernels_test_mode", |
| 3672 | hdrs = INTERNAL_HDRS, |
| 3673 | copts = [ |
| 3674 | "-UNDEBUG", |
| 3675 | "-DXNN_TEST_MODE=1", |
| 3676 | ], |
| 3677 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3678 | gcc_x86_copts = ["-mavx"], |
| 3679 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3680 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3681 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3682 | x86_srcs = AVX_UKERNELS, |
| 3683 | deps = [ |
| 3684 | ":tables", |
| 3685 | "@FP16", |
| 3686 | "@pthreadpool", |
| 3687 | ], |
| 3688 | ) |
| 3689 | |
| 3690 | xnnpack_cc_library( |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3691 | name = "xop_ukernels", |
| 3692 | hdrs = INTERNAL_HDRS, |
| 3693 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3694 | gcc_x86_copts = ["-mxop"], |
| 3695 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3696 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3697 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3698 | x86_srcs = XOP_UKERNELS, |
| 3699 | deps = [ |
| 3700 | ":tables", |
| 3701 | "@FP16", |
| 3702 | "@pthreadpool", |
| 3703 | ], |
| 3704 | ) |
| 3705 | |
| 3706 | xnnpack_cc_library( |
| 3707 | name = "xop_ukernels_test_mode", |
| 3708 | hdrs = INTERNAL_HDRS, |
| 3709 | copts = [ |
| 3710 | "-UNDEBUG", |
| 3711 | "-DXNN_TEST_MODE=1", |
| 3712 | ], |
| 3713 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3714 | gcc_x86_copts = ["-mxop"], |
| 3715 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3716 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3717 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3718 | x86_srcs = XOP_UKERNELS, |
| 3719 | deps = [ |
| 3720 | ":tables", |
| 3721 | "@FP16", |
| 3722 | "@pthreadpool", |
| 3723 | ], |
| 3724 | ) |
| 3725 | |
| 3726 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3727 | name = "fma3_ukernels", |
| 3728 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3729 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3730 | gcc_x86_copts = ["-mfma"], |
| 3731 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3732 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3733 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3734 | x86_srcs = FMA3_UKERNELS, |
| 3735 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3736 | ":tables", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3737 | "@FP16", |
| 3738 | "@pthreadpool", |
| 3739 | ], |
| 3740 | ) |
| 3741 | |
| 3742 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3743 | name = "fma3_ukernels_test_mode", |
| 3744 | hdrs = INTERNAL_HDRS, |
| 3745 | copts = [ |
| 3746 | "-UNDEBUG", |
| 3747 | "-DXNN_TEST_MODE=1", |
| 3748 | ], |
| 3749 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3750 | gcc_x86_copts = ["-mfma"], |
| 3751 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3752 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3753 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3754 | x86_srcs = FMA3_UKERNELS, |
| 3755 | deps = [ |
| 3756 | ":tables", |
| 3757 | "@FP16", |
| 3758 | "@pthreadpool", |
| 3759 | ], |
| 3760 | ) |
| 3761 | |
| 3762 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3763 | name = "avx2_ukernels", |
| 3764 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3765 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3766 | gcc_x86_copts = [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3767 | "-mfma", |
| 3768 | "-mavx2", |
| 3769 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3770 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3771 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 3772 | msvc_x86_64_copts = ["/arch:AVX2"], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3773 | x86_srcs = AVX2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3774 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3775 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3776 | "@FP16", |
| 3777 | "@pthreadpool", |
| 3778 | ], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3779 | ) |
| 3780 | |
| 3781 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3782 | name = "avx2_ukernels_test_mode", |
| 3783 | hdrs = INTERNAL_HDRS, |
| 3784 | copts = [ |
| 3785 | "-UNDEBUG", |
| 3786 | "-DXNN_TEST_MODE=1", |
| 3787 | ], |
| 3788 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3789 | gcc_x86_copts = [ |
| 3790 | "-mfma", |
| 3791 | "-mavx2", |
| 3792 | ], |
| 3793 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3794 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 3795 | msvc_x86_64_copts = ["/arch:AVX2"], |
| 3796 | x86_srcs = AVX2_UKERNELS, |
| 3797 | deps = [ |
| 3798 | ":tables", |
| 3799 | "@FP16", |
| 3800 | "@pthreadpool", |
| 3801 | ], |
| 3802 | ) |
| 3803 | |
| 3804 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3805 | name = "avx512f_ukernels", |
| 3806 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3807 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3808 | gcc_x86_copts = ["-mavx512f"], |
| 3809 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3810 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3811 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3812 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3813 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3814 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3815 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3816 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3817 | "@FP16", |
| 3818 | "@pthreadpool", |
| 3819 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3820 | ) |
| 3821 | |
| 3822 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3823 | name = "avx512f_ukernels_test_mode", |
| 3824 | hdrs = INTERNAL_HDRS, |
| 3825 | copts = [ |
| 3826 | "-UNDEBUG", |
| 3827 | "-DXNN_TEST_MODE=1", |
| 3828 | ], |
| 3829 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3830 | gcc_x86_copts = ["-mavx512f"], |
| 3831 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3832 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3833 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3834 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3835 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3836 | x86_srcs = AVX512F_UKERNELS, |
| 3837 | deps = [ |
| 3838 | ":tables", |
| 3839 | "@FP16", |
| 3840 | "@pthreadpool", |
| 3841 | ], |
| 3842 | ) |
| 3843 | |
| 3844 | xnnpack_cc_library( |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3845 | name = "avx512skx_ukernels", |
| 3846 | hdrs = INTERNAL_HDRS, |
| 3847 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3848 | gcc_x86_copts = [ |
| 3849 | "-mavx512f", |
| 3850 | "-mavx512cd", |
| 3851 | "-mavx512bw", |
| 3852 | "-mavx512dq", |
| 3853 | "-mavx512vl", |
| 3854 | ], |
| 3855 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3856 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3857 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3858 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3859 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3860 | x86_srcs = AVX512SKX_UKERNELS, |
| 3861 | deps = [ |
| 3862 | ":tables", |
| 3863 | "@FP16", |
| 3864 | "@pthreadpool", |
| 3865 | ], |
| 3866 | ) |
| 3867 | |
| 3868 | xnnpack_cc_library( |
| 3869 | name = "avx512skx_ukernels_test_mode", |
| 3870 | hdrs = INTERNAL_HDRS, |
| 3871 | copts = [ |
| 3872 | "-UNDEBUG", |
| 3873 | "-DXNN_TEST_MODE=1", |
| 3874 | ], |
| 3875 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3876 | gcc_x86_copts = [ |
| 3877 | "-mavx512f", |
| 3878 | "-mavx512cd", |
| 3879 | "-mavx512bw", |
| 3880 | "-mavx512dq", |
| 3881 | "-mavx512vl", |
| 3882 | ], |
| 3883 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3884 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3885 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3886 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3887 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3888 | x86_srcs = AVX512SKX_UKERNELS, |
| 3889 | deps = [ |
| 3890 | ":tables", |
| 3891 | "@FP16", |
| 3892 | "@pthreadpool", |
| 3893 | ], |
| 3894 | ) |
| 3895 | |
| 3896 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3897 | name = "asm_ukernels", |
| 3898 | hdrs = ["src/xnnpack/assembly.h"], |
| 3899 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
Frank Barchard | 31bb45b | 2020-10-06 00:26:33 -0700 | [diff] [blame] | 3900 | aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3901 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3902 | wasm_srcs = WASM32_ASM_UKERNELS, |
| 3903 | wasmsimd_srcs = WASM32_ASM_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3904 | ) |
| 3905 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3906 | xnnpack_cc_library( |
| 3907 | name = "logging_utils", |
| 3908 | srcs = LOGGING_SRCS, |
| 3909 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3910 | copts = LOGGING_COPTS + [ |
| 3911 | "-Isrc", |
| 3912 | "-Iinclude", |
| 3913 | ] + select({ |
| 3914 | ":debug_build": [], |
| 3915 | "//conditions:default": xnnpack_min_size_copts(), |
| 3916 | }), |
| 3917 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3918 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3919 | visibility = xnnpack_visibility(), |
| 3920 | deps = [ |
| 3921 | "@FP16", |
| 3922 | "@clog", |
| 3923 | "@pthreadpool", |
| 3924 | ], |
| 3925 | ) |
| 3926 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3927 | xnnpack_aggregate_library( |
| 3928 | name = "ukernels", |
| 3929 | aarch32_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3930 | ":neon_ukernels", |
| 3931 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3932 | ":neonv8_ukernels", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3933 | ":neondot_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3934 | ":asm_ukernels", |
| 3935 | ], |
| 3936 | aarch64_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3937 | ":neon_ukernels", |
| 3938 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3939 | ":neonv8_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3940 | ":neonfp16arith_ukernels", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3941 | ":neondot_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3942 | ":asm_ukernels", |
| 3943 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3944 | generic_deps = [ |
| 3945 | ":scalar_ukernels", |
| 3946 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3947 | psimd_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3948 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3949 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3950 | wasm_deps = [ |
| 3951 | ":wasm_ukernels", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3952 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3953 | ], |
| 3954 | wasmsimd_deps = [ |
| 3955 | ":wasm_ukernels", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3956 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3957 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3958 | x86_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3959 | ":sse2_ukernels", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3960 | ":ssse3_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3961 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3962 | ":avx_ukernels", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3963 | ":xop_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3964 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3965 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3966 | ":avx512f_ukernels", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3967 | ":avx512skx_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3968 | ], |
| 3969 | ) |
| 3970 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3971 | xnnpack_aggregate_library( |
| 3972 | name = "ukernels_test_mode", |
| 3973 | aarch32_deps = [ |
| 3974 | ":neon_ukernels_test_mode", |
| 3975 | ":neonfma_ukernels_test_mode", |
| 3976 | ":neonv8_ukernels_test_mode", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3977 | ":neondot_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3978 | ":asm_ukernels", |
| 3979 | ], |
| 3980 | aarch64_deps = [ |
| 3981 | ":neon_ukernels_test_mode", |
| 3982 | ":neonfma_ukernels_test_mode", |
| 3983 | ":neonv8_ukernels_test_mode", |
| 3984 | ":neonfp16arith_ukernels_test_mode", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3985 | ":neondot_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3986 | ":asm_ukernels", |
| 3987 | ], |
| 3988 | generic_deps = [ |
| 3989 | ":scalar_ukernels_test_mode", |
| 3990 | ], |
| 3991 | psimd_deps = [ |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3992 | ":psimd_accmath_ukernels_test_mode", |
| 3993 | ], |
| 3994 | wasm_deps = [ |
| 3995 | ":wasm_ukernels_test_mode", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3996 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3997 | ], |
| 3998 | wasmsimd_deps = [ |
| 3999 | ":wasm_ukernels_test_mode", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 4000 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4001 | ], |
| 4002 | x86_deps = [ |
| 4003 | ":sse2_ukernels_test_mode", |
| 4004 | ":ssse3_ukernels_test_mode", |
| 4005 | ":sse41_ukernels_test_mode", |
| 4006 | ":avx_ukernels_test_mode", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 4007 | ":xop_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4008 | ":fma3_ukernels_test_mode", |
| 4009 | ":avx2_ukernels_test_mode", |
| 4010 | ":avx512f_ukernels_test_mode", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 4011 | ":avx512skx_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4012 | ], |
| 4013 | ) |
| 4014 | |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 4015 | # This target must be compatible with Android NDK r17 and compile with Clang 6. |
| 4016 | # It is similar to :ukernels target, but excludes microkernels which require |
| 4017 | # NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics. |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4018 | xnnpack_aggregate_library( |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 4019 | name = "ukernels_max_compatibility", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4020 | aarch32_deps = [ |
| 4021 | ":neon_ukernels", |
| 4022 | ":neonfma_ukernels", |
| 4023 | ":neonv8_ukernels", |
| 4024 | ":asm_ukernels", |
| 4025 | ], |
| 4026 | aarch64_deps = [ |
| 4027 | ":neon_ukernels", |
| 4028 | ":neonfma_ukernels", |
| 4029 | ":neonv8_ukernels", |
| 4030 | ":asm_ukernels", |
| 4031 | ], |
| 4032 | generic_deps = [ |
| 4033 | ":scalar_ukernels", |
| 4034 | ], |
| 4035 | psimd_deps = [ |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4036 | ":psimd_accmath_ukernels", |
| 4037 | ], |
| 4038 | wasm_deps = [ |
| 4039 | ":wasm_ukernels", |
| 4040 | ], |
| 4041 | wasmsimd_deps = [ |
| 4042 | ":wasm_ukernels", |
| 4043 | ], |
| 4044 | x86_deps = [ |
| 4045 | ":sse2_ukernels", |
| 4046 | ":ssse3_ukernels", |
| 4047 | ":sse41_ukernels", |
| 4048 | ":avx_ukernels", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 4049 | ":xop_ukernels", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4050 | ":fma3_ukernels", |
| 4051 | ":avx2_ukernels", |
| 4052 | ":avx512f_ukernels", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 4053 | ":avx512skx_ukernels", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4054 | ], |
| 4055 | ) |
| 4056 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4057 | xnnpack_cc_library( |
| 4058 | name = "im2col", |
| 4059 | srcs = ["src/im2col.c"], |
| 4060 | hdrs = [ |
| 4061 | "src/xnnpack/common.h", |
| 4062 | "src/xnnpack/im2col.h", |
| 4063 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4064 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4065 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4066 | ) |
| 4067 | |
| 4068 | xnnpack_cc_library( |
| 4069 | name = "indirection", |
| 4070 | srcs = ["src/indirection.c"], |
| 4071 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4072 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4073 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4074 | deps = [ |
| 4075 | "@FP16", |
| 4076 | "@FXdiv", |
| 4077 | "@pthreadpool", |
| 4078 | ], |
| 4079 | ) |
| 4080 | |
| 4081 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4082 | name = "indirection_test_mode", |
| 4083 | srcs = ["src/indirection.c"], |
| 4084 | hdrs = INTERNAL_HDRS, |
| 4085 | copts = [ |
| 4086 | "-UNDEBUG", |
| 4087 | "-DXNN_TEST_MODE=1", |
| 4088 | ], |
| 4089 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4090 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4091 | deps = [ |
| 4092 | "@FP16", |
| 4093 | "@FXdiv", |
| 4094 | "@pthreadpool", |
| 4095 | ], |
| 4096 | ) |
| 4097 | |
| 4098 | xnnpack_cc_library( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4099 | name = "packing", |
| 4100 | srcs = ["src/packing.c"], |
| 4101 | hdrs = INTERNAL_HDRS, |
| 4102 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4103 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4104 | deps = [ |
| 4105 | "@FP16", |
| 4106 | "@FXdiv", |
| 4107 | "@pthreadpool", |
| 4108 | ], |
| 4109 | ) |
| 4110 | |
| 4111 | xnnpack_cc_library( |
| 4112 | name = "packing_test_mode", |
| 4113 | srcs = ["src/packing.c"], |
| 4114 | hdrs = INTERNAL_HDRS, |
| 4115 | copts = [ |
| 4116 | "-UNDEBUG", |
| 4117 | "-DXNN_TEST_MODE=1", |
| 4118 | ], |
| 4119 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4120 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4121 | deps = [ |
| 4122 | "@FP16", |
| 4123 | "@FXdiv", |
| 4124 | "@pthreadpool", |
| 4125 | ], |
| 4126 | ) |
| 4127 | |
| 4128 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4129 | name = "operator_run", |
| 4130 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 4131 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4132 | copts = LOGGING_COPTS + select({ |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4133 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4134 | "//conditions:default": [], |
| 4135 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4136 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4137 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4138 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4139 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4140 | "@FP16", |
| 4141 | "@FXdiv", |
| 4142 | "@clog", |
| 4143 | "@pthreadpool", |
| 4144 | ], |
| 4145 | ) |
| 4146 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4147 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4148 | name = "operator_run_test_mode", |
| 4149 | srcs = ["src/operator-run.c"], |
| 4150 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 4151 | copts = LOGGING_COPTS + [ |
| 4152 | "-UNDEBUG", |
| 4153 | "-DXNN_TEST_MODE=1", |
| 4154 | ] + select({ |
| 4155 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4156 | "//conditions:default": [], |
| 4157 | }), |
| 4158 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4159 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4160 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4161 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4162 | "@FP16", |
| 4163 | "@FXdiv", |
| 4164 | "@clog", |
| 4165 | "@pthreadpool", |
| 4166 | ], |
| 4167 | ) |
| 4168 | |
| 4169 | xnnpack_cc_library( |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4170 | name = "memory_planner", |
| 4171 | srcs = ["src/memory-planner.c"], |
| 4172 | hdrs = INTERNAL_HDRS, |
| 4173 | defines = select({ |
| 4174 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 4175 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 4176 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 4177 | }), |
| 4178 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4179 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4180 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4181 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4182 | "@pthreadpool", |
| 4183 | ], |
| 4184 | ) |
| 4185 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4186 | xnnpack_cc_library( |
| 4187 | name = "memory_planner_test_mode", |
| 4188 | srcs = ["src/memory-planner.c"], |
| 4189 | hdrs = INTERNAL_HDRS, |
| 4190 | copts = [ |
| 4191 | "-UNDEBUG", |
| 4192 | "-DXNN_TEST_MODE=1", |
| 4193 | ], |
| 4194 | defines = select({ |
| 4195 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 4196 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 4197 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 4198 | }), |
| 4199 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4200 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4201 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4202 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4203 | "@pthreadpool", |
| 4204 | ], |
| 4205 | ) |
| 4206 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4207 | cc_library( |
| 4208 | name = "enable_assembly", |
| 4209 | defines = select({ |
| 4210 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 4211 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 4212 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4213 | }), |
| 4214 | ) |
| 4215 | |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4216 | cc_library( |
| 4217 | name = "enable_sparse", |
| 4218 | defines = select({ |
| 4219 | ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"], |
| 4220 | ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"], |
Marat Dukhan | 8b283aa | 2020-06-18 16:48:57 -0700 | [diff] [blame] | 4221 | "//conditions:default": ["XNN_ENABLE_SPARSE=0"], |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4222 | }), |
| 4223 | ) |
| 4224 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4225 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4226 | name = "operators", |
| 4227 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 4228 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4229 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4230 | ], |
| 4231 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4232 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4233 | "-Isrc", |
| 4234 | "-Iinclude", |
| 4235 | ] + select({ |
| 4236 | ":debug_build": [], |
| 4237 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4238 | }) + select({ |
| 4239 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4240 | "//conditions:default": [], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4241 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4242 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4243 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4244 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4245 | ":indirection", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4246 | ":logging_utils", |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4247 | ":packing", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4248 | "@FP16", |
| 4249 | "@FXdiv", |
| 4250 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4251 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4252 | ], |
| 4253 | ) |
| 4254 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4255 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4256 | name = "operators_test_mode", |
| 4257 | srcs = OPERATOR_SRCS + [ |
| 4258 | "src/memory.c", |
| 4259 | "src/operator-delete.c", |
| 4260 | ], |
| 4261 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 4262 | copts = LOGGING_COPTS + [ |
| 4263 | "-Isrc", |
| 4264 | "-Iinclude", |
| 4265 | "-UNDEBUG", |
| 4266 | "-DXNN_TEST_MODE=1", |
| 4267 | ] + select({ |
| 4268 | ":debug_build": [], |
| 4269 | "//conditions:default": xnnpack_min_size_copts(), |
| 4270 | }) + select({ |
| 4271 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4272 | "//conditions:default": [], |
| 4273 | }), |
| 4274 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4275 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4276 | deps = [ |
| 4277 | ":indirection_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4278 | ":logging_utils", |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4279 | ":packing_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4280 | "@FP16", |
| 4281 | "@FXdiv", |
| 4282 | "@clog", |
| 4283 | "@pthreadpool", |
| 4284 | ], |
| 4285 | ) |
| 4286 | |
| 4287 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4288 | name = "XNNPACK", |
| 4289 | srcs = [ |
| 4290 | "src/init.c", |
Marat Dukhan | ccfdbd1 | 2020-02-03 14:27:45 -0800 | [diff] [blame] | 4291 | "src/runtime.c", |
| 4292 | "src/subgraph.c", |
| 4293 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 4294 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4295 | hdrs = ["include/xnnpack.h"], |
| 4296 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4297 | "-Isrc", |
| 4298 | "-Iinclude", |
| 4299 | ] + select({ |
| 4300 | ":debug_build": [], |
| 4301 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4302 | }) + select({ |
| 4303 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4304 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4305 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4306 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4307 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4308 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4309 | visibility = xnnpack_visibility(), |
| 4310 | deps = [ |
| 4311 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4312 | ":enable_sparse", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4313 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4314 | ":memory_planner", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4315 | ":operator_run", |
| 4316 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4317 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4318 | "@clog", |
Marat Dukhan | ab2946c | 2020-05-21 20:04:13 -0700 | [diff] [blame] | 4319 | "@FP16", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4320 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 4321 | ] + select({ |
| 4322 | ":emscripten": [], |
| 4323 | "//conditions:default": ["@cpuinfo"], |
| 4324 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4325 | ) |
| 4326 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4327 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4328 | name = "XNNPACK_test_mode", |
| 4329 | srcs = [ |
| 4330 | "src/init.c", |
| 4331 | "src/runtime.c", |
| 4332 | "src/subgraph.c", |
| 4333 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 4334 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4335 | hdrs = ["include/xnnpack.h"], |
| 4336 | copts = LOGGING_COPTS + [ |
| 4337 | "-Isrc", |
| 4338 | "-Iinclude", |
| 4339 | "-UNDEBUG", |
| 4340 | "-DXNN_TEST_MODE=1", |
| 4341 | ] + select({ |
| 4342 | ":debug_build": [], |
| 4343 | "//conditions:default": xnnpack_min_size_copts(), |
| 4344 | }) + select({ |
| 4345 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4346 | "//conditions:default": [], |
| 4347 | }), |
| 4348 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4349 | includes = ["include"], |
| 4350 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4351 | visibility = xnnpack_visibility(), |
| 4352 | deps = [ |
| 4353 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4354 | ":enable_sparse", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4355 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4356 | ":memory_planner_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4357 | ":operator_run_test_mode", |
| 4358 | ":operators_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4359 | ":ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4360 | "@clog", |
| 4361 | "@FP16", |
| 4362 | "@pthreadpool", |
| 4363 | ] + select({ |
| 4364 | ":emscripten": [], |
| 4365 | "//conditions:default": ["@cpuinfo"], |
| 4366 | }), |
| 4367 | ) |
| 4368 | |
| 4369 | xnnpack_cc_library( |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4370 | name = "xnnpack_f32", |
| 4371 | srcs = [ |
| 4372 | "src/init.c", |
| 4373 | "src/runtime.c", |
| 4374 | "src/subgraph.c", |
| 4375 | "src/tensor.c", |
| 4376 | ] + SUBGRAPH_SRCS, |
| 4377 | hdrs = ["include/xnnpack.h"], |
| 4378 | copts = LOGGING_COPTS + [ |
| 4379 | "-Isrc", |
| 4380 | "-Iinclude", |
| 4381 | ] + select({ |
| 4382 | ":debug_build": [], |
| 4383 | "//conditions:default": xnnpack_min_size_copts(), |
| 4384 | }) + select({ |
| 4385 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4386 | "//conditions:default": [], |
| 4387 | }), |
| 4388 | defines = [ |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 4389 | "XNN_NO_QS8_OPERATORS", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4390 | "XNN_NO_QU8_OPERATORS", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4391 | "XNN_NO_U8_OPERATORS", |
| 4392 | "XNN_NO_X8_OPERATORS", |
| 4393 | "XNN_NO_F16_OPERATORS", |
| 4394 | "XNN_NO_X16_OPERATORS", |
| 4395 | ], |
| 4396 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4397 | includes = ["include"], |
| 4398 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4399 | visibility = xnnpack_visibility(), |
| 4400 | deps = [ |
| 4401 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4402 | ":enable_sparse", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4403 | ":logging_utils", |
| 4404 | ":memory_planner", |
| 4405 | ":operator_run", |
| 4406 | ":operators", |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 4407 | ":ukernels_max_compatibility", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4408 | "@clog", |
| 4409 | "@FP16", |
| 4410 | "@pthreadpool", |
| 4411 | ] + select({ |
| 4412 | ":emscripten": [], |
| 4413 | "//conditions:default": ["@cpuinfo"], |
| 4414 | }), |
| 4415 | ) |
| 4416 | |
| 4417 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4418 | name = "xnnpack_operators_nhwc_f32", |
| 4419 | srcs = [ |
| 4420 | "src/init.c", |
| 4421 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4422 | hdrs = ["include/xnnpack.h"], |
| 4423 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4424 | "-Isrc", |
| 4425 | "-Iinclude", |
| 4426 | ] + select({ |
| 4427 | ":debug_build": [], |
| 4428 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4429 | }) + select({ |
| 4430 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4431 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4432 | }), |
| 4433 | defines = [ |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 4434 | "XNN_NO_QS8_OPERATORS", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4435 | "XNN_NO_QU8_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4436 | "XNN_NO_U8_OPERATORS", |
| 4437 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4438 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4439 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4440 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4441 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4442 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4443 | visibility = xnnpack_visibility(), |
| 4444 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4445 | ":enable_assembly", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4446 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4447 | ":operator_run", |
| 4448 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4449 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4450 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4451 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4452 | ] + select({ |
| 4453 | ":emscripten": [], |
| 4454 | "//conditions:default": ["@cpuinfo"], |
| 4455 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4456 | ) |
| 4457 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4458 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4459 | name = "bench_utils", |
| 4460 | srcs = ["bench/utils.cc"], |
| 4461 | hdrs = ["bench/utils.h"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 4462 | deps = [ |
| 4463 | "@com_google_benchmark//:benchmark", |
| 4464 | "@cpuinfo", |
| 4465 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4466 | ) |
| 4467 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 4468 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4469 | |
| 4470 | xnnpack_benchmark( |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 4471 | name = "qs8_gemm_bench", |
| 4472 | srcs = [ |
| 4473 | "bench/gemm.h", |
| 4474 | "bench/qs8-gemm.cc", |
| 4475 | "src/xnnpack/AlignedAllocator.h", |
| 4476 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Frank Barchard | 31328cb | 2020-10-12 11:55:18 -0700 | [diff] [blame] | 4477 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 4478 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(), |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 4479 | ) |
| 4480 | |
| 4481 | xnnpack_benchmark( |
Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4482 | name = "qs8_requantization_bench", |
| 4483 | srcs = [ |
| 4484 | "bench/qs8-requantization.cc", |
Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4485 | "src/xnnpack/AlignedAllocator.h", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 4486 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4487 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4488 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4489 | ) |
| 4490 | |
| 4491 | xnnpack_benchmark( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4492 | name = "qu8_gemm_bench", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4493 | srcs = [ |
| 4494 | "bench/gemm.h", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4495 | "bench/qu8-gemm.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4496 | "src/xnnpack/AlignedAllocator.h", |
| 4497 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4498 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4499 | 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] | 4500 | ) |
| 4501 | |
| 4502 | xnnpack_benchmark( |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4503 | name = "qu8_requantization_bench", |
| 4504 | srcs = [ |
| 4505 | "bench/qu8-requantization.cc", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4506 | "src/xnnpack/AlignedAllocator.h", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 4507 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4508 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4509 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4510 | ) |
| 4511 | |
| 4512 | xnnpack_benchmark( |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 4513 | name = "f16_igemm_bench", |
| 4514 | srcs = [ |
| 4515 | "bench/f16-igemm.cc", |
| 4516 | "bench/conv.h", |
| 4517 | "bench/google/conv.h", |
| 4518 | "src/xnnpack/AlignedAllocator.h", |
| 4519 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4520 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4521 | ":indirection", |
| 4522 | ":packing", |
| 4523 | ], |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 4524 | ) |
| 4525 | |
| 4526 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4527 | name = "f16_gemm_bench", |
| 4528 | srcs = [ |
| 4529 | "bench/f16-gemm.cc", |
| 4530 | "bench/gemm.h", |
| 4531 | "src/xnnpack/AlignedAllocator.h", |
| 4532 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4533 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4534 | ":packing", |
| 4535 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4536 | ) |
| 4537 | |
| 4538 | xnnpack_benchmark( |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4539 | name = "f16_spmm_bench", |
| 4540 | srcs = [ |
| 4541 | "bench/f16-spmm.cc", |
| 4542 | "bench/gemm.h", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4543 | "src/xnnpack/AlignedAllocator.h", |
| 4544 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4545 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4546 | ) |
| 4547 | |
| 4548 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4549 | name = "f32_igemm_bench", |
| 4550 | srcs = [ |
| 4551 | "bench/f32-igemm.cc", |
| 4552 | "bench/conv.h", |
| 4553 | "src/xnnpack/AlignedAllocator.h", |
| 4554 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4555 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4556 | ":indirection", |
| 4557 | ":packing", |
| 4558 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4559 | ) |
| 4560 | |
| 4561 | xnnpack_benchmark( |
Frank Barchard | c201006 | 2020-07-19 10:04:12 -0700 | [diff] [blame] | 4562 | name = "f16_relu_bench", |
| 4563 | srcs = [ |
| 4564 | "bench/f16-relu.cc", |
| 4565 | "src/xnnpack/AlignedAllocator.h", |
| 4566 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4567 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4568 | ) |
| 4569 | |
| 4570 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4571 | name = "f32_conv_hwc_bench", |
| 4572 | srcs = [ |
| 4573 | "bench/f32-conv-hwc.cc", |
| 4574 | "bench/dconv.h", |
| 4575 | "src/xnnpack/AlignedAllocator.h", |
| 4576 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4577 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4578 | ":packing", |
| 4579 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4580 | ) |
| 4581 | |
| 4582 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4583 | name = "f32_conv_hwc2chw_bench", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4584 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4585 | "bench/f32-conv-hwc2chw.cc", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4586 | "bench/dconv.h", |
| 4587 | "src/xnnpack/AlignedAllocator.h", |
| 4588 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4589 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4590 | ":packing", |
| 4591 | ], |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4592 | ) |
| 4593 | |
| 4594 | xnnpack_benchmark( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4595 | name = "f16_dwconv_bench", |
| 4596 | srcs = [ |
| 4597 | "bench/f16-dwconv.cc", |
| 4598 | "bench/dwconv.h", |
| 4599 | "bench/google/dwconv.h", |
| 4600 | "src/xnnpack/AlignedAllocator.h", |
| 4601 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4602 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4603 | ":indirection", |
| 4604 | ":packing", |
| 4605 | ], |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4606 | ) |
| 4607 | |
| 4608 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4609 | name = "f32_dwconv_bench", |
| 4610 | srcs = [ |
| 4611 | "bench/f32-dwconv.cc", |
| 4612 | "bench/dwconv.h", |
| 4613 | "src/xnnpack/AlignedAllocator.h", |
| 4614 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4615 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4616 | ":indirection", |
| 4617 | ":packing", |
| 4618 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4619 | ) |
| 4620 | |
| 4621 | xnnpack_benchmark( |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 4622 | name = "f32_dwconv2d_chw_bench", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4623 | srcs = [ |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 4624 | "bench/f32-dwconv2d-chw.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4625 | "bench/dwconv.h", |
| 4626 | "src/xnnpack/AlignedAllocator.h", |
| 4627 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4628 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4629 | ":indirection", |
| 4630 | ":packing", |
| 4631 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4632 | ) |
| 4633 | |
| 4634 | xnnpack_benchmark( |
| 4635 | name = "f32_gemm_bench", |
| 4636 | srcs = [ |
| 4637 | "bench/f32-gemm.cc", |
| 4638 | "bench/gemm.h", |
| 4639 | "src/xnnpack/AlignedAllocator.h", |
| 4640 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4641 | copts = xnnpack_optional_ruy_copts(), |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4642 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4643 | ) |
| 4644 | |
| 4645 | xnnpack_benchmark( |
Marat Dukhan | a11ca34 | 2020-06-25 23:45:07 -0700 | [diff] [blame] | 4646 | name = "f32_hswish_bench", |
| 4647 | srcs = [ |
| 4648 | "bench/f32-hswish.cc", |
| 4649 | "src/xnnpack/AlignedAllocator.h", |
| 4650 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4651 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4652 | ) |
| 4653 | |
| 4654 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 4655 | name = "f32_raddexpminusmax_bench", |
| 4656 | srcs = [ |
| 4657 | "bench/f32-raddexpminusmax.cc", |
| 4658 | "src/xnnpack/AlignedAllocator.h", |
| 4659 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4660 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4661 | ) |
| 4662 | |
| 4663 | xnnpack_benchmark( |
| 4664 | name = "f32_raddextexp_bench", |
| 4665 | srcs = [ |
| 4666 | "bench/f32-raddextexp.cc", |
| 4667 | "src/xnnpack/AlignedAllocator.h", |
| 4668 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4669 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4670 | ) |
| 4671 | |
| 4672 | xnnpack_benchmark( |
| 4673 | name = "f32_raddstoreexpminusmax_bench", |
| 4674 | srcs = [ |
| 4675 | "bench/f32-raddstoreexpminusmax.cc", |
| 4676 | "src/xnnpack/AlignedAllocator.h", |
| 4677 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4678 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4679 | ) |
| 4680 | |
| 4681 | xnnpack_benchmark( |
Frank Barchard | 3c4a952 | 2020-07-18 16:11:22 -0700 | [diff] [blame] | 4682 | name = "f32_relu_bench", |
| 4683 | srcs = [ |
| 4684 | "bench/f32-relu.cc", |
| 4685 | "src/xnnpack/AlignedAllocator.h", |
| 4686 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4687 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4688 | ) |
| 4689 | |
| 4690 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4691 | name = "f32_rmax_bench", |
| 4692 | srcs = [ |
| 4693 | "bench/f32-rmax.cc", |
| 4694 | "src/xnnpack/AlignedAllocator.h", |
| 4695 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4696 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4697 | ) |
| 4698 | |
| 4699 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 4700 | name = "f32_sigmoid_bench", |
| 4701 | srcs = [ |
| 4702 | "bench/f32-sigmoid.cc", |
| 4703 | "src/xnnpack/AlignedAllocator.h", |
| 4704 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 4705 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4706 | ) |
| 4707 | |
| 4708 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4709 | name = "f32_spmm_bench", |
| 4710 | srcs = [ |
| 4711 | "bench/f32-spmm.cc", |
| 4712 | "bench/gemm.h", |
| 4713 | "src/xnnpack/AlignedAllocator.h", |
| 4714 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4715 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4716 | ) |
| 4717 | |
| 4718 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4719 | name = "f32_softmax_bench", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4720 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4721 | "bench/f32-softmax.cc", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4722 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4723 | copts = xnnpack_optional_dnnl_copts(), |
Marat Dukhan | 8d3c693 | 2020-03-06 20:27:27 -0800 | [diff] [blame] | 4724 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(), |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4725 | ) |
| 4726 | |
| 4727 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 4728 | name = "f32_vscaleexpminusmax_bench", |
| 4729 | srcs = [ |
| 4730 | "bench/f32-vscaleexpminusmax.cc", |
| 4731 | "src/xnnpack/AlignedAllocator.h", |
| 4732 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4733 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4734 | ) |
| 4735 | |
| 4736 | xnnpack_benchmark( |
| 4737 | name = "f32_vscaleextexp_bench", |
| 4738 | srcs = [ |
| 4739 | "bench/f32-vscaleextexp.cc", |
| 4740 | "src/xnnpack/AlignedAllocator.h", |
| 4741 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4742 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4743 | ) |
| 4744 | |
| 4745 | xnnpack_benchmark( |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 4746 | name = "f32_vsqrt_bench", |
| 4747 | srcs = [ |
| 4748 | "bench/f32-vsqrt.cc", |
| 4749 | "src/xnnpack/AlignedAllocator.h", |
| 4750 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4751 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4752 | ) |
| 4753 | |
| 4754 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4755 | name = "f32_im2col_gemm_bench", |
| 4756 | srcs = [ |
| 4757 | "bench/f32-im2col-gemm.cc", |
| 4758 | "bench/conv.h", |
| 4759 | "src/xnnpack/AlignedAllocator.h", |
| 4760 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4761 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4762 | ":im2col", |
| 4763 | ":packing", |
| 4764 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4765 | ) |
| 4766 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 4767 | xnnpack_benchmark( |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4768 | name = "rounding_bench", |
| 4769 | srcs = [ |
| 4770 | "bench/rounding.cc", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4771 | "src/xnnpack/AlignedAllocator.h", |
Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 4772 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4773 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4774 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4775 | ) |
| 4776 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4777 | ########################### Benchmarks for operators ########################### |
| 4778 | |
| 4779 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4780 | name = "average_pooling_bench", |
| 4781 | srcs = ["bench/average-pooling.cc"], |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 4782 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4783 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4784 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4785 | ) |
| 4786 | |
| 4787 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4788 | name = "bankers_rounding_bench", |
| 4789 | srcs = ["bench/bankers-rounding.cc"], |
| 4790 | copts = xnnpack_optional_tflite_copts(), |
| 4791 | tags = ["nowin32"], |
| 4792 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4793 | ) |
| 4794 | |
| 4795 | xnnpack_benchmark( |
| 4796 | name = "ceiling_bench", |
| 4797 | srcs = ["bench/ceiling.cc"], |
| 4798 | copts = xnnpack_optional_tflite_copts(), |
| 4799 | tags = ["nowin32"], |
| 4800 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4801 | ) |
| 4802 | |
| 4803 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4804 | name = "channel_shuffle_bench", |
| 4805 | srcs = ["bench/channel-shuffle.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4806 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4807 | ) |
| 4808 | |
| 4809 | xnnpack_benchmark( |
| 4810 | name = "convolution_bench", |
| 4811 | srcs = ["bench/convolution.cc"], |
| 4812 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4813 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4814 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4815 | ) |
| 4816 | |
| 4817 | xnnpack_benchmark( |
| 4818 | name = "deconvolution_bench", |
| 4819 | srcs = ["bench/deconvolution.cc"], |
| 4820 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4821 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4822 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4823 | ) |
| 4824 | |
| 4825 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4826 | name = "floor_bench", |
| 4827 | srcs = ["bench/floor.cc"], |
| 4828 | copts = xnnpack_optional_tflite_copts(), |
| 4829 | tags = ["nowin32"], |
| 4830 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4831 | ) |
| 4832 | |
| 4833 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4834 | name = "global_average_pooling_bench", |
| 4835 | srcs = ["bench/global-average-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4836 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4837 | ) |
| 4838 | |
| 4839 | xnnpack_benchmark( |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 4840 | name = "hardswish_bench", |
| 4841 | srcs = ["bench/hardswish.cc"], |
| 4842 | copts = xnnpack_optional_tflite_copts(), |
| 4843 | tags = ["nowin32"], |
| 4844 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4845 | ) |
| 4846 | |
| 4847 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4848 | name = "max_pooling_bench", |
| 4849 | srcs = ["bench/max-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4850 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4851 | ) |
| 4852 | |
| 4853 | xnnpack_benchmark( |
| 4854 | name = "sigmoid_bench", |
| 4855 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 4856 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 4857 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4858 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4859 | ) |
| 4860 | |
| 4861 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 4862 | name = "prelu_bench", |
| 4863 | srcs = ["bench/prelu.cc"], |
| 4864 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4865 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4866 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 4867 | ) |
| 4868 | |
| 4869 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4870 | name = "softmax_bench", |
| 4871 | srcs = ["bench/softmax.cc"], |
Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame] | 4872 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 4873 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4874 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4875 | ) |
| 4876 | |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4877 | xnnpack_benchmark( |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 4878 | name = "square_root_bench", |
| 4879 | srcs = ["bench/square-root.cc"], |
| 4880 | copts = xnnpack_optional_tflite_copts(), |
| 4881 | tags = ["nowin32"], |
| 4882 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4883 | ) |
| 4884 | |
| 4885 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4886 | name = "truncation_bench", |
| 4887 | srcs = ["bench/truncation.cc"], |
| 4888 | deps = OPERATOR_BENCHMARK_DEPS, |
| 4889 | ) |
| 4890 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4891 | ############################# End-to-end benchmarks ############################ |
| 4892 | |
| 4893 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4894 | name = "fp32_mobilenet_v1", |
| 4895 | srcs = ["models/fp32-mobilenet-v1.cc"], |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4896 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4897 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4898 | linkstatic = True, |
| 4899 | deps = [ |
| 4900 | ":XNNPACK", |
| 4901 | "@pthreadpool", |
| 4902 | ], |
| 4903 | ) |
| 4904 | |
| 4905 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4906 | name = "fp16_mobilenet_v1", |
| 4907 | srcs = ["models/fp16-mobilenet-v1.cc"], |
| 4908 | hdrs = ["models/models.h"], |
| 4909 | copts = xnnpack_std_cxxopts(), |
| 4910 | linkstatic = True, |
| 4911 | deps = [ |
| 4912 | ":XNNPACK", |
| 4913 | "@FP16", |
| 4914 | "@pthreadpool", |
| 4915 | ], |
| 4916 | ) |
| 4917 | |
| 4918 | cc_library( |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 4919 | name = "qs8_mobilenet_v1", |
| 4920 | srcs = ["models/qs8-mobilenet-v1.cc"], |
| 4921 | hdrs = ["models/models.h"], |
| 4922 | copts = xnnpack_std_cxxopts(), |
| 4923 | linkstatic = True, |
| 4924 | deps = [ |
| 4925 | ":XNNPACK", |
| 4926 | "@pthreadpool", |
| 4927 | ], |
| 4928 | ) |
| 4929 | |
| 4930 | cc_library( |
Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 4931 | name = "qs8_mobilenet_v2", |
| 4932 | srcs = ["models/qs8-mobilenet-v2.cc"], |
| 4933 | hdrs = ["models/models.h"], |
| 4934 | copts = xnnpack_std_cxxopts(), |
| 4935 | linkstatic = True, |
| 4936 | deps = [ |
| 4937 | ":XNNPACK", |
| 4938 | "@pthreadpool", |
| 4939 | ], |
| 4940 | ) |
| 4941 | |
| 4942 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4943 | name = "fp32_mobilenet_v2", |
| 4944 | srcs = ["models/fp32-mobilenet-v2.cc"], |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4945 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4946 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4947 | linkstatic = True, |
| 4948 | deps = [ |
| 4949 | ":XNNPACK", |
| 4950 | "@pthreadpool", |
| 4951 | ], |
| 4952 | ) |
| 4953 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4954 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4955 | name = "fp16_mobilenet_v2", |
| 4956 | srcs = ["models/fp16-mobilenet-v2.cc"], |
| 4957 | hdrs = ["models/models.h"], |
| 4958 | copts = xnnpack_std_cxxopts(), |
| 4959 | linkstatic = True, |
| 4960 | deps = [ |
| 4961 | ":XNNPACK", |
| 4962 | "@FP16", |
| 4963 | "@pthreadpool", |
| 4964 | ], |
| 4965 | ) |
| 4966 | |
| 4967 | cc_library( |
| 4968 | name = "fp32_mobilenet_v3_large", |
| 4969 | srcs = ["models/fp32-mobilenet-v3-large.cc"], |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4970 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4971 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4972 | linkstatic = True, |
| 4973 | deps = [ |
| 4974 | ":XNNPACK", |
| 4975 | "@pthreadpool", |
| 4976 | ], |
| 4977 | ) |
| 4978 | |
| 4979 | cc_library( |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4980 | name = "fp16_mobilenet_v3_large", |
| 4981 | srcs = ["models/fp16-mobilenet-v3-large.cc"], |
| 4982 | hdrs = ["models/models.h"], |
| 4983 | copts = xnnpack_std_cxxopts(), |
| 4984 | linkstatic = True, |
| 4985 | deps = [ |
| 4986 | ":XNNPACK", |
| 4987 | "@FP16", |
| 4988 | "@pthreadpool", |
| 4989 | ], |
| 4990 | ) |
| 4991 | |
| 4992 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4993 | name = "fp32_mobilenet_v3_small", |
| 4994 | srcs = ["models/fp32-mobilenet-v3-small.cc"], |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4995 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4996 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4997 | linkstatic = True, |
| 4998 | deps = [ |
| 4999 | ":XNNPACK", |
| 5000 | "@pthreadpool", |
| 5001 | ], |
| 5002 | ) |
| 5003 | |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 5004 | cc_library( |
| 5005 | name = "fp16_mobilenet_v3_small", |
| 5006 | srcs = ["models/fp16-mobilenet-v3-small.cc"], |
| 5007 | hdrs = ["models/models.h"], |
| 5008 | copts = xnnpack_std_cxxopts(), |
| 5009 | linkstatic = True, |
| 5010 | deps = [ |
| 5011 | ":XNNPACK", |
| 5012 | "@FP16", |
| 5013 | "@pthreadpool", |
| 5014 | ], |
| 5015 | ) |
| 5016 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 5017 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 5018 | name = "f32_dwconv_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 5019 | srcs = [ |
| 5020 | "bench/f32-dwconv-e2e.cc", |
| 5021 | "bench/end2end.h", |
| 5022 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 5023 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 5024 | ":XNNPACK", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 5025 | ":fp32_mobilenet_v1", |
| 5026 | ":fp32_mobilenet_v2", |
| 5027 | ":fp32_mobilenet_v3_large", |
| 5028 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 5029 | ], |
| 5030 | ) |
| 5031 | |
| 5032 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 5033 | name = "f32_gemm_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 5034 | srcs = [ |
| 5035 | "bench/f32-gemm-e2e.cc", |
| 5036 | "bench/end2end.h", |
| 5037 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 5038 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 5039 | ":XNNPACK", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 5040 | ":fp32_mobilenet_v1", |
| 5041 | ":fp32_mobilenet_v2", |
| 5042 | ":fp32_mobilenet_v3_large", |
| 5043 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 5044 | ], |
| 5045 | ) |
| 5046 | |
| 5047 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 5048 | name = "end2end_bench", |
| 5049 | srcs = ["bench/end2end.cc"], |
| 5050 | deps = [ |
| 5051 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 5052 | ":bench_utils", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 5053 | ":fp16_mobilenet_v1", |
| 5054 | ":fp16_mobilenet_v2", |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 5055 | ":fp16_mobilenet_v3_large", |
| 5056 | ":fp16_mobilenet_v3_small", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 5057 | ":fp32_mobilenet_v1", |
| 5058 | ":fp32_mobilenet_v2", |
| 5059 | ":fp32_mobilenet_v3_large", |
| 5060 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 5061 | ":qs8_mobilenet_v1", |
Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 5062 | ":qs8_mobilenet_v2", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 5063 | "@pthreadpool", |
| 5064 | ], |
| 5065 | ) |
| 5066 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 5067 | #################### Accuracy evaluation for math functions #################### |
| 5068 | |
| 5069 | xnnpack_benchmark( |
| 5070 | name = "f32_exp_eval", |
| 5071 | srcs = [ |
| 5072 | "eval/f32-exp.cc", |
| 5073 | "src/xnnpack/AlignedAllocator.h", |
| 5074 | ] + ACCURACY_EVAL_HDRS, |
| 5075 | deps = ACCURACY_EVAL_DEPS, |
| 5076 | ) |
| 5077 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 5078 | xnnpack_benchmark( |
| 5079 | name = "f32_expminus_eval", |
| 5080 | srcs = [ |
| 5081 | "eval/f32-expminus.cc", |
| 5082 | "src/xnnpack/AlignedAllocator.h", |
| 5083 | ] + ACCURACY_EVAL_HDRS, |
| 5084 | deps = ACCURACY_EVAL_DEPS, |
| 5085 | ) |
| 5086 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 5087 | xnnpack_benchmark( |
Marat Dukhan | a438aca | 2020-11-20 15:45:01 -0800 | [diff] [blame] | 5088 | name = "f32_expm1minus_eval", |
| 5089 | srcs = [ |
| 5090 | "eval/f32-expm1minus.cc", |
| 5091 | "src/xnnpack/AlignedAllocator.h", |
| 5092 | ] + ACCURACY_EVAL_HDRS, |
| 5093 | deps = ACCURACY_EVAL_DEPS, |
| 5094 | ) |
| 5095 | |
| 5096 | xnnpack_benchmark( |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 5097 | name = "f32_extexp_eval", |
| 5098 | srcs = [ |
| 5099 | "eval/f32-extexp.cc", |
| 5100 | "src/xnnpack/AlignedAllocator.h", |
| 5101 | ] + ACCURACY_EVAL_HDRS, |
| 5102 | deps = ACCURACY_EVAL_DEPS, |
| 5103 | ) |
| 5104 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 5105 | xnnpack_unit_test( |
| 5106 | name = "f32_roundne_eval", |
| 5107 | srcs = [ |
| 5108 | "eval/f32-roundne.cc", |
| 5109 | "src/xnnpack/AlignedAllocator.h", |
| 5110 | "src/xnnpack/math-stubs.h", |
| 5111 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | 22eed3d | 2020-05-11 20:13:37 -0700 | [diff] [blame] | 5112 | automatic = False, |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 5113 | deps = MICROKERNEL_TEST_DEPS, |
| 5114 | ) |
| 5115 | |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 5116 | xnnpack_unit_test( |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 5117 | name = "f32_roundd_eval", |
| 5118 | srcs = [ |
| 5119 | "eval/f32-roundd.cc", |
| 5120 | "src/xnnpack/AlignedAllocator.h", |
| 5121 | "src/xnnpack/math-stubs.h", |
| 5122 | ] + MICROKERNEL_TEST_HDRS, |
| 5123 | automatic = False, |
| 5124 | deps = MICROKERNEL_TEST_DEPS, |
| 5125 | ) |
| 5126 | |
| 5127 | xnnpack_unit_test( |
| 5128 | name = "f32_roundu_eval", |
| 5129 | srcs = [ |
| 5130 | "eval/f32-roundu.cc", |
| 5131 | "src/xnnpack/AlignedAllocator.h", |
| 5132 | "src/xnnpack/math-stubs.h", |
| 5133 | ] + MICROKERNEL_TEST_HDRS, |
| 5134 | automatic = False, |
| 5135 | deps = MICROKERNEL_TEST_DEPS, |
| 5136 | ) |
| 5137 | |
| 5138 | xnnpack_unit_test( |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 5139 | name = "f32_roundz_eval", |
| 5140 | srcs = [ |
| 5141 | "eval/f32-roundz.cc", |
| 5142 | "src/xnnpack/AlignedAllocator.h", |
| 5143 | "src/xnnpack/math-stubs.h", |
| 5144 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 5145 | automatic = False, |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 5146 | deps = MICROKERNEL_TEST_DEPS, |
| 5147 | ) |
| 5148 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5149 | xnnpack_benchmark( |
| 5150 | name = "f32_sigmoid_eval", |
| 5151 | srcs = [ |
| 5152 | "eval/f32-sigmoid.cc", |
| 5153 | "src/xnnpack/AlignedAllocator.h", |
| 5154 | ] + ACCURACY_EVAL_HDRS, |
| 5155 | deps = ACCURACY_EVAL_DEPS, |
| 5156 | ) |
| 5157 | |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 5158 | xnnpack_benchmark( |
| 5159 | name = "f32_sqrt_eval", |
| 5160 | srcs = [ |
| 5161 | "eval/f32-sqrt.cc", |
| 5162 | "src/xnnpack/AlignedAllocator.h", |
| 5163 | ] + ACCURACY_EVAL_HDRS, |
| 5164 | deps = ACCURACY_EVAL_DEPS + [ |
| 5165 | ":bench_utils", |
| 5166 | ], |
| 5167 | ) |
| 5168 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5169 | ######################### Unit tests for micro-kernels ######################### |
| 5170 | |
| 5171 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 5172 | name = "f16_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5173 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 5174 | "test/f16-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5175 | "test/gemm-microkernel-tester.h", |
| 5176 | "src/xnnpack/AlignedAllocator.h", |
| 5177 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5178 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5179 | ) |
| 5180 | |
| 5181 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5182 | name = "f16_spmm_minmax_test", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 5183 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5184 | "test/f16-spmm-minmax.cc", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 5185 | "test/spmm-microkernel-tester.h", |
| 5186 | "src/xnnpack/AlignedAllocator.h", |
| 5187 | ] + MICROKERNEL_TEST_HDRS, |
| 5188 | deps = MICROKERNEL_TEST_DEPS, |
| 5189 | ) |
| 5190 | |
| 5191 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5192 | name = "f32_argmaxpool_test", |
| 5193 | srcs = [ |
| 5194 | "test/f32-argmaxpool.cc", |
| 5195 | "test/argmaxpool-microkernel-tester.h", |
| 5196 | "src/xnnpack/AlignedAllocator.h", |
| 5197 | ] + MICROKERNEL_TEST_HDRS, |
| 5198 | deps = MICROKERNEL_TEST_DEPS, |
| 5199 | ) |
| 5200 | |
| 5201 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5202 | name = "f32_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5203 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5204 | "test/f32-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5205 | "test/avgpool-microkernel-tester.h", |
| 5206 | "src/xnnpack/AlignedAllocator.h", |
| 5207 | ] + MICROKERNEL_TEST_HDRS, |
| 5208 | deps = MICROKERNEL_TEST_DEPS, |
| 5209 | ) |
| 5210 | |
| 5211 | xnnpack_unit_test( |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 5212 | name = "f32_ibilinear_test", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 5213 | srcs = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 5214 | "test/f32-ibilinear.cc", |
| 5215 | "test/ibilinear-microkernel-tester.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 5216 | "src/xnnpack/AlignedAllocator.h", |
| 5217 | ] + MICROKERNEL_TEST_HDRS, |
| 5218 | deps = MICROKERNEL_TEST_DEPS, |
| 5219 | ) |
| 5220 | |
| 5221 | xnnpack_unit_test( |
XNNPACK Team | 2143267 | 2020-10-19 19:58:48 -0700 | [diff] [blame] | 5222 | name = "f32_ibilinear_chw_test", |
| 5223 | srcs = [ |
| 5224 | "test/f32-ibilinear-chw.cc", |
XNNPACK Team | 6be46b2 | 2020-10-22 23:34:54 -0700 | [diff] [blame] | 5225 | "test/ibilinear-microkernel-tester.h", |
XNNPACK Team | 2143267 | 2020-10-19 19:58:48 -0700 | [diff] [blame] | 5226 | "src/xnnpack/AlignedAllocator.h", |
| 5227 | ] + MICROKERNEL_TEST_HDRS, |
| 5228 | deps = MICROKERNEL_TEST_DEPS, |
| 5229 | ) |
| 5230 | |
| 5231 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5232 | name = "f16_clamp_test", |
| 5233 | srcs = [ |
| 5234 | "test/f16-clamp.cc", |
| 5235 | "test/clamp-microkernel-tester.h", |
| 5236 | ] + MICROKERNEL_TEST_HDRS, |
| 5237 | deps = MICROKERNEL_TEST_DEPS, |
| 5238 | ) |
| 5239 | |
| 5240 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5241 | name = "f32_clamp_test", |
| 5242 | srcs = [ |
| 5243 | "test/f32-clamp.cc", |
| 5244 | "test/clamp-microkernel-tester.h", |
| 5245 | ] + MICROKERNEL_TEST_HDRS, |
| 5246 | deps = MICROKERNEL_TEST_DEPS, |
| 5247 | ) |
| 5248 | |
| 5249 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5250 | name = "f32_igemm_test", |
| 5251 | srcs = [ |
| 5252 | "test/f32-igemm.cc", |
| 5253 | "test/gemm-microkernel-tester.h", |
| 5254 | "src/xnnpack/AlignedAllocator.h", |
| 5255 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5256 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5257 | ) |
| 5258 | |
| 5259 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5260 | name = "f32_igemm_relu_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5261 | srcs = [ |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5262 | "test/f32-igemm-relu.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5263 | "test/gemm-microkernel-tester.h", |
| 5264 | "src/xnnpack/AlignedAllocator.h", |
| 5265 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5266 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5267 | ) |
| 5268 | |
| 5269 | xnnpack_unit_test( |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 5270 | name = "f32_igemm_minmax_test", |
| 5271 | srcs = [ |
| 5272 | "test/f32-igemm-minmax.cc", |
| 5273 | "test/gemm-microkernel-tester.h", |
| 5274 | "src/xnnpack/AlignedAllocator.h", |
| 5275 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5276 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 5277 | ) |
| 5278 | |
| 5279 | xnnpack_unit_test( |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 5280 | name = "f16_igemm_minmax_test", |
| 5281 | srcs = [ |
| 5282 | "test/f16-igemm-minmax.cc", |
| 5283 | "test/gemm-microkernel-tester.h", |
| 5284 | "src/xnnpack/AlignedAllocator.h", |
| 5285 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5286 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 5287 | ) |
| 5288 | |
| 5289 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5290 | name = "f32_conv_hwc_test", |
| 5291 | srcs = [ |
| 5292 | "test/f32-conv-hwc.cc", |
| 5293 | "test/conv-hwc-microkernel-tester.h", |
| 5294 | "src/xnnpack/AlignedAllocator.h", |
| 5295 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5296 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5297 | ) |
| 5298 | |
| 5299 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5300 | name = "f32_conv_hwc2chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5301 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5302 | "test/f32-conv-hwc2chw.cc", |
| 5303 | "test/conv-hwc2chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5304 | "src/xnnpack/AlignedAllocator.h", |
| 5305 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5306 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5307 | ) |
| 5308 | |
| 5309 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5310 | name = "f32_dwconv_test", |
| 5311 | srcs = [ |
| 5312 | "test/f32-dwconv.cc", |
| 5313 | "test/dwconv-microkernel-tester.h", |
| 5314 | "src/xnnpack/AlignedAllocator.h", |
| 5315 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5316 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5317 | ) |
| 5318 | |
| 5319 | xnnpack_unit_test( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 5320 | name = "f16_dwconv_minmax_test", |
| 5321 | srcs = [ |
| 5322 | "test/f16-dwconv-minmax.cc", |
| 5323 | "test/dwconv-microkernel-tester.h", |
| 5324 | "src/xnnpack/AlignedAllocator.h", |
| 5325 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5326 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 5327 | ) |
| 5328 | |
| 5329 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5330 | name = "f32_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5331 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5332 | "test/f32-dwconv-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5333 | "test/dwconv-microkernel-tester.h", |
| 5334 | "src/xnnpack/AlignedAllocator.h", |
| 5335 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5336 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5337 | ) |
| 5338 | |
| 5339 | xnnpack_unit_test( |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 5340 | name = "f32_dwconv2d_chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5341 | srcs = [ |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 5342 | "test/f32-dwconv2d-chw.cc", |
| 5343 | "test/dwconv2d-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5344 | "src/xnnpack/AlignedAllocator.h", |
| 5345 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5346 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5347 | ) |
| 5348 | |
| 5349 | xnnpack_unit_test( |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 5350 | name = "f16_gavgpool_minmax_test", |
| 5351 | srcs = [ |
| 5352 | "test/f16-gavgpool-minmax.cc", |
| 5353 | "test/gavgpool-microkernel-tester.h", |
| 5354 | "src/xnnpack/AlignedAllocator.h", |
| 5355 | ] + MICROKERNEL_TEST_HDRS, |
| 5356 | deps = MICROKERNEL_TEST_DEPS, |
| 5357 | ) |
| 5358 | |
| 5359 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5360 | name = "f32_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5361 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5362 | "test/f32-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5363 | "test/gavgpool-microkernel-tester.h", |
| 5364 | "src/xnnpack/AlignedAllocator.h", |
| 5365 | ] + MICROKERNEL_TEST_HDRS, |
| 5366 | deps = MICROKERNEL_TEST_DEPS, |
| 5367 | ) |
| 5368 | |
| 5369 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5370 | name = "f32_gavgpool_cw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5371 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5372 | "test/f32-gavgpool-cw.cc", |
| 5373 | "test/gavgpool-cw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5374 | "src/xnnpack/AlignedAllocator.h", |
| 5375 | ] + MICROKERNEL_TEST_HDRS, |
| 5376 | deps = MICROKERNEL_TEST_DEPS, |
| 5377 | ) |
| 5378 | |
| 5379 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5380 | name = "f32_gemm_test", |
| 5381 | srcs = [ |
| 5382 | "test/f32-gemm.cc", |
| 5383 | "test/gemm-microkernel-tester.h", |
| 5384 | "src/xnnpack/AlignedAllocator.h", |
| 5385 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5386 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5387 | ) |
| 5388 | |
| 5389 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5390 | name = "f32_gemm_relu_test", |
| 5391 | srcs = [ |
| 5392 | "test/f32-gemm-relu.cc", |
| 5393 | "test/gemm-microkernel-tester.h", |
| 5394 | "src/xnnpack/AlignedAllocator.h", |
| 5395 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5396 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5397 | ) |
| 5398 | |
| 5399 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5400 | name = "f32_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5401 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5402 | "test/f32-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5403 | "test/gemm-microkernel-tester.h", |
| 5404 | "src/xnnpack/AlignedAllocator.h", |
| 5405 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5406 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5407 | ) |
| 5408 | |
| 5409 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5410 | name = "f32_gemminc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5411 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5412 | "test/f32-gemminc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5413 | "test/gemm-microkernel-tester.h", |
| 5414 | "src/xnnpack/AlignedAllocator.h", |
| 5415 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5416 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5417 | ) |
| 5418 | |
| 5419 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5420 | name = "f16_hswish_test", |
| 5421 | srcs = [ |
| 5422 | "test/f16-hswish.cc", |
| 5423 | "test/hswish-microkernel-tester.h", |
| 5424 | ] + MICROKERNEL_TEST_HDRS, |
| 5425 | deps = MICROKERNEL_TEST_DEPS, |
| 5426 | ) |
| 5427 | |
| 5428 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5429 | name = "f32_hswish_test", |
| 5430 | srcs = [ |
| 5431 | "test/f32-hswish.cc", |
| 5432 | "test/hswish-microkernel-tester.h", |
| 5433 | ] + MICROKERNEL_TEST_HDRS, |
| 5434 | deps = MICROKERNEL_TEST_DEPS, |
| 5435 | ) |
| 5436 | |
| 5437 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5438 | name = "f32_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5439 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5440 | "test/f32-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5441 | "test/maxpool-microkernel-tester.h", |
| 5442 | ] + MICROKERNEL_TEST_HDRS, |
| 5443 | deps = MICROKERNEL_TEST_DEPS, |
| 5444 | ) |
| 5445 | |
| 5446 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5447 | name = "f32_pavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5448 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5449 | "test/f32-pavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5450 | "test/avgpool-microkernel-tester.h", |
| 5451 | "src/xnnpack/AlignedAllocator.h", |
| 5452 | ] + MICROKERNEL_TEST_HDRS, |
| 5453 | deps = MICROKERNEL_TEST_DEPS, |
| 5454 | ) |
| 5455 | |
| 5456 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5457 | name = "f32_ppmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5458 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5459 | "test/f32-ppmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5460 | "test/gemm-microkernel-tester.h", |
| 5461 | "src/xnnpack/AlignedAllocator.h", |
| 5462 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5463 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5464 | ) |
| 5465 | |
| 5466 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5467 | name = "f16_prelu_test", |
| 5468 | srcs = [ |
| 5469 | "test/f16-prelu.cc", |
| 5470 | "test/prelu-microkernel-tester.h", |
| 5471 | "src/xnnpack/AlignedAllocator.h", |
| 5472 | ] + MICROKERNEL_TEST_HDRS, |
| 5473 | deps = MICROKERNEL_TEST_DEPS, |
| 5474 | ) |
| 5475 | |
| 5476 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5477 | name = "f32_prelu_test", |
| 5478 | srcs = [ |
| 5479 | "test/f32-prelu.cc", |
| 5480 | "test/prelu-microkernel-tester.h", |
| 5481 | "src/xnnpack/AlignedAllocator.h", |
| 5482 | ] + MICROKERNEL_TEST_HDRS, |
| 5483 | deps = MICROKERNEL_TEST_DEPS, |
| 5484 | ) |
| 5485 | |
| 5486 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5487 | name = "f32_raddexpminusmax_test", |
| 5488 | srcs = [ |
| 5489 | "test/f32-raddexpminusmax.cc", |
| 5490 | "test/raddexpminusmax-microkernel-tester.h", |
| 5491 | ] + MICROKERNEL_TEST_HDRS, |
| 5492 | deps = MICROKERNEL_TEST_DEPS, |
| 5493 | ) |
| 5494 | |
| 5495 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 5496 | name = "f32_raddextexp_test", |
| 5497 | srcs = [ |
| 5498 | "test/f32-raddextexp.cc", |
| 5499 | "test/raddextexp-microkernel-tester.h", |
| 5500 | ] + MICROKERNEL_TEST_HDRS, |
| 5501 | deps = MICROKERNEL_TEST_DEPS, |
| 5502 | ) |
| 5503 | |
| 5504 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5505 | name = "f32_raddstoreexpminusmax_test", |
| 5506 | srcs = [ |
| 5507 | "test/f32-raddstoreexpminusmax.cc", |
| 5508 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 5509 | ] + MICROKERNEL_TEST_HDRS, |
| 5510 | deps = MICROKERNEL_TEST_DEPS, |
| 5511 | ) |
| 5512 | |
| 5513 | xnnpack_unit_test( |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 5514 | name = "f32_relu_test", |
| 5515 | srcs = [ |
| 5516 | "test/f32-relu.cc", |
| 5517 | "test/vunary-microkernel-tester.h", |
| 5518 | ] + MICROKERNEL_TEST_HDRS, |
| 5519 | deps = MICROKERNEL_TEST_DEPS, |
| 5520 | ) |
| 5521 | |
| 5522 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5523 | name = "f32_rmax_test", |
| 5524 | srcs = [ |
| 5525 | "test/f32-rmax.cc", |
| 5526 | "test/rmax-microkernel-tester.h", |
| 5527 | ] + MICROKERNEL_TEST_HDRS, |
| 5528 | deps = MICROKERNEL_TEST_DEPS, |
| 5529 | ) |
| 5530 | |
| 5531 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5532 | name = "f32_sigmoid_test", |
| 5533 | srcs = [ |
| 5534 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5535 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5536 | ] + MICROKERNEL_TEST_HDRS, |
| 5537 | deps = MICROKERNEL_TEST_DEPS, |
| 5538 | ) |
| 5539 | |
| 5540 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5541 | name = "f32_spmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5542 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5543 | "test/f32-spmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5544 | "test/spmm-microkernel-tester.h", |
| 5545 | "src/xnnpack/AlignedAllocator.h", |
| 5546 | ] + MICROKERNEL_TEST_HDRS, |
| 5547 | deps = MICROKERNEL_TEST_DEPS, |
| 5548 | ) |
| 5549 | |
| 5550 | xnnpack_unit_test( |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 5551 | name = "f16_vadd_minmax_test", |
| 5552 | srcs = [ |
| 5553 | "test/f16-vadd-minmax.cc", |
| 5554 | "test/vbinary-microkernel-tester.h", |
| 5555 | ] + MICROKERNEL_TEST_HDRS, |
| 5556 | deps = MICROKERNEL_TEST_DEPS, |
| 5557 | ) |
| 5558 | |
| 5559 | xnnpack_unit_test( |
| 5560 | name = "f16_vaddc_minmax_test", |
| 5561 | srcs = [ |
| 5562 | "test/f16-vaddc-minmax.cc", |
| 5563 | "test/vbinaryc-microkernel-tester.h", |
| 5564 | ] + MICROKERNEL_TEST_HDRS, |
| 5565 | deps = MICROKERNEL_TEST_DEPS, |
| 5566 | ) |
| 5567 | |
| 5568 | xnnpack_unit_test( |
| 5569 | name = "f16_vdiv_minmax_test", |
| 5570 | srcs = [ |
| 5571 | "test/f16-vdiv-minmax.cc", |
| 5572 | "test/vbinary-microkernel-tester.h", |
| 5573 | ] + MICROKERNEL_TEST_HDRS, |
| 5574 | deps = MICROKERNEL_TEST_DEPS, |
| 5575 | ) |
| 5576 | |
| 5577 | xnnpack_unit_test( |
| 5578 | name = "f16_vdivc_minmax_test", |
| 5579 | srcs = [ |
| 5580 | "test/f16-vdivc-minmax.cc", |
| 5581 | "test/vbinaryc-microkernel-tester.h", |
| 5582 | ] + MICROKERNEL_TEST_HDRS, |
| 5583 | deps = MICROKERNEL_TEST_DEPS, |
| 5584 | ) |
| 5585 | |
| 5586 | xnnpack_unit_test( |
| 5587 | name = "f16_vrdivc_minmax_test", |
| 5588 | srcs = [ |
| 5589 | "test/f16-vrdivc-minmax.cc", |
| 5590 | "test/vbinaryc-microkernel-tester.h", |
| 5591 | ] + MICROKERNEL_TEST_HDRS, |
| 5592 | deps = MICROKERNEL_TEST_DEPS, |
| 5593 | ) |
| 5594 | |
| 5595 | xnnpack_unit_test( |
| 5596 | name = "f16_vmax_test", |
| 5597 | srcs = [ |
| 5598 | "test/f16-vmax.cc", |
| 5599 | "test/vbinary-microkernel-tester.h", |
| 5600 | ] + MICROKERNEL_TEST_HDRS, |
| 5601 | deps = MICROKERNEL_TEST_DEPS, |
| 5602 | ) |
| 5603 | |
| 5604 | xnnpack_unit_test( |
| 5605 | name = "f16_vmaxc_test", |
| 5606 | srcs = [ |
| 5607 | "test/f16-vmaxc.cc", |
| 5608 | "test/vbinaryc-microkernel-tester.h", |
| 5609 | ] + MICROKERNEL_TEST_HDRS, |
| 5610 | deps = MICROKERNEL_TEST_DEPS, |
| 5611 | ) |
| 5612 | |
| 5613 | xnnpack_unit_test( |
| 5614 | name = "f16_vmin_test", |
| 5615 | srcs = [ |
| 5616 | "test/f16-vmin.cc", |
| 5617 | "test/vbinary-microkernel-tester.h", |
| 5618 | ] + MICROKERNEL_TEST_HDRS, |
| 5619 | deps = MICROKERNEL_TEST_DEPS, |
| 5620 | ) |
| 5621 | |
| 5622 | xnnpack_unit_test( |
| 5623 | name = "f16_vminc_test", |
| 5624 | srcs = [ |
| 5625 | "test/f16-vminc.cc", |
| 5626 | "test/vbinaryc-microkernel-tester.h", |
| 5627 | ] + MICROKERNEL_TEST_HDRS, |
| 5628 | deps = MICROKERNEL_TEST_DEPS, |
| 5629 | ) |
| 5630 | |
| 5631 | xnnpack_unit_test( |
| 5632 | name = "f16_vmul_minmax_test", |
| 5633 | srcs = [ |
| 5634 | "test/f16-vmul-minmax.cc", |
| 5635 | "test/vbinary-microkernel-tester.h", |
| 5636 | ] + MICROKERNEL_TEST_HDRS, |
| 5637 | deps = MICROKERNEL_TEST_DEPS, |
| 5638 | ) |
| 5639 | |
| 5640 | xnnpack_unit_test( |
| 5641 | name = "f16_vmulc_minmax_test", |
| 5642 | srcs = [ |
| 5643 | "test/f16-vmulc-minmax.cc", |
| 5644 | "test/vbinaryc-microkernel-tester.h", |
| 5645 | ] + MICROKERNEL_TEST_HDRS, |
| 5646 | deps = MICROKERNEL_TEST_DEPS, |
| 5647 | ) |
| 5648 | |
| 5649 | xnnpack_unit_test( |
| 5650 | name = "f16_vsub_minmax_test", |
| 5651 | srcs = [ |
| 5652 | "test/f16-vsub-minmax.cc", |
| 5653 | "test/vbinary-microkernel-tester.h", |
| 5654 | ] + MICROKERNEL_TEST_HDRS, |
| 5655 | deps = MICROKERNEL_TEST_DEPS, |
| 5656 | ) |
| 5657 | |
| 5658 | xnnpack_unit_test( |
| 5659 | name = "f16_vsubc_minmax_test", |
| 5660 | srcs = [ |
| 5661 | "test/f16-vsubc-minmax.cc", |
| 5662 | "test/vbinaryc-microkernel-tester.h", |
| 5663 | ] + MICROKERNEL_TEST_HDRS, |
| 5664 | deps = MICROKERNEL_TEST_DEPS, |
| 5665 | ) |
| 5666 | |
| 5667 | xnnpack_unit_test( |
| 5668 | name = "f16_vrsubc_minmax_test", |
| 5669 | srcs = [ |
| 5670 | "test/f16-vrsubc-minmax.cc", |
| 5671 | "test/vbinaryc-microkernel-tester.h", |
| 5672 | ] + MICROKERNEL_TEST_HDRS, |
| 5673 | deps = MICROKERNEL_TEST_DEPS, |
| 5674 | ) |
| 5675 | |
| 5676 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5677 | name = "f32_vabs_test", |
| 5678 | srcs = [ |
| 5679 | "test/f32-vabs.cc", |
| 5680 | "test/vunary-microkernel-tester.h", |
| 5681 | ] + MICROKERNEL_TEST_HDRS, |
| 5682 | deps = MICROKERNEL_TEST_DEPS, |
| 5683 | ) |
| 5684 | |
| 5685 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5686 | name = "f32_vadd_test", |
| 5687 | srcs = [ |
| 5688 | "test/f32-vadd.cc", |
| 5689 | "test/vbinary-microkernel-tester.h", |
| 5690 | ] + MICROKERNEL_TEST_HDRS, |
| 5691 | deps = MICROKERNEL_TEST_DEPS, |
| 5692 | ) |
| 5693 | |
| 5694 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5695 | name = "f32_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5696 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5697 | "test/f32-vadd-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5698 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5699 | ] + MICROKERNEL_TEST_HDRS, |
| 5700 | deps = MICROKERNEL_TEST_DEPS, |
| 5701 | ) |
| 5702 | |
| 5703 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5704 | name = "f32_vadd_relu_test", |
| 5705 | srcs = [ |
| 5706 | "test/f32-vadd-relu.cc", |
| 5707 | "test/vbinary-microkernel-tester.h", |
| 5708 | ] + MICROKERNEL_TEST_HDRS, |
| 5709 | deps = MICROKERNEL_TEST_DEPS, |
| 5710 | ) |
| 5711 | |
| 5712 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5713 | name = "f32_vaddc_test", |
| 5714 | srcs = [ |
| 5715 | "test/f32-vaddc.cc", |
| 5716 | "test/vbinaryc-microkernel-tester.h", |
| 5717 | ] + MICROKERNEL_TEST_HDRS, |
| 5718 | deps = MICROKERNEL_TEST_DEPS, |
| 5719 | ) |
| 5720 | |
| 5721 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5722 | name = "f32_vaddc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5723 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5724 | "test/f32-vaddc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5725 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5726 | ] + MICROKERNEL_TEST_HDRS, |
| 5727 | deps = MICROKERNEL_TEST_DEPS, |
| 5728 | ) |
| 5729 | |
| 5730 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5731 | name = "f32_vaddc_relu_test", |
| 5732 | srcs = [ |
| 5733 | "test/f32-vaddc-relu.cc", |
| 5734 | "test/vbinaryc-microkernel-tester.h", |
| 5735 | ] + MICROKERNEL_TEST_HDRS, |
| 5736 | deps = MICROKERNEL_TEST_DEPS, |
| 5737 | ) |
| 5738 | |
| 5739 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5740 | name = "f32_vdiv_test", |
| 5741 | srcs = [ |
| 5742 | "test/f32-vdiv.cc", |
| 5743 | "test/vbinary-microkernel-tester.h", |
| 5744 | ] + MICROKERNEL_TEST_HDRS, |
| 5745 | deps = MICROKERNEL_TEST_DEPS, |
| 5746 | ) |
| 5747 | |
| 5748 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5749 | name = "f32_vdiv_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5750 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5751 | "test/f32-vdiv-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5752 | "test/vbinary-microkernel-tester.h", |
| 5753 | ] + MICROKERNEL_TEST_HDRS, |
| 5754 | deps = MICROKERNEL_TEST_DEPS, |
| 5755 | ) |
| 5756 | |
| 5757 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5758 | name = "f32_vdiv_relu_test", |
| 5759 | srcs = [ |
| 5760 | "test/f32-vdiv-relu.cc", |
| 5761 | "test/vbinary-microkernel-tester.h", |
| 5762 | ] + MICROKERNEL_TEST_HDRS, |
| 5763 | deps = MICROKERNEL_TEST_DEPS, |
| 5764 | ) |
| 5765 | |
| 5766 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5767 | name = "f32_vdivc_test", |
| 5768 | srcs = [ |
| 5769 | "test/f32-vdivc.cc", |
| 5770 | "test/vbinaryc-microkernel-tester.h", |
| 5771 | ] + MICROKERNEL_TEST_HDRS, |
| 5772 | deps = MICROKERNEL_TEST_DEPS, |
| 5773 | ) |
| 5774 | |
| 5775 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5776 | name = "f32_vdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5777 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5778 | "test/f32-vdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5779 | "test/vbinaryc-microkernel-tester.h", |
| 5780 | ] + MICROKERNEL_TEST_HDRS, |
| 5781 | deps = MICROKERNEL_TEST_DEPS, |
| 5782 | ) |
| 5783 | |
| 5784 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5785 | name = "f32_vdivc_relu_test", |
| 5786 | srcs = [ |
| 5787 | "test/f32-vdivc-relu.cc", |
| 5788 | "test/vbinaryc-microkernel-tester.h", |
| 5789 | ] + MICROKERNEL_TEST_HDRS, |
| 5790 | deps = MICROKERNEL_TEST_DEPS, |
| 5791 | ) |
| 5792 | |
| 5793 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5794 | name = "f32_vrdivc_test", |
| 5795 | srcs = [ |
| 5796 | "test/f32-vrdivc.cc", |
| 5797 | "test/vbinaryc-microkernel-tester.h", |
| 5798 | ] + MICROKERNEL_TEST_HDRS, |
| 5799 | deps = MICROKERNEL_TEST_DEPS, |
| 5800 | ) |
| 5801 | |
| 5802 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5803 | name = "f32_vrdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5804 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5805 | "test/f32-vrdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5806 | "test/vbinaryc-microkernel-tester.h", |
| 5807 | ] + MICROKERNEL_TEST_HDRS, |
| 5808 | deps = MICROKERNEL_TEST_DEPS, |
| 5809 | ) |
| 5810 | |
| 5811 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5812 | name = "f32_vrdivc_relu_test", |
| 5813 | srcs = [ |
| 5814 | "test/f32-vrdivc-relu.cc", |
| 5815 | "test/vbinaryc-microkernel-tester.h", |
| 5816 | ] + MICROKERNEL_TEST_HDRS, |
| 5817 | deps = MICROKERNEL_TEST_DEPS, |
| 5818 | ) |
| 5819 | |
| 5820 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 5821 | name = "f32_vmax_test", |
| 5822 | srcs = [ |
| 5823 | "test/f32-vmax.cc", |
| 5824 | "test/vbinary-microkernel-tester.h", |
| 5825 | ] + MICROKERNEL_TEST_HDRS, |
| 5826 | deps = MICROKERNEL_TEST_DEPS, |
| 5827 | ) |
| 5828 | |
| 5829 | xnnpack_unit_test( |
| 5830 | name = "f32_vmaxc_test", |
| 5831 | srcs = [ |
| 5832 | "test/f32-vmaxc.cc", |
| 5833 | "test/vbinaryc-microkernel-tester.h", |
| 5834 | ] + MICROKERNEL_TEST_HDRS, |
| 5835 | deps = MICROKERNEL_TEST_DEPS, |
| 5836 | ) |
| 5837 | |
| 5838 | xnnpack_unit_test( |
| 5839 | name = "f32_vmin_test", |
| 5840 | srcs = [ |
| 5841 | "test/f32-vmin.cc", |
| 5842 | "test/vbinary-microkernel-tester.h", |
| 5843 | ] + MICROKERNEL_TEST_HDRS, |
| 5844 | deps = MICROKERNEL_TEST_DEPS, |
| 5845 | ) |
| 5846 | |
| 5847 | xnnpack_unit_test( |
| 5848 | name = "f32_vminc_test", |
| 5849 | srcs = [ |
| 5850 | "test/f32-vminc.cc", |
| 5851 | "test/vbinaryc-microkernel-tester.h", |
| 5852 | ] + MICROKERNEL_TEST_HDRS, |
| 5853 | deps = MICROKERNEL_TEST_DEPS, |
| 5854 | ) |
| 5855 | |
| 5856 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5857 | name = "f32_vmul_test", |
| 5858 | srcs = [ |
| 5859 | "test/f32-vmul.cc", |
| 5860 | "test/vbinary-microkernel-tester.h", |
| 5861 | ] + MICROKERNEL_TEST_HDRS, |
| 5862 | deps = MICROKERNEL_TEST_DEPS, |
| 5863 | ) |
| 5864 | |
| 5865 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5866 | name = "f32_vmul_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5867 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5868 | "test/f32-vmul-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5869 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5870 | ] + MICROKERNEL_TEST_HDRS, |
| 5871 | deps = MICROKERNEL_TEST_DEPS, |
| 5872 | ) |
| 5873 | |
| 5874 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5875 | name = "f32_vmul_relu_test", |
| 5876 | srcs = [ |
| 5877 | "test/f32-vmul-relu.cc", |
| 5878 | "test/vbinary-microkernel-tester.h", |
| 5879 | ] + MICROKERNEL_TEST_HDRS, |
| 5880 | deps = MICROKERNEL_TEST_DEPS, |
| 5881 | ) |
| 5882 | |
| 5883 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5884 | name = "f32_vmulc_test", |
| 5885 | srcs = [ |
| 5886 | "test/f32-vmulc.cc", |
| 5887 | "test/vbinaryc-microkernel-tester.h", |
| 5888 | ] + MICROKERNEL_TEST_HDRS, |
| 5889 | deps = MICROKERNEL_TEST_DEPS, |
| 5890 | ) |
| 5891 | |
| 5892 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5893 | name = "f32_vmulc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5894 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5895 | "test/f32-vmulc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5896 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5897 | ] + MICROKERNEL_TEST_HDRS, |
| 5898 | deps = MICROKERNEL_TEST_DEPS, |
| 5899 | ) |
| 5900 | |
| 5901 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5902 | name = "f32_vmulc_relu_test", |
| 5903 | srcs = [ |
| 5904 | "test/f32-vmulc-relu.cc", |
| 5905 | "test/vbinaryc-microkernel-tester.h", |
| 5906 | ] + MICROKERNEL_TEST_HDRS, |
| 5907 | deps = MICROKERNEL_TEST_DEPS, |
| 5908 | ) |
| 5909 | |
| 5910 | xnnpack_unit_test( |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 5911 | name = "f16_vmulcaddc_minmax_test", |
| 5912 | srcs = [ |
| 5913 | "test/f16-vmulcaddc-minmax.cc", |
| 5914 | "test/vmulcaddc-microkernel-tester.h", |
| 5915 | "src/xnnpack/AlignedAllocator.h", |
| 5916 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5917 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 5918 | ) |
| 5919 | |
| 5920 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5921 | name = "f32_vmulcaddc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5922 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5923 | "test/f32-vmulcaddc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5924 | "test/vmulcaddc-microkernel-tester.h", |
| 5925 | "src/xnnpack/AlignedAllocator.h", |
| 5926 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5927 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5928 | ) |
| 5929 | |
| 5930 | xnnpack_unit_test( |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 5931 | name = "f32_vlrelu_test", |
| 5932 | srcs = [ |
| 5933 | "test/f32-vlrelu.cc", |
| 5934 | "test/vunary-microkernel-tester.h", |
| 5935 | ] + MICROKERNEL_TEST_HDRS, |
| 5936 | deps = MICROKERNEL_TEST_DEPS, |
| 5937 | ) |
| 5938 | |
| 5939 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5940 | name = "f32_vneg_test", |
| 5941 | srcs = [ |
| 5942 | "test/f32-vneg.cc", |
| 5943 | "test/vunary-microkernel-tester.h", |
| 5944 | ] + MICROKERNEL_TEST_HDRS, |
| 5945 | deps = MICROKERNEL_TEST_DEPS, |
| 5946 | ) |
| 5947 | |
| 5948 | xnnpack_unit_test( |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 5949 | name = "f32_vrndne_test", |
| 5950 | srcs = [ |
| 5951 | "test/f32-vrndne.cc", |
| 5952 | "test/vunary-microkernel-tester.h", |
| 5953 | ] + MICROKERNEL_TEST_HDRS, |
| 5954 | deps = MICROKERNEL_TEST_DEPS, |
| 5955 | ) |
| 5956 | |
| 5957 | xnnpack_unit_test( |
| 5958 | name = "f32_vrndz_test", |
| 5959 | srcs = [ |
| 5960 | "test/f32-vrndz.cc", |
| 5961 | "test/vunary-microkernel-tester.h", |
| 5962 | ] + MICROKERNEL_TEST_HDRS, |
| 5963 | deps = MICROKERNEL_TEST_DEPS, |
| 5964 | ) |
| 5965 | |
| 5966 | xnnpack_unit_test( |
| 5967 | name = "f32_vrndu_test", |
| 5968 | srcs = [ |
| 5969 | "test/f32-vrndu.cc", |
| 5970 | "test/vunary-microkernel-tester.h", |
| 5971 | ] + MICROKERNEL_TEST_HDRS, |
| 5972 | deps = MICROKERNEL_TEST_DEPS, |
| 5973 | ) |
| 5974 | |
| 5975 | xnnpack_unit_test( |
| 5976 | name = "f32_vrndd_test", |
| 5977 | srcs = [ |
| 5978 | "test/f32-vrndd.cc", |
| 5979 | "test/vunary-microkernel-tester.h", |
| 5980 | ] + MICROKERNEL_TEST_HDRS, |
| 5981 | deps = MICROKERNEL_TEST_DEPS, |
| 5982 | ) |
| 5983 | |
| 5984 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 5985 | name = "f32_vscale_test", |
| 5986 | srcs = [ |
| 5987 | "test/f32-vscale.cc", |
| 5988 | "test/vscale-microkernel-tester.h", |
| 5989 | ] + MICROKERNEL_TEST_HDRS, |
| 5990 | deps = MICROKERNEL_TEST_DEPS, |
| 5991 | ) |
| 5992 | |
| 5993 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5994 | name = "f32_vscaleexpminusmax_test", |
| 5995 | srcs = [ |
| 5996 | "test/f32-vscaleexpminusmax.cc", |
| 5997 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 5998 | ] + MICROKERNEL_TEST_HDRS, |
| 5999 | deps = MICROKERNEL_TEST_DEPS, |
| 6000 | ) |
| 6001 | |
| 6002 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 6003 | name = "f32_vscaleextexp_test", |
| 6004 | srcs = [ |
| 6005 | "test/f32-vscaleextexp.cc", |
| 6006 | "test/vscaleextexp-microkernel-tester.h", |
| 6007 | ] + MICROKERNEL_TEST_HDRS, |
| 6008 | deps = MICROKERNEL_TEST_DEPS, |
| 6009 | ) |
| 6010 | |
| 6011 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6012 | name = "f32_vsqr_test", |
| 6013 | srcs = [ |
| 6014 | "test/f32-vsqr.cc", |
| 6015 | "test/vunary-microkernel-tester.h", |
| 6016 | ] + MICROKERNEL_TEST_HDRS, |
| 6017 | deps = MICROKERNEL_TEST_DEPS, |
| 6018 | ) |
| 6019 | |
| 6020 | xnnpack_unit_test( |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 6021 | name = "f32_vsqrdiff_test", |
| 6022 | srcs = [ |
| 6023 | "test/f32-vsqrdiff.cc", |
| 6024 | "test/vbinary-microkernel-tester.h", |
| 6025 | ] + MICROKERNEL_TEST_HDRS, |
| 6026 | deps = MICROKERNEL_TEST_DEPS, |
| 6027 | ) |
| 6028 | |
| 6029 | xnnpack_unit_test( |
| 6030 | name = "f32_vsqrdiffc_test", |
| 6031 | srcs = [ |
| 6032 | "test/f32-vsqrdiffc.cc", |
| 6033 | "test/vbinaryc-microkernel-tester.h", |
| 6034 | ] + MICROKERNEL_TEST_HDRS, |
| 6035 | deps = MICROKERNEL_TEST_DEPS, |
| 6036 | ) |
| 6037 | |
| 6038 | xnnpack_unit_test( |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 6039 | name = "f32_vsqrt_test", |
| 6040 | srcs = [ |
| 6041 | "test/f32-vsqrt.cc", |
| 6042 | "test/vunary-microkernel-tester.h", |
| 6043 | ] + MICROKERNEL_TEST_HDRS, |
| 6044 | deps = MICROKERNEL_TEST_DEPS, |
| 6045 | ) |
| 6046 | |
| 6047 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 6048 | name = "f32_vsub_test", |
| 6049 | srcs = [ |
| 6050 | "test/f32-vsub.cc", |
| 6051 | "test/vbinary-microkernel-tester.h", |
| 6052 | ] + MICROKERNEL_TEST_HDRS, |
| 6053 | deps = MICROKERNEL_TEST_DEPS, |
| 6054 | ) |
| 6055 | |
| 6056 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6057 | name = "f32_vsub_minmax_test", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 6058 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6059 | "test/f32-vsub-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 6060 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 6061 | ] + MICROKERNEL_TEST_HDRS, |
| 6062 | deps = MICROKERNEL_TEST_DEPS, |
| 6063 | ) |
| 6064 | |
| 6065 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 6066 | name = "f32_vsub_relu_test", |
| 6067 | srcs = [ |
| 6068 | "test/f32-vsub-relu.cc", |
| 6069 | "test/vbinary-microkernel-tester.h", |
| 6070 | ] + MICROKERNEL_TEST_HDRS, |
| 6071 | deps = MICROKERNEL_TEST_DEPS, |
| 6072 | ) |
| 6073 | |
| 6074 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 6075 | name = "f32_vsubc_test", |
| 6076 | srcs = [ |
| 6077 | "test/f32-vsubc.cc", |
| 6078 | "test/vbinaryc-microkernel-tester.h", |
| 6079 | ] + MICROKERNEL_TEST_HDRS, |
| 6080 | deps = MICROKERNEL_TEST_DEPS, |
| 6081 | ) |
| 6082 | |
| 6083 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6084 | name = "f32_vsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 6085 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6086 | "test/f32-vsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 6087 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 6088 | ] + MICROKERNEL_TEST_HDRS, |
| 6089 | deps = MICROKERNEL_TEST_DEPS, |
| 6090 | ) |
| 6091 | |
| 6092 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 6093 | name = "f32_vsubc_relu_test", |
| 6094 | srcs = [ |
| 6095 | "test/f32-vsubc-relu.cc", |
| 6096 | "test/vbinaryc-microkernel-tester.h", |
| 6097 | ] + MICROKERNEL_TEST_HDRS, |
| 6098 | deps = MICROKERNEL_TEST_DEPS, |
| 6099 | ) |
| 6100 | |
| 6101 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 6102 | name = "f32_vrsubc_test", |
| 6103 | srcs = [ |
| 6104 | "test/f32-vrsubc.cc", |
| 6105 | "test/vbinaryc-microkernel-tester.h", |
| 6106 | ] + MICROKERNEL_TEST_HDRS, |
| 6107 | deps = MICROKERNEL_TEST_DEPS, |
| 6108 | ) |
| 6109 | |
| 6110 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6111 | name = "f32_vrsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 6112 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6113 | "test/f32-vrsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 6114 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 6115 | ] + MICROKERNEL_TEST_HDRS, |
| 6116 | deps = MICROKERNEL_TEST_DEPS, |
| 6117 | ) |
| 6118 | |
| 6119 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 6120 | name = "f32_vrsubc_relu_test", |
| 6121 | srcs = [ |
| 6122 | "test/f32-vrsubc-relu.cc", |
| 6123 | "test/vbinaryc-microkernel-tester.h", |
| 6124 | ] + MICROKERNEL_TEST_HDRS, |
| 6125 | deps = MICROKERNEL_TEST_DEPS, |
| 6126 | ) |
| 6127 | |
| 6128 | xnnpack_unit_test( |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6129 | name = "qs8_dwconv_minmax_test", |
| 6130 | srcs = [ |
| 6131 | "test/qs8-dwconv-minmax.cc", |
| 6132 | "test/dwconv-microkernel-tester.h", |
| 6133 | "src/xnnpack/AlignedAllocator.h", |
| 6134 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 6135 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 6136 | ) |
| 6137 | |
| 6138 | xnnpack_unit_test( |
Marat Dukhan | 4ed53f4 | 2020-08-06 01:12:55 -0700 | [diff] [blame] | 6139 | name = "qs8_gavgpool_minmax_test", |
| 6140 | srcs = [ |
| 6141 | "test/qs8-gavgpool-minmax.cc", |
| 6142 | "test/gavgpool-microkernel-tester.h", |
| 6143 | "src/xnnpack/AlignedAllocator.h", |
| 6144 | ] + MICROKERNEL_TEST_HDRS, |
| 6145 | deps = MICROKERNEL_TEST_DEPS, |
| 6146 | ) |
| 6147 | |
| 6148 | xnnpack_unit_test( |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 6149 | name = "qs8_gemm_minmax_test", |
| 6150 | srcs = [ |
| 6151 | "test/qs8-gemm-minmax.cc", |
| 6152 | "test/gemm-microkernel-tester.h", |
| 6153 | "src/xnnpack/AlignedAllocator.h", |
| 6154 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 6155 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 6156 | ) |
| 6157 | |
| 6158 | xnnpack_unit_test( |
Marat Dukhan | f948068 | 2020-07-31 14:50:24 -0700 | [diff] [blame] | 6159 | name = "qs8_igemm_minmax_test", |
| 6160 | srcs = [ |
| 6161 | "test/qs8-igemm-minmax.cc", |
| 6162 | "test/gemm-microkernel-tester.h", |
| 6163 | "src/xnnpack/AlignedAllocator.h", |
| 6164 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 6165 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 6166 | ) |
| 6167 | |
| 6168 | xnnpack_unit_test( |
| 6169 | name = "qs8_requantization_test", |
| 6170 | srcs = [ |
| 6171 | "src/xnnpack/requantization-stubs.h", |
| 6172 | "test/qs8-requantization.cc", |
| 6173 | "test/requantization-tester.h", |
| 6174 | ] + MICROKERNEL_TEST_HDRS, |
| 6175 | deps = MICROKERNEL_TEST_DEPS, |
| 6176 | ) |
| 6177 | |
| 6178 | xnnpack_unit_test( |
Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 6179 | name = "qs8_vadd_minmax_test", |
| 6180 | srcs = [ |
| 6181 | "test/qs8-vadd-minmax.cc", |
| 6182 | "test/vadd-microkernel-tester.h", |
| 6183 | ] + MICROKERNEL_TEST_HDRS, |
| 6184 | deps = MICROKERNEL_TEST_DEPS, |
| 6185 | ) |
| 6186 | |
| 6187 | xnnpack_unit_test( |
Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 6188 | name = "qs8_vaddc_minmax_test", |
| 6189 | srcs = [ |
| 6190 | "test/qs8-vaddc-minmax.cc", |
| 6191 | "test/vaddc-microkernel-tester.h", |
| 6192 | ] + MICROKERNEL_TEST_HDRS, |
| 6193 | deps = MICROKERNEL_TEST_DEPS, |
| 6194 | ) |
| 6195 | |
| 6196 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6197 | name = "qu8_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6198 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6199 | "test/qu8-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6200 | "test/avgpool-microkernel-tester.h", |
| 6201 | "src/xnnpack/AlignedAllocator.h", |
| 6202 | ] + MICROKERNEL_TEST_HDRS, |
| 6203 | deps = MICROKERNEL_TEST_DEPS, |
| 6204 | ) |
| 6205 | |
| 6206 | xnnpack_unit_test( |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6207 | name = "qu8_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6208 | srcs = [ |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6209 | "test/qu8-dwconv-minmax.cc", |
| 6210 | "test/dwconv-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6211 | "src/xnnpack/AlignedAllocator.h", |
| 6212 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 6213 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6214 | ) |
| 6215 | |
| 6216 | xnnpack_unit_test( |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6217 | name = "qu8_igemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6218 | srcs = [ |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6219 | "test/qu8-igemm-minmax.cc", |
| 6220 | "test/gemm-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6221 | "src/xnnpack/AlignedAllocator.h", |
| 6222 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 6223 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6224 | ) |
| 6225 | |
| 6226 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6227 | name = "qu8_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6228 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6229 | "test/qu8-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6230 | "test/gavgpool-microkernel-tester.h", |
| 6231 | "src/xnnpack/AlignedAllocator.h", |
| 6232 | ] + MICROKERNEL_TEST_HDRS, |
| 6233 | deps = MICROKERNEL_TEST_DEPS, |
| 6234 | ) |
| 6235 | |
| 6236 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6237 | name = "qu8_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6238 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6239 | "test/qu8-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6240 | "test/gemm-microkernel-tester.h", |
| 6241 | "src/xnnpack/AlignedAllocator.h", |
| 6242 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 6243 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6244 | ) |
| 6245 | |
| 6246 | xnnpack_unit_test( |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 6247 | name = "qu8_requantization_test", |
| 6248 | srcs = [ |
| 6249 | "src/xnnpack/requantization-stubs.h", |
| 6250 | "test/qu8-requantization.cc", |
| 6251 | "test/requantization-tester.h", |
| 6252 | ] + MICROKERNEL_TEST_HDRS, |
| 6253 | deps = MICROKERNEL_TEST_DEPS, |
| 6254 | ) |
| 6255 | |
| 6256 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6257 | name = "qu8_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6258 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6259 | "test/qu8-vadd-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6260 | "test/vadd-microkernel-tester.h", |
| 6261 | ] + MICROKERNEL_TEST_HDRS, |
| 6262 | deps = MICROKERNEL_TEST_DEPS, |
| 6263 | ) |
| 6264 | |
| 6265 | xnnpack_unit_test( |
| 6266 | name = "u8_clamp_test", |
| 6267 | srcs = [ |
| 6268 | "test/u8-clamp.cc", |
| 6269 | "test/clamp-microkernel-tester.h", |
| 6270 | ] + MICROKERNEL_TEST_HDRS, |
| 6271 | deps = MICROKERNEL_TEST_DEPS, |
| 6272 | ) |
| 6273 | |
| 6274 | xnnpack_unit_test( |
| 6275 | name = "u8_lut32norm_test", |
| 6276 | srcs = [ |
| 6277 | "test/u8-lut32norm.cc", |
| 6278 | "test/lut-norm-microkernel-tester.h", |
| 6279 | ] + MICROKERNEL_TEST_HDRS, |
| 6280 | deps = MICROKERNEL_TEST_DEPS, |
| 6281 | ) |
| 6282 | |
| 6283 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 6284 | name = "u8_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6285 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 6286 | "test/u8-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6287 | "test/maxpool-microkernel-tester.h", |
| 6288 | ] + MICROKERNEL_TEST_HDRS, |
| 6289 | deps = MICROKERNEL_TEST_DEPS, |
| 6290 | ) |
| 6291 | |
| 6292 | xnnpack_unit_test( |
| 6293 | name = "u8_rmax_test", |
| 6294 | srcs = [ |
| 6295 | "test/u8-rmax.cc", |
| 6296 | "test/rmax-microkernel-tester.h", |
| 6297 | ] + MICROKERNEL_TEST_HDRS, |
| 6298 | deps = MICROKERNEL_TEST_DEPS, |
| 6299 | ) |
| 6300 | |
| 6301 | xnnpack_unit_test( |
Marat Dukhan | ad71b9a | 2020-11-20 00:01:51 -0800 | [diff] [blame] | 6302 | name = "x32_depthtospace2d_chw2hwc_test", |
Yury Kartynnik | e784186 | 2020-11-04 18:22:18 -0800 | [diff] [blame] | 6303 | srcs = [ |
Marat Dukhan | ad71b9a | 2020-11-20 00:01:51 -0800 | [diff] [blame] | 6304 | "test/x32-depthtospace2d-chw2hwc.cc", |
| 6305 | "test/depthtospace-microkernel-tester.h", |
Yury Kartynnik | e784186 | 2020-11-04 18:22:18 -0800 | [diff] [blame] | 6306 | ] + MICROKERNEL_TEST_HDRS, |
| 6307 | deps = MICROKERNEL_TEST_DEPS, |
| 6308 | ) |
| 6309 | |
| 6310 | xnnpack_unit_test( |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 6311 | name = "x32_fill_test", |
| 6312 | srcs = [ |
| 6313 | "test/x32-fill.cc", |
| 6314 | "test/fill-microkernel-tester.h", |
| 6315 | ] + MICROKERNEL_TEST_HDRS, |
| 6316 | deps = MICROKERNEL_TEST_DEPS, |
| 6317 | ) |
| 6318 | |
| 6319 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6320 | name = "x32_packx_test", |
| 6321 | srcs = [ |
| 6322 | "test/x32-packx.cc", |
| 6323 | "test/pack-microkernel-tester.h", |
| 6324 | "src/xnnpack/AlignedAllocator.h", |
| 6325 | ] + MICROKERNEL_TEST_HDRS, |
| 6326 | deps = MICROKERNEL_TEST_DEPS, |
| 6327 | ) |
| 6328 | |
| 6329 | xnnpack_unit_test( |
| 6330 | name = "x32_pad_test", |
| 6331 | srcs = [ |
| 6332 | "test/x32-pad.cc", |
| 6333 | "test/pad-microkernel-tester.h", |
| 6334 | ] + MICROKERNEL_TEST_HDRS, |
| 6335 | deps = MICROKERNEL_TEST_DEPS, |
| 6336 | ) |
| 6337 | |
| 6338 | xnnpack_unit_test( |
| 6339 | name = "x32_unpool_test", |
| 6340 | srcs = [ |
| 6341 | "test/x32-unpool.cc", |
| 6342 | "test/unpool-microkernel-tester.h", |
| 6343 | ] + MICROKERNEL_TEST_HDRS, |
| 6344 | deps = MICROKERNEL_TEST_DEPS, |
| 6345 | ) |
| 6346 | |
| 6347 | xnnpack_unit_test( |
| 6348 | name = "x32_zip_test", |
| 6349 | srcs = [ |
| 6350 | "test/x32-zip.cc", |
| 6351 | "test/zip-microkernel-tester.h", |
| 6352 | ] + MICROKERNEL_TEST_HDRS, |
| 6353 | deps = MICROKERNEL_TEST_DEPS, |
| 6354 | ) |
| 6355 | |
| 6356 | xnnpack_unit_test( |
| 6357 | name = "x8_lut_test", |
| 6358 | srcs = [ |
| 6359 | "test/x8-lut.cc", |
| 6360 | "test/lut-microkernel-tester.h", |
| 6361 | ] + MICROKERNEL_TEST_HDRS, |
| 6362 | deps = MICROKERNEL_TEST_DEPS, |
| 6363 | ) |
| 6364 | |
| 6365 | xnnpack_unit_test( |
| 6366 | name = "x8_zip_test", |
| 6367 | srcs = [ |
| 6368 | "test/x8-zip.cc", |
| 6369 | "test/zip-microkernel-tester.h", |
| 6370 | ] + MICROKERNEL_TEST_HDRS, |
| 6371 | deps = MICROKERNEL_TEST_DEPS, |
| 6372 | ) |
| 6373 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6374 | ########################## Size tests for the library ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6375 | |
| 6376 | xnnpack_binary( |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6377 | name = "operator_size_test", |
| 6378 | srcs = ["test/operator-size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 6379 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6380 | ) |
| 6381 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6382 | xnnpack_binary( |
| 6383 | name = "subgraph_size_test", |
| 6384 | srcs = ["test/subgraph-size.c"], |
| 6385 | deps = [":XNNPACK"], |
| 6386 | ) |
| 6387 | |
| 6388 | ########################### Unit tests for operators ########################## |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6389 | |
| 6390 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6391 | name = "abs_nc_test", |
| 6392 | srcs = [ |
| 6393 | "test/abs-nc.cc", |
| 6394 | "test/abs-operator-tester.h", |
| 6395 | ], |
| 6396 | deps = OPERATOR_TEST_DEPS, |
| 6397 | ) |
| 6398 | |
| 6399 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6400 | name = "add_nd_test", |
| 6401 | srcs = [ |
| 6402 | "test/add-nd.cc", |
| 6403 | "test/binary-elementwise-operator-tester.h", |
| 6404 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6405 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6406 | ) |
| 6407 | |
| 6408 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6409 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6410 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6411 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6412 | "test/argmax-pooling-operator-tester.h", |
| 6413 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6414 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6415 | ) |
| 6416 | |
| 6417 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6418 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6419 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6420 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6421 | "test/average-pooling-operator-tester.h", |
| 6422 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6423 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6424 | ) |
| 6425 | |
| 6426 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6427 | name = "bankers_rounding_nc_test", |
| 6428 | srcs = [ |
| 6429 | "test/bankers-rounding-nc.cc", |
| 6430 | "test/bankers-rounding-operator-tester.h", |
| 6431 | ], |
| 6432 | deps = OPERATOR_TEST_DEPS, |
| 6433 | ) |
| 6434 | |
| 6435 | xnnpack_unit_test( |
| 6436 | name = "ceiling_nc_test", |
| 6437 | srcs = [ |
| 6438 | "test/ceiling-nc.cc", |
| 6439 | "test/ceiling-operator-tester.h", |
| 6440 | ], |
| 6441 | deps = OPERATOR_TEST_DEPS, |
| 6442 | ) |
| 6443 | |
| 6444 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6445 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6446 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6447 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6448 | "test/channel-shuffle-operator-tester.h", |
| 6449 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6450 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6451 | ) |
| 6452 | |
| 6453 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6454 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6455 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6456 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6457 | "test/clamp-operator-tester.h", |
| 6458 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6459 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6460 | ) |
| 6461 | |
| 6462 | xnnpack_unit_test( |
Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 6463 | name = "constant_pad_nd_test", |
| 6464 | srcs = [ |
| 6465 | "test/constant-pad-nd.cc", |
| 6466 | "test/constant-pad-operator-tester.h", |
| 6467 | ], |
| 6468 | deps = OPERATOR_TEST_DEPS, |
| 6469 | ) |
| 6470 | |
| 6471 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6472 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6473 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6474 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6475 | "test/convolution-operator-tester.h", |
| 6476 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6477 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6478 | ) |
| 6479 | |
| 6480 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6481 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6482 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6483 | "test/convolution-nchw.cc", |
| 6484 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6485 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6486 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6487 | ) |
| 6488 | |
| 6489 | xnnpack_unit_test( |
Marat Dukhan | 4e21b27 | 2020-06-04 18:45:01 -0700 | [diff] [blame] | 6490 | name = "copy_nc_test", |
| 6491 | srcs = [ |
| 6492 | "test/copy-nc.cc", |
| 6493 | "test/copy-operator-tester.h", |
| 6494 | ], |
| 6495 | deps = OPERATOR_TEST_DEPS, |
| 6496 | ) |
| 6497 | |
| 6498 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6499 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6500 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6501 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6502 | "test/deconvolution-operator-tester.h", |
| 6503 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6504 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6505 | ) |
| 6506 | |
| 6507 | xnnpack_unit_test( |
Artsiom Ablavatski | 0f1dc18 | 2020-11-05 19:21:50 -0800 | [diff] [blame] | 6508 | name = "depth_to_space_test", |
| 6509 | srcs = [ |
| 6510 | "test/depth-to-space.cc", |
| 6511 | "test/depth-to-space-operator-tester.h", |
| 6512 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 6513 | deps = OPERATOR_TEST_DEPS, |
| 6514 | ) |
| 6515 | |
| 6516 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 6517 | name = "divide_nd_test", |
| 6518 | srcs = [ |
| 6519 | "test/binary-elementwise-operator-tester.h", |
| 6520 | "test/divide-nd.cc", |
| 6521 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6522 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 6523 | ) |
| 6524 | |
| 6525 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6526 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6527 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6528 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6529 | "test/fully-connected-operator-tester.h", |
| 6530 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6531 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6532 | ) |
| 6533 | |
| 6534 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6535 | name = "floor_nc_test", |
| 6536 | srcs = [ |
| 6537 | "test/floor-nc.cc", |
| 6538 | "test/floor-operator-tester.h", |
| 6539 | ], |
| 6540 | deps = OPERATOR_TEST_DEPS, |
| 6541 | ) |
| 6542 | |
| 6543 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6544 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6545 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6546 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6547 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | ef61d02 | 2020-06-19 13:54:49 -0700 | [diff] [blame] | 6548 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6549 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6550 | ) |
| 6551 | |
| 6552 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6553 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6554 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6555 | "test/global-average-pooling-ncw.cc", |
| 6556 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6557 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6558 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6559 | ) |
| 6560 | |
| 6561 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6562 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6563 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6564 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6565 | "test/hardswish-operator-tester.h", |
| 6566 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6567 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6568 | ) |
| 6569 | |
| 6570 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6571 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6572 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6573 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6574 | "test/leaky-relu-operator-tester.h", |
| 6575 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6576 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6577 | ) |
| 6578 | |
| 6579 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6580 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6581 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6582 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6583 | "test/max-pooling-operator-tester.h", |
| 6584 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6585 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6586 | ) |
| 6587 | |
| 6588 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6589 | name = "maximum_nd_test", |
| 6590 | srcs = [ |
| 6591 | "test/binary-elementwise-operator-tester.h", |
| 6592 | "test/maximum-nd.cc", |
| 6593 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6594 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6595 | ) |
| 6596 | |
| 6597 | xnnpack_unit_test( |
| 6598 | name = "minimum_nd_test", |
| 6599 | srcs = [ |
| 6600 | "test/binary-elementwise-operator-tester.h", |
| 6601 | "test/minimum-nd.cc", |
| 6602 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6603 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6604 | ) |
| 6605 | |
| 6606 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6607 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6608 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6609 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6610 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6611 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6612 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6613 | ) |
| 6614 | |
| 6615 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6616 | name = "negate_nc_test", |
| 6617 | srcs = [ |
| 6618 | "test/negate-nc.cc", |
| 6619 | "test/negate-operator-tester.h", |
| 6620 | ], |
| 6621 | deps = OPERATOR_TEST_DEPS, |
| 6622 | ) |
| 6623 | |
| 6624 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6625 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6626 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6627 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6628 | "test/prelu-operator-tester.h", |
| 6629 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6630 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6631 | ) |
| 6632 | |
| 6633 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6634 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6635 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6636 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6637 | "test/resize-bilinear-operator-tester.h", |
| 6638 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6639 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6640 | ) |
| 6641 | |
| 6642 | xnnpack_unit_test( |
Artsiom Ablavatski | 9791810 | 2020-10-27 15:52:59 -0700 | [diff] [blame] | 6643 | name = "resize_bilinear_nchw_test", |
| 6644 | srcs = [ |
| 6645 | "test/resize-bilinear-nchw.cc", |
| 6646 | "test/resize-bilinear-operator-tester.h", |
| 6647 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| 6648 | deps = OPERATOR_TEST_DEPS, |
| 6649 | ) |
| 6650 | |
| 6651 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6652 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6653 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6654 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6655 | "test/sigmoid-operator-tester.h", |
| 6656 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6657 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6658 | ) |
| 6659 | |
| 6660 | xnnpack_unit_test( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 6661 | name = "softmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6662 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 6663 | "test/softmax-nc.cc", |
| 6664 | "test/softmax-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6665 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6666 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6667 | ) |
| 6668 | |
| 6669 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6670 | name = "square_nc_test", |
| 6671 | srcs = [ |
| 6672 | "test/square-nc.cc", |
| 6673 | "test/square-operator-tester.h", |
| 6674 | ], |
| 6675 | deps = OPERATOR_TEST_DEPS, |
| 6676 | ) |
| 6677 | |
| 6678 | xnnpack_unit_test( |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 6679 | name = "square_root_nc_test", |
| 6680 | srcs = [ |
| 6681 | "test/square-root-nc.cc", |
| 6682 | "test/square-root-operator-tester.h", |
| 6683 | ], |
| 6684 | deps = OPERATOR_TEST_DEPS, |
| 6685 | ) |
| 6686 | |
| 6687 | xnnpack_unit_test( |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 6688 | name = "squared_difference_nd_test", |
| 6689 | srcs = [ |
| 6690 | "test/binary-elementwise-operator-tester.h", |
| 6691 | "test/squared-difference-nd.cc", |
| 6692 | ], |
| 6693 | deps = OPERATOR_TEST_DEPS, |
| 6694 | ) |
| 6695 | |
| 6696 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 6697 | name = "subtract_nd_test", |
| 6698 | srcs = [ |
| 6699 | "test/binary-elementwise-operator-tester.h", |
| 6700 | "test/subtract-nd.cc", |
| 6701 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6702 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 6703 | ) |
| 6704 | |
| 6705 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6706 | name = "truncation_nc_test", |
| 6707 | srcs = [ |
| 6708 | "test/truncation-nc.cc", |
| 6709 | "test/truncation-operator-tester.h", |
| 6710 | ], |
| 6711 | deps = OPERATOR_TEST_DEPS, |
| 6712 | ) |
| 6713 | |
| 6714 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6715 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6716 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6717 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6718 | "test/unpooling-operator-tester.h", |
| 6719 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6720 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6721 | ) |
| 6722 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 6723 | ############################### Misc unit tests ############################### |
| 6724 | |
| 6725 | xnnpack_unit_test( |
| 6726 | name = "memory_planner_test", |
| 6727 | srcs = [ |
| 6728 | "test/memory-planner-test.cc", |
| 6729 | ], |
| 6730 | deps = [ |
| 6731 | ":XNNPACK", |
| 6732 | ":memory_planner", |
| 6733 | ], |
| 6734 | ) |
| 6735 | |
XNNPACK Team | ab8c4c8 | 2020-10-09 08:05:51 -0700 | [diff] [blame] | 6736 | xnnpack_unit_test( |
| 6737 | name = "subgraph_nchw_test", |
| 6738 | srcs = [ |
| 6739 | "src/xnnpack/subgraph.h", |
| 6740 | "test/subgraph-nchw.cc", |
| 6741 | "test/subgraph-tester.h", |
| 6742 | ], |
| 6743 | deps = [ |
| 6744 | ":XNNPACK", |
| 6745 | ], |
| 6746 | ) |
| 6747 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6748 | ############################# Build configurations ############################# |
| 6749 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6750 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6751 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6752 | name = "xnn_enable_assembly_explicit_true", |
| 6753 | define_values = {"xnn_enable_assembly": "true"}, |
| 6754 | ) |
| 6755 | |
| 6756 | # Disables usage of assembly kernels. |
| 6757 | config_setting( |
| 6758 | name = "xnn_enable_assembly_explicit_false", |
| 6759 | define_values = {"xnn_enable_assembly": "false"}, |
| 6760 | ) |
| 6761 | |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 6762 | # Enables usage of sparse inference. |
| 6763 | config_setting( |
| 6764 | name = "xnn_enable_sparse_explicit_true", |
| 6765 | define_values = {"xnn_enable_sparse": "true"}, |
| 6766 | ) |
| 6767 | |
| 6768 | # Disables usage of sparse inference. |
| 6769 | config_setting( |
| 6770 | name = "xnn_enable_sparse_explicit_false", |
| 6771 | define_values = {"xnn_enable_sparse": "false"}, |
| 6772 | ) |
| 6773 | |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 6774 | # Disables usage of HMP-aware optimizations. |
| 6775 | config_setting( |
| 6776 | name = "xnn_enable_hmp_explicit_false", |
| 6777 | define_values = {"xnn_enable_hmp": "false"}, |
| 6778 | ) |
| 6779 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 6780 | # Enable usage of optimized memory allocation |
| 6781 | config_setting( |
| 6782 | name = "xnn_enable_memopt_explicit_true", |
| 6783 | define_values = {"xnnpack_enable_memopt": "true"}, |
| 6784 | ) |
| 6785 | |
| 6786 | # Disable usage of optimized memory allocation |
| 6787 | config_setting( |
| 6788 | name = "xnn_enable_memopt_explicit_false", |
| 6789 | define_values = {"xnnpack_enable_memopt": "false"}, |
| 6790 | ) |
| 6791 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6792 | # Builds with -c dbg |
| 6793 | config_setting( |
| 6794 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6795 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6796 | "compilation_mode": "dbg", |
| 6797 | }, |
| 6798 | ) |
| 6799 | |
| 6800 | # Builds with -c opt |
| 6801 | config_setting( |
| 6802 | name = "optimized_build", |
| 6803 | values = { |
| 6804 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6805 | }, |
| 6806 | ) |
| 6807 | |
| 6808 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6809 | name = "linux_k8", |
| 6810 | values = {"cpu": "k8"}, |
| 6811 | ) |
| 6812 | |
| 6813 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 6814 | name = "linux_arm", |
| 6815 | values = {"cpu": "arm"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 6816 | ) |
| 6817 | |
| 6818 | config_setting( |
Marat Dukhan | f0bd4de | 2020-06-15 15:53:19 -0700 | [diff] [blame] | 6819 | name = "linux_armeabi", |
| 6820 | values = {"cpu": "armeabi"}, |
| 6821 | ) |
| 6822 | |
| 6823 | config_setting( |
Terry Heo | 68eef3f | 2020-04-13 22:53:52 -0700 | [diff] [blame] | 6824 | name = "linux_armhf", |
| 6825 | values = {"cpu": "armhf"}, |
| 6826 | ) |
| 6827 | |
| 6828 | config_setting( |
Marat Dukhan | a720e93 | 2020-06-10 13:01:11 -0700 | [diff] [blame] | 6829 | name = "linux_armv7a", |
| 6830 | values = {"cpu": "armv7a"}, |
| 6831 | ) |
| 6832 | |
| 6833 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 6834 | name = "linux_aarch64", |
| 6835 | values = {"cpu": "aarch64"}, |
| 6836 | ) |
| 6837 | |
| 6838 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6839 | name = "android", |
| 6840 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 6841 | ) |
| 6842 | |
| 6843 | config_setting( |
| 6844 | name = "android_armv7", |
| 6845 | values = { |
| 6846 | "crosstool_top": "//external:android/crosstool", |
| 6847 | "cpu": "armeabi-v7a", |
| 6848 | }, |
| 6849 | ) |
| 6850 | |
| 6851 | config_setting( |
| 6852 | name = "android_arm64", |
| 6853 | values = { |
| 6854 | "crosstool_top": "//external:android/crosstool", |
| 6855 | "cpu": "arm64-v8a", |
| 6856 | }, |
| 6857 | ) |
| 6858 | |
| 6859 | config_setting( |
| 6860 | name = "android_x86", |
| 6861 | values = { |
| 6862 | "crosstool_top": "//external:android/crosstool", |
| 6863 | "cpu": "x86", |
| 6864 | }, |
| 6865 | ) |
| 6866 | |
| 6867 | config_setting( |
| 6868 | name = "android_x86_64", |
| 6869 | values = { |
| 6870 | "crosstool_top": "//external:android/crosstool", |
| 6871 | "cpu": "x86_64", |
| 6872 | }, |
| 6873 | ) |
| 6874 | |
| 6875 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 6876 | name = "windows_x86_64", |
| 6877 | values = {"cpu": "x64_windows"}, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 6878 | ) |
| 6879 | |
| 6880 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 6881 | name = "windows_x86_64_clang", |
| 6882 | values = { |
| 6883 | "compiler": "clang-cl", |
| 6884 | "cpu": "x64_windows", |
| 6885 | }, |
| 6886 | ) |
| 6887 | |
| 6888 | config_setting( |
| 6889 | name = "windows_x86_64_mingw", |
| 6890 | values = { |
| 6891 | "compiler": "mingw-gcc", |
| 6892 | "cpu": "x64_windows", |
| 6893 | }, |
| 6894 | ) |
| 6895 | |
| 6896 | config_setting( |
| 6897 | name = "windows_x86_64_msys", |
| 6898 | values = { |
| 6899 | "compiler": "msys-gcc", |
| 6900 | "cpu": "x64_windows", |
| 6901 | }, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 6902 | ) |
| 6903 | |
| 6904 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 6905 | name = "macos_x86_64", |
| 6906 | values = { |
| 6907 | "apple_platform_type": "macos", |
| 6908 | "cpu": "darwin", |
| 6909 | }, |
| 6910 | ) |
| 6911 | |
| 6912 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6913 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6914 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6915 | ) |
| 6916 | |
| 6917 | config_setting( |
| 6918 | name = "emscripten_wasm", |
| 6919 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6920 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6921 | "cpu": "wasm", |
| 6922 | }, |
| 6923 | ) |
| 6924 | |
| 6925 | config_setting( |
| 6926 | name = "emscripten_wasmsimd", |
| 6927 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6928 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6929 | "cpu": "wasm", |
Marat Dukhan | 81c6260 | 2020-05-29 13:22:49 -0700 | [diff] [blame] | 6930 | "copt": "-msimd128", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6931 | }, |
| 6932 | ) |
| 6933 | |
| 6934 | config_setting( |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6935 | name = "ios_armv7", |
| 6936 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6937 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6938 | "cpu": "ios_armv7", |
| 6939 | }, |
| 6940 | ) |
| 6941 | |
| 6942 | config_setting( |
| 6943 | name = "ios_arm64", |
| 6944 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6945 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6946 | "cpu": "ios_arm64", |
| 6947 | }, |
| 6948 | ) |
| 6949 | |
| 6950 | config_setting( |
| 6951 | name = "ios_arm64e", |
| 6952 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6953 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6954 | "cpu": "ios_arm64e", |
| 6955 | }, |
| 6956 | ) |
| 6957 | |
| 6958 | config_setting( |
| 6959 | name = "ios_x86", |
| 6960 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6961 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6962 | "cpu": "ios_i386", |
| 6963 | }, |
| 6964 | ) |
| 6965 | |
| 6966 | config_setting( |
| 6967 | name = "ios_x86_64", |
| 6968 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6969 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6970 | "cpu": "ios_x86_64", |
| 6971 | }, |
| 6972 | ) |
| 6973 | |
| 6974 | config_setting( |
| 6975 | name = "watchos_armv7k", |
| 6976 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6977 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6978 | "cpu": "watchos_armv7k", |
| 6979 | }, |
| 6980 | ) |
| 6981 | |
| 6982 | config_setting( |
| 6983 | name = "watchos_arm64_32", |
| 6984 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6985 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6986 | "cpu": "watchos_arm64_32", |
| 6987 | }, |
| 6988 | ) |
| 6989 | |
| 6990 | config_setting( |
| 6991 | name = "watchos_x86", |
| 6992 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6993 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6994 | "cpu": "watchos_i386", |
| 6995 | }, |
| 6996 | ) |
| 6997 | |
| 6998 | config_setting( |
| 6999 | name = "watchos_x86_64", |
| 7000 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 7001 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 7002 | "cpu": "watchos_x86_64", |
| 7003 | }, |
| 7004 | ) |
| 7005 | |
| 7006 | config_setting( |
| 7007 | name = "tvos_arm64", |
| 7008 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 7009 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 7010 | "cpu": "tvos_arm64", |
| 7011 | }, |
| 7012 | ) |
| 7013 | |
| 7014 | config_setting( |
| 7015 | name = "tvos_x86_64", |
| 7016 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 7017 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 7018 | "cpu": "tvos_x86_64", |
| 7019 | }, |
| 7020 | ) |