| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1 | # Copyright 2019 Google LLC |
| 2 | # |
| 3 | # This source code is licensed under the BSD-style license found in the |
| 4 | # LICENSE file in the root directory of this source tree. |
| 5 | # |
| 6 | # Description: |
| 7 | # XNNPACK - optimized floating-point neural network operators library |
| 8 | |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 9 | load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility") |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 10 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 11 | licenses(["notice"]) |
| 12 | |
| 13 | exports_files(["LICENSE"]) |
| 14 | |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 15 | OPERATOR_BENCHMARK_DEPS = [ |
| 16 | ":XNNPACK", |
| 17 | ":bench_utils", |
| 18 | "@cpuinfo", |
| Frank Barchard | 0c84973 | 2020-06-12 13:31:32 -0700 | [diff] [blame] | 19 | "@FP16", |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 20 | "@pthreadpool", |
| 21 | ] |
| 22 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 23 | MICROKERNEL_BENCHMARK_DEPS = [ |
| 24 | ":ukernels", |
| 25 | ":bench_utils", |
| Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 26 | ":enable_assembly", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 27 | "@cpuinfo", |
| 28 | "@FP16", |
| 29 | "@pthreadpool", |
| 30 | ] |
| 31 | |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 32 | ACCURACY_EVAL_DEPS = [ |
| 33 | ":XNNPACK", |
| 34 | ":ukernels", |
| 35 | "@FP16", |
| 36 | "@pthreadpool", |
| 37 | ] |
| 38 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 39 | MICROKERNEL_TEST_DEPS = [ |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 40 | ":ukernels_test_mode", |
| Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 41 | ":enable_assembly", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 42 | "@cpuinfo", |
| 43 | "@FP16", |
| 44 | "@pthreadpool", |
| 45 | ] |
| 46 | |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 47 | OPERATOR_TEST_DEPS = [ |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 48 | ":XNNPACK_test_mode", |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 49 | "@pthreadpool", |
| 50 | "@FP16", |
| 51 | ] |
| 52 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 53 | OPERATOR_SRCS = [ |
| Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 54 | "src/operators/argmax-pooling-nhwc.c", |
| 55 | "src/operators/average-pooling-nhwc.c", |
| 56 | "src/operators/binary-elementwise-nd.c", |
| Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 57 | "src/operators/channel-shuffle-nc.c", |
| Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 58 | "src/operators/constant-pad-nd.c", |
| Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 59 | "src/operators/convolution-nchw.c", |
| 60 | "src/operators/convolution-nhwc.c", |
| 61 | "src/operators/deconvolution-nhwc.c", |
| 62 | "src/operators/fully-connected-nc.c", |
| 63 | "src/operators/global-average-pooling-ncw.c", |
| 64 | "src/operators/global-average-pooling-nwc.c", |
| Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 65 | "src/operators/leaky-relu-nc.c", |
| 66 | "src/operators/max-pooling-nhwc.c", |
| 67 | "src/operators/prelu-nc.c", |
| 68 | "src/operators/resize-bilinear-nhwc.c", |
| 69 | "src/operators/sigmoid-nc.c", |
| 70 | "src/operators/softmax-nc.c", |
| Marat Dukhan | c3065f5 | 2020-06-04 13:33:32 -0700 | [diff] [blame] | 71 | "src/operators/unary-elementwise-nc.c", |
| Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 72 | "src/operators/unpooling-nhwc.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 73 | ] |
| 74 | |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 75 | SUBGRAPH_SRCS = [ |
| Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 76 | "src/subgraph/abs.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 77 | "src/subgraph/add2.c", |
| 78 | "src/subgraph/argmax-pooling-2d.c", |
| 79 | "src/subgraph/average-pooling-2d.c", |
| Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 80 | "src/subgraph/bankers-rounding.c", |
| 81 | "src/subgraph/ceiling.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 82 | "src/subgraph/clamp.c", |
| 83 | "src/subgraph/convolution-2d.c", |
| 84 | "src/subgraph/deconvolution-2d.c", |
| 85 | "src/subgraph/depthwise-convolution-2d.c", |
| Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 86 | "src/subgraph/divide.c", |
| Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 87 | "src/subgraph/floor.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 88 | "src/subgraph/fully-connected.c", |
| Marat Dukhan | a059b7d | 2020-06-11 11:41:27 -0700 | [diff] [blame] | 89 | "src/subgraph/global-average-pooling-2d.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 90 | "src/subgraph/hardswish.c", |
| Marat Dukhan | 5bbebac | 2020-06-10 19:42:15 -0700 | [diff] [blame] | 91 | "src/subgraph/leaky-relu.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 92 | "src/subgraph/max-pooling-2d.c", |
| Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 93 | "src/subgraph/maximum2.c", |
| 94 | "src/subgraph/minimum2.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 95 | "src/subgraph/multiply2.c", |
| Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 96 | "src/subgraph/negate.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 97 | "src/subgraph/prelu.c", |
| 98 | "src/subgraph/sigmoid.c", |
| 99 | "src/subgraph/softmax.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 100 | "src/subgraph/square-root.c", |
| 101 | "src/subgraph/square.c", |
| 102 | "src/subgraph/squared-difference.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 103 | "src/subgraph/static-constant-pad.c", |
| Marat Dukhan | d27202d | 2020-07-09 23:43:40 -0700 | [diff] [blame] | 104 | "src/subgraph/static-reshape.c", |
| Marat Dukhan | aff24e2 | 2020-07-23 01:43:58 -0700 | [diff] [blame] | 105 | "src/subgraph/static-resize-bilinear-2d.c", |
| Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 106 | "src/subgraph/subtract.c", |
| Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 107 | "src/subgraph/unpooling-2d.c", |
| 108 | ] |
| 109 | |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 110 | TABLE_SRCS = [ |
| 111 | "src/tables/exp2-k-over-64.c", |
| 112 | "src/tables/exp2-k-over-2048.c", |
| Marat Dukhan | 1f256fc | 2020-09-24 21:27:14 -0700 | [diff] [blame] | 113 | "src/tables/exp2minus-k-over-64.c", |
| 114 | "src/tables/exp2minus-k-over-2048.c", |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 115 | ] |
| 116 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 117 | SCALAR_UKERNELS = [ |
| Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 118 | "src/f32-argmaxpool/4x-scalar-c1.c", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 119 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
| Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 120 | "src/f32-argmaxpool/9x-scalar-c1.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 121 | "src/f32-avgpool/9p8x-minmax-scalar-c1.c", |
| 122 | "src/f32-avgpool/9x-minmax-scalar-c1.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 123 | "src/f32-clamp/gen/scalar-x1.c", |
| 124 | "src/f32-clamp/gen/scalar-x2.c", |
| 125 | "src/f32-clamp/gen/scalar-x4.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 126 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 127 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 128 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 129 | "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 130 | "src/f32-dwconv/gen/up1x4-minmax-scalar.c", |
| 131 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 132 | "src/f32-dwconv/gen/up1x4-scalar.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 133 | "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 134 | "src/f32-dwconv/gen/up1x9-minmax-scalar.c", |
| 135 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 136 | "src/f32-dwconv/gen/up1x9-scalar.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 137 | "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 138 | "src/f32-dwconv/gen/up1x25-minmax-scalar.c", |
| 139 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 140 | "src/f32-dwconv/gen/up1x25-scalar.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 141 | "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 142 | "src/f32-dwconv/gen/up2x4-minmax-scalar.c", |
| 143 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 144 | "src/f32-dwconv/gen/up2x4-scalar.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 145 | "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 146 | "src/f32-dwconv/gen/up2x9-minmax-scalar.c", |
| 147 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 148 | "src/f32-dwconv/gen/up2x9-scalar.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 149 | "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 150 | "src/f32-dwconv/gen/up2x25-minmax-scalar.c", |
| 151 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 152 | "src/f32-dwconv/gen/up2x25-scalar.c", |
| Marat Dukhan | 91249d2 | 2020-10-24 12:02:51 -0700 | [diff] [blame] | 153 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c", |
| 154 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c", |
| 155 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c", |
| 156 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c", |
| 157 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c", |
| 158 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c", |
| 159 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c", |
| 160 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c", |
| 161 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c", |
| 162 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c", |
| Marat Dukhan | cf5b3c3 | 2020-10-25 19:21:10 -0700 | [diff] [blame] | 163 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c", |
| 164 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c", |
| 165 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c", |
| 166 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c", |
| 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", |
| 170 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c", |
| Marat Dukhan | c4efb00 | 2020-10-25 23:14:47 -0700 | [diff] [blame^] | 171 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c", |
| 172 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c", |
| 173 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c", |
| 174 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c", |
| 175 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c", |
| 176 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c", |
| 177 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c", |
| 178 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c", |
| 179 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c", |
| 180 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c", |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 181 | "src/f32-dwconv2d-chw/5x5s2p2-scalar-1x1-acc5.c", |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 182 | "src/f32-gavgpool-cw/scalar-x1.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 183 | "src/f32-gavgpool/7p7x-minmax-scalar-c1.c", |
| 184 | "src/f32-gavgpool/7x-minmax-scalar-c1.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 185 | "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c", |
| 186 | "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c", |
| 187 | "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 188 | "src/f32-gemm/gen/1x4-minmax-scalar.c", |
| 189 | "src/f32-gemm/gen/1x4-relu-scalar.c", |
| 190 | "src/f32-gemm/gen/1x4-scalar.c", |
| 191 | "src/f32-gemm/gen/2x4-minmax-scalar.c", |
| 192 | "src/f32-gemm/gen/2x4-relu-scalar.c", |
| 193 | "src/f32-gemm/gen/2x4-scalar.c", |
| 194 | "src/f32-gemm/gen/4x2-minmax-scalar.c", |
| 195 | "src/f32-gemm/gen/4x2-relu-scalar.c", |
| 196 | "src/f32-gemm/gen/4x2-scalar.c", |
| 197 | "src/f32-gemm/gen/4x4-minmax-scalar.c", |
| 198 | "src/f32-gemm/gen/4x4-relu-scalar.c", |
| 199 | "src/f32-gemm/gen/4x4-scalar.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 200 | "src/f32-hswish/gen/hswish-scalar-x1.c", |
| 201 | "src/f32-hswish/gen/hswish-scalar-x2.c", |
| 202 | "src/f32-hswish/gen/hswish-scalar-x4.c", |
| XNNPACK Team | 2143267 | 2020-10-19 19:58:48 -0700 | [diff] [blame] | 203 | "src/f32-ibilinear-chw/gen/scalar-p1.c", |
| 204 | "src/f32-ibilinear-chw/gen/scalar-p2.c", |
| 205 | "src/f32-ibilinear-chw/gen/scalar-p4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 206 | "src/f32-ibilinear/gen/scalar-c1.c", |
| 207 | "src/f32-ibilinear/gen/scalar-c2.c", |
| 208 | "src/f32-ibilinear/gen/scalar-c4.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 209 | "src/f32-igemm/gen/1x4-minmax-scalar.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 210 | "src/f32-igemm/gen/1x4-relu-scalar.c", |
| 211 | "src/f32-igemm/gen/1x4-scalar.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 212 | "src/f32-igemm/gen/2x4-minmax-scalar.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 213 | "src/f32-igemm/gen/2x4-relu-scalar.c", |
| 214 | "src/f32-igemm/gen/2x4-scalar.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 215 | "src/f32-igemm/gen/4x2-minmax-scalar.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 216 | "src/f32-igemm/gen/4x2-relu-scalar.c", |
| 217 | "src/f32-igemm/gen/4x2-scalar.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 218 | "src/f32-igemm/gen/4x4-minmax-scalar.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 219 | "src/f32-igemm/gen/4x4-relu-scalar.c", |
| 220 | "src/f32-igemm/gen/4x4-scalar.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 221 | "src/f32-maxpool/9p8x-minmax-scalar-c1.c", |
| 222 | "src/f32-pavgpool/9p8x-minmax-scalar-c1.c", |
| 223 | "src/f32-pavgpool/9x-minmax-scalar-c1.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 224 | "src/f32-ppmm/gen/2x4-minmax-scalar.c", |
| 225 | "src/f32-ppmm/gen/3x3-minmax-scalar.c", |
| 226 | "src/f32-ppmm/gen/4x2-minmax-scalar.c", |
| 227 | "src/f32-ppmm/gen/4x4-minmax-scalar.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 228 | "src/f32-prelu/gen/scalar-2x1.c", |
| 229 | "src/f32-prelu/gen/scalar-2x4.c", |
| Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 230 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
| Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 231 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 232 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
| Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 233 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 234 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 235 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
| Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 236 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
| Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 237 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 238 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
| Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 239 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 240 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 241 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 242 | "src/f32-relu/gen/scalar-x1.c", |
| 243 | "src/f32-relu/gen/scalar-x2.c", |
| 244 | "src/f32-relu/gen/scalar-x4.c", |
| Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 245 | "src/f32-relu/gen/scalar-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 246 | "src/f32-rmax/scalar.c", |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 247 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 248 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 249 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 250 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 251 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 252 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 253 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 254 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 255 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 256 | "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c", |
| 257 | "src/f32-spmm/gen/1x1-minmax-scalar.c", |
| 258 | "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c", |
| 259 | "src/f32-spmm/gen/2x1-minmax-scalar.c", |
| 260 | "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c", |
| 261 | "src/f32-spmm/gen/4x1-minmax-scalar.c", |
| 262 | "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c", |
| 263 | "src/f32-spmm/gen/8x1-minmax-scalar.c", |
| 264 | "src/f32-spmm/gen/8x2-minmax-scalar.c", |
| 265 | "src/f32-spmm/gen/8x4-minmax-scalar.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 266 | "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c", |
| 267 | "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c", |
| 268 | "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 269 | "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 270 | "src/f32-vbinary/gen/vadd-relu-scalar-x1.c", |
| 271 | "src/f32-vbinary/gen/vadd-relu-scalar-x2.c", |
| 272 | "src/f32-vbinary/gen/vadd-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 273 | "src/f32-vbinary/gen/vadd-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 274 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 275 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 276 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 277 | "src/f32-vbinary/gen/vadd-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 278 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c", |
| 279 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c", |
| 280 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 281 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 282 | "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c", |
| 283 | "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c", |
| 284 | "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 285 | "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 286 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 287 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 288 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 289 | "src/f32-vbinary/gen/vaddc-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 290 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c", |
| 291 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c", |
| 292 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 293 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 294 | "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c", |
| 295 | "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c", |
| 296 | "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 297 | "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 298 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 299 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 300 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 301 | "src/f32-vbinary/gen/vdiv-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 302 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c", |
| 303 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c", |
| 304 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 305 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 306 | "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c", |
| 307 | "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c", |
| 308 | "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 309 | "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 310 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 311 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 312 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 313 | "src/f32-vbinary/gen/vdivc-scalar-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 314 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 315 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 316 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 317 | "src/f32-vbinary/gen/vmax-scalar-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 318 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 319 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 320 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 321 | "src/f32-vbinary/gen/vmaxc-scalar-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 322 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 323 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 324 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 325 | "src/f32-vbinary/gen/vmin-scalar-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 326 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 327 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 328 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 329 | "src/f32-vbinary/gen/vminc-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 330 | "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c", |
| 331 | "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c", |
| 332 | "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 333 | "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 334 | "src/f32-vbinary/gen/vmul-relu-scalar-x1.c", |
| 335 | "src/f32-vbinary/gen/vmul-relu-scalar-x2.c", |
| 336 | "src/f32-vbinary/gen/vmul-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 337 | "src/f32-vbinary/gen/vmul-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 338 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 339 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 340 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 341 | "src/f32-vbinary/gen/vmul-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 342 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c", |
| 343 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c", |
| 344 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 345 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 346 | "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c", |
| 347 | "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c", |
| 348 | "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 349 | "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 350 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 351 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 352 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 353 | "src/f32-vbinary/gen/vmulc-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 354 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c", |
| 355 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c", |
| 356 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 357 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 358 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c", |
| 359 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c", |
| 360 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 361 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 362 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 363 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 364 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 365 | "src/f32-vbinary/gen/vrdivc-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 366 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c", |
| 367 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c", |
| 368 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 369 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 370 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c", |
| 371 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c", |
| 372 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 373 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 374 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 375 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 376 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 377 | "src/f32-vbinary/gen/vrsubc-scalar-x8.c", |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 378 | "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c", |
| 379 | "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c", |
| 380 | "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 381 | "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c", |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 382 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c", |
| 383 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c", |
| 384 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 385 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 386 | "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c", |
| 387 | "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c", |
| 388 | "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 389 | "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 390 | "src/f32-vbinary/gen/vsub-relu-scalar-x1.c", |
| 391 | "src/f32-vbinary/gen/vsub-relu-scalar-x2.c", |
| 392 | "src/f32-vbinary/gen/vsub-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 393 | "src/f32-vbinary/gen/vsub-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 394 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 395 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 396 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 397 | "src/f32-vbinary/gen/vsub-scalar-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 398 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c", |
| 399 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c", |
| 400 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 401 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 402 | "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c", |
| 403 | "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c", |
| 404 | "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 405 | "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c", |
| Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 406 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 407 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 408 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 409 | "src/f32-vbinary/gen/vsubc-scalar-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 410 | "src/f32-vlrelu/gen/vlrelu-scalar-x1.c", |
| 411 | "src/f32-vlrelu/gen/vlrelu-scalar-x2.c", |
| 412 | "src/f32-vlrelu/gen/vlrelu-scalar-x4.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 413 | "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c", |
| 414 | "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c", |
| 415 | "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c", |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 416 | "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c", |
| 417 | "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c", |
| 418 | "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 419 | "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c", |
| 420 | "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c", |
| 421 | "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c", |
| 422 | "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c", |
| 423 | "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c", |
| 424 | "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c", |
| 425 | "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c", |
| 426 | "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c", |
| 427 | "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 428 | "src/f32-vsqrt/gen/scalar-sqrt-x1.c", |
| 429 | "src/f32-vsqrt/gen/scalar-sqrt-x2.c", |
| 430 | "src/f32-vsqrt/gen/scalar-sqrt-x4.c", |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 431 | "src/f32-vunary/gen/vabs-scalar-x1.c", |
| 432 | "src/f32-vunary/gen/vabs-scalar-x2.c", |
| 433 | "src/f32-vunary/gen/vabs-scalar-x4.c", |
| 434 | "src/f32-vunary/gen/vneg-scalar-x1.c", |
| 435 | "src/f32-vunary/gen/vneg-scalar-x2.c", |
| 436 | "src/f32-vunary/gen/vneg-scalar-x4.c", |
| 437 | "src/f32-vunary/gen/vsqr-scalar-x1.c", |
| 438 | "src/f32-vunary/gen/vsqr-scalar-x2.c", |
| 439 | "src/f32-vunary/gen/vsqr-scalar-x4.c", |
| Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 440 | "src/math/expminus-scalar-lut64-p2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 441 | "src/math/expminus-scalar-lut2048-p1.c", |
| Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 442 | "src/math/expminus-scalar-p5.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 443 | "src/math/roundd-scalar-addsub.c", |
| 444 | "src/math/roundd-scalar-cvt.c", |
| Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 445 | "src/math/roundd-scalar-floor.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 446 | "src/math/roundne-scalar-addsub.c", |
| 447 | "src/math/roundne-scalar-nearbyint.c", |
| 448 | "src/math/roundne-scalar-rint.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 449 | "src/math/roundu-scalar-addsub.c", |
| Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 450 | "src/math/roundu-scalar-ceil.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 451 | "src/math/roundu-scalar-cvt.c", |
| Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 452 | "src/math/roundz-scalar-addsub.c", |
| 453 | "src/math/roundz-scalar-cvt.c", |
| Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 454 | "src/math/roundz-scalar-trunc.c", |
| Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 455 | "src/math/sigmoid-scalar-rr2-lut64-p2-div.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 456 | "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c", |
| Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 457 | "src/math/sigmoid-scalar-rr2-p5-div.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 458 | "src/qs8-gemm/gen/8x8c4-minmax-scalar.c", |
| 459 | "src/qs8-gemm/gen/12x4c4-minmax-scalar.c", |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 460 | "src/qs8-requantization/fp32-scalar-lrintf.c", |
| 461 | "src/qs8-requantization/fp32-scalar-magic.c", |
| 462 | "src/qs8-requantization/precise-scalar-signed64.c", |
| 463 | "src/qs8-requantization/precise-scalar-unsigned32.c", |
| 464 | "src/qs8-requantization/precise-scalar-unsigned64.c", |
| 465 | "src/qs8-requantization/q31-scalar.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 466 | "src/qu8-avgpool/9p8x-minmax-scalar-c1.c", |
| 467 | "src/qu8-avgpool/9x-minmax-scalar-c1.c", |
| 468 | "src/qu8-dwconv/up1x9-minmax-scalar.c", |
| 469 | "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c", |
| 470 | "src/qu8-gavgpool/7x-minmax-scalar-c1.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 471 | "src/qu8-gemm/2x2-minmax-scalar.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 472 | "src/qu8-gemm/gen/8x8c4-minmax-scalar.c", |
| 473 | "src/qu8-gemm/gen/12x4c4-minmax-scalar.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 474 | "src/qu8-igemm/2x2-minmax-scalar.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 475 | "src/qu8-requantization/fp32-scalar-lrintf.c", |
| 476 | "src/qu8-requantization/fp32-scalar-magic.c", |
| 477 | "src/qu8-requantization/precise-scalar-signed64.c", |
| 478 | "src/qu8-requantization/precise-scalar-unsigned32.c", |
| 479 | "src/qu8-requantization/precise-scalar-unsigned64.c", |
| 480 | "src/qu8-requantization/q31-scalar.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 481 | "src/qu8-vadd/minmax-scalar.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 482 | "src/u8-clamp/scalar-x4.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 483 | "src/u8-lut32norm/scalar.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 484 | "src/u8-maxpool/9p8x-minmax-scalar-c1.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 485 | "src/u8-rmax/scalar.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 486 | "src/x8-lut/scalar.c", |
| 487 | "src/x8-zip/x2-scalar.c", |
| 488 | "src/x8-zip/x3-scalar.c", |
| 489 | "src/x8-zip/x4-scalar.c", |
| 490 | "src/x8-zip/xm-scalar.c", |
| Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 491 | "src/x32-fill/scalar-float.c", |
| 492 | "src/x32-fill/scalar-int.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 493 | "src/x32-packx/x2-scalar.c", |
| 494 | "src/x32-packx/x3-scalar.c", |
| 495 | "src/x32-packx/x4-scalar.c", |
| Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 496 | "src/x32-pad/scalar-float.c", |
| Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 497 | "src/x32-pad/scalar-int.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 498 | "src/x32-unpool/scalar.c", |
| 499 | "src/x32-zip/x2-scalar.c", |
| 500 | "src/x32-zip/x3-scalar.c", |
| 501 | "src/x32-zip/x4-scalar.c", |
| 502 | "src/x32-zip/xm-scalar.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 503 | ] |
| 504 | |
| Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 505 | WASM_UKERNELS = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 506 | "src/f32-avgpool/9p8x-minmax-wasm-c1.c", |
| 507 | "src/f32-avgpool/9x-minmax-wasm-c1.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 508 | "src/f32-clamp/gen/wasm-x1.c", |
| 509 | "src/f32-clamp/gen/wasm-x2.c", |
| 510 | "src/f32-clamp/gen/wasm-x4.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 511 | "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c", |
| 512 | "src/f32-dwconv/gen/up1x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 513 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 514 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 515 | "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c", |
| 516 | "src/f32-dwconv/gen/up1x9-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 517 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 518 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 519 | "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c", |
| 520 | "src/f32-dwconv/gen/up1x25-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 521 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 522 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 523 | "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c", |
| 524 | "src/f32-dwconv/gen/up2x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 525 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 526 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 527 | "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c", |
| 528 | "src/f32-dwconv/gen/up2x9-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 529 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 530 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 531 | "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c", |
| 532 | "src/f32-dwconv/gen/up2x25-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 533 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 534 | "src/f32-dwconv/gen/up2x25-wasm.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 535 | "src/f32-gavgpool/7p7x-minmax-wasm-c1.c", |
| 536 | "src/f32-gavgpool/7x-minmax-wasm-c1.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 537 | "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c", |
| 538 | "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c", |
| 539 | "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c", |
| 540 | "src/f32-gemm/gen/1x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 541 | "src/f32-gemm/gen/1x4-relu-wasm.c", |
| 542 | "src/f32-gemm/gen/1x4-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 543 | "src/f32-gemm/gen/2x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 544 | "src/f32-gemm/gen/2x4-relu-wasm.c", |
| 545 | "src/f32-gemm/gen/2x4-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 546 | "src/f32-gemm/gen/4x2-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 547 | "src/f32-gemm/gen/4x2-relu-wasm.c", |
| 548 | "src/f32-gemm/gen/4x2-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 549 | "src/f32-gemm/gen/4x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 550 | "src/f32-gemm/gen/4x4-relu-wasm.c", |
| 551 | "src/f32-gemm/gen/4x4-wasm.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 552 | "src/f32-hswish/gen/hswish-wasm-x1.c", |
| 553 | "src/f32-hswish/gen/hswish-wasm-x2.c", |
| 554 | "src/f32-hswish/gen/hswish-wasm-x4.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 555 | "src/f32-igemm/gen/1x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 556 | "src/f32-igemm/gen/1x4-relu-wasm.c", |
| 557 | "src/f32-igemm/gen/1x4-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 558 | "src/f32-igemm/gen/2x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 559 | "src/f32-igemm/gen/2x4-relu-wasm.c", |
| 560 | "src/f32-igemm/gen/2x4-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 561 | "src/f32-igemm/gen/4x2-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 562 | "src/f32-igemm/gen/4x2-relu-wasm.c", |
| 563 | "src/f32-igemm/gen/4x2-wasm.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 564 | "src/f32-igemm/gen/4x4-minmax-wasm.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 565 | "src/f32-igemm/gen/4x4-relu-wasm.c", |
| 566 | "src/f32-igemm/gen/4x4-wasm.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 567 | "src/f32-maxpool/9p8x-minmax-wasm-c1.c", |
| 568 | "src/f32-pavgpool/9p8x-minmax-wasm-c1.c", |
| 569 | "src/f32-pavgpool/9x-minmax-wasm-c1.c", |
| Marat Dukhan | 7c1f808 | 2020-06-25 13:26:20 -0700 | [diff] [blame] | 570 | "src/f32-prelu/gen/wasm-2x1.c", |
| 571 | "src/f32-prelu/gen/wasm-2x4.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 572 | "src/f32-relu/gen/wasm-x1.c", |
| 573 | "src/f32-relu/gen/wasm-x2.c", |
| 574 | "src/f32-relu/gen/wasm-x4.c", |
| Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 575 | "src/f32-relu/gen/wasm-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 576 | "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c", |
| 577 | "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c", |
| 578 | "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 579 | "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 580 | "src/f32-vbinary/gen/vadd-relu-wasm-x1.c", |
| 581 | "src/f32-vbinary/gen/vadd-relu-wasm-x2.c", |
| 582 | "src/f32-vbinary/gen/vadd-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 583 | "src/f32-vbinary/gen/vadd-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 584 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c", |
| 585 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c", |
| 586 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c", |
| 587 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 588 | "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c", |
| 589 | "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c", |
| 590 | "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 591 | "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 592 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c", |
| 593 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c", |
| 594 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c", |
| 595 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 596 | "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c", |
| 597 | "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c", |
| 598 | "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 599 | "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 600 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c", |
| 601 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c", |
| 602 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c", |
| 603 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 604 | "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c", |
| 605 | "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c", |
| 606 | "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 607 | "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 608 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 609 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 610 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 611 | "src/f32-vbinary/gen/vmax-wasm-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 612 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 613 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 614 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 615 | "src/f32-vbinary/gen/vmaxc-wasm-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 616 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 617 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 618 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 619 | "src/f32-vbinary/gen/vmin-wasm-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 620 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 621 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 622 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 623 | "src/f32-vbinary/gen/vminc-wasm-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 624 | "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c", |
| 625 | "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c", |
| 626 | "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 627 | "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 628 | "src/f32-vbinary/gen/vmul-relu-wasm-x1.c", |
| 629 | "src/f32-vbinary/gen/vmul-relu-wasm-x2.c", |
| 630 | "src/f32-vbinary/gen/vmul-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 631 | "src/f32-vbinary/gen/vmul-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 632 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c", |
| 633 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c", |
| 634 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c", |
| 635 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 636 | "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c", |
| 637 | "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c", |
| 638 | "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 639 | "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 640 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c", |
| 641 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c", |
| 642 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c", |
| 643 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 644 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c", |
| 645 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c", |
| 646 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 647 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 648 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c", |
| 649 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c", |
| 650 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c", |
| 651 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 652 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c", |
| 653 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c", |
| 654 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 655 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 656 | "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c", |
| 657 | "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c", |
| 658 | "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c", |
| 659 | "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 660 | "src/f32-vbinary/gen/vsub-relu-wasm-x1.c", |
| 661 | "src/f32-vbinary/gen/vsub-relu-wasm-x2.c", |
| 662 | "src/f32-vbinary/gen/vsub-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 663 | "src/f32-vbinary/gen/vsub-relu-wasm-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 664 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c", |
| 665 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c", |
| 666 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c", |
| 667 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 668 | "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c", |
| 669 | "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c", |
| 670 | "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 671 | "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 672 | "src/f32-vlrelu/gen/vlrelu-wasm-x1.c", |
| 673 | "src/f32-vlrelu/gen/vlrelu-wasm-x2.c", |
| 674 | "src/f32-vlrelu/gen/vlrelu-wasm-x4.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 675 | "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c", |
| 676 | "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c", |
| 677 | "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c", |
| Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 678 | ] |
| 679 | |
| Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 680 | WASMSIMD_UKERNELS = [ |
| Marat Dukhan | 40f0552 | 2020-07-16 22:33:12 -0700 | [diff] [blame] | 681 | "src/f32-argmaxpool/4x-wasmsimd-c4.c", |
| 682 | "src/f32-argmaxpool/9p8x-wasmsimd-c4.c", |
| 683 | "src/f32-argmaxpool/9x-wasmsimd-c4.c", |
| Marat Dukhan | 3b7432d | 2020-07-16 17:46:32 -0700 | [diff] [blame] | 684 | "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 685 | "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| 686 | "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c", |
| 687 | "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c", |
| Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 688 | "src/f32-clamp/gen/wasmsimd-arm-x4.c", |
| 689 | "src/f32-clamp/gen/wasmsimd-arm-x8.c", |
| 690 | "src/f32-clamp/gen/wasmsimd-x86-x4.c", |
| 691 | "src/f32-clamp/gen/wasmsimd-x86-x8.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 692 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 693 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 694 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 695 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 696 | "src/f32-dwconv/gen/up4x4-wasmsimd.c", |
| 697 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 698 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 699 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 700 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 701 | "src/f32-dwconv/gen/up4x9-wasmsimd.c", |
| 702 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c", |
| 703 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c", |
| 704 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c", |
| 705 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c", |
| 706 | "src/f32-dwconv/gen/up4x25-wasmsimd.c", |
| 707 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 708 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 709 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 710 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 711 | "src/f32-dwconv/gen/up8x4-wasmsimd.c", |
| 712 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 713 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 714 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c", |
| Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 715 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 716 | "src/f32-dwconv/gen/up8x9-wasmsimd.c", |
| 717 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c", |
| 718 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c", |
| 719 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c", |
| 720 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c", |
| 721 | "src/f32-dwconv/gen/up8x25-wasmsimd.c", |
| 722 | "src/f32-gavgpool-cw/wasmsimd-arm-x4.c", |
| 723 | "src/f32-gavgpool-cw/wasmsimd-x86-x4.c", |
| Marat Dukhan | c601680 | 2020-07-16 18:51:28 -0700 | [diff] [blame] | 724 | "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c", |
| 725 | "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c", |
| 726 | "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c", |
| 727 | "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 728 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 729 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 730 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c", |
| 731 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c", |
| 732 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c", |
| 733 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c", |
| 734 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 735 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 736 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c", |
| 737 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c", |
| 738 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c", |
| 739 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c", |
| 740 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 741 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 742 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c", |
| 743 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c", |
| 744 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c", |
| 745 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c", |
| 746 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 747 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 748 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c", |
| 749 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c", |
| 750 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c", |
| 751 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c", |
| 752 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 753 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 754 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c", |
| 755 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c", |
| 756 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c", |
| 757 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c", |
| 758 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 759 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 760 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 761 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 762 | "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c", |
| 763 | "src/f32-gemm/gen/1x8-wasmsimd-splat.c", |
| 764 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 765 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 766 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 767 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 768 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 769 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 770 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 771 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 772 | "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c", |
| 773 | "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c", |
| 774 | "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c", |
| 775 | "src/f32-gemm/gen/4x2c4-wasmsimd.c", |
| 776 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 777 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 778 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 779 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 780 | "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c", |
| 781 | "src/f32-gemm/gen/4x8-wasmsimd-splat.c", |
| 782 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 783 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 784 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 785 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 786 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 787 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 788 | "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c", |
| 789 | "src/f32-gemm/gen/5x8-wasmsimd-splat.c", |
| 790 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 791 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 792 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 793 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 794 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 795 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 796 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 797 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 798 | "src/f32-hswish/gen/hswish-wasmsimd-x4.c", |
| 799 | "src/f32-hswish/gen/hswish-wasmsimd-x8.c", |
| 800 | "src/f32-hswish/gen/hswish-wasmsimd-x16.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 801 | "src/f32-ibilinear/gen/wasmsimd-c4.c", |
| 802 | "src/f32-ibilinear/gen/wasmsimd-c8.c", |
| XNNPACK Team | 965272b | 2020-10-23 21:10:15 -0700 | [diff] [blame] | 803 | "src/f32-ibilinear-chw/gen/wasmsimd-p4.c", |
| 804 | "src/f32-ibilinear-chw/gen/wasmsimd-p8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 805 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 806 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 807 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 808 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 809 | "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c", |
| 810 | "src/f32-igemm/gen/1x8-wasmsimd-splat.c", |
| 811 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 812 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 813 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 814 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 815 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 816 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 817 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 818 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 819 | "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c", |
| 820 | "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c", |
| 821 | "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c", |
| 822 | "src/f32-igemm/gen/4x2c4-wasmsimd.c", |
| 823 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 824 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 825 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 826 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 827 | "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c", |
| 828 | "src/f32-igemm/gen/4x8-wasmsimd-splat.c", |
| 829 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 830 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 831 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 832 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 833 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 834 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 835 | "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c", |
| 836 | "src/f32-igemm/gen/5x8-wasmsimd-splat.c", |
| 837 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 838 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 839 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 840 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 841 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 842 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 843 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 844 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
| Marat Dukhan | f6e2480 | 2020-07-08 22:20:40 -0700 | [diff] [blame] | 845 | "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 846 | "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| Marat Dukhan | 1483c53 | 2020-07-16 18:08:19 -0700 | [diff] [blame] | 847 | "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 848 | "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| 849 | "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c", |
| 850 | "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c", |
| Marat Dukhan | 0d0d882 | 2020-07-23 23:37:56 -0700 | [diff] [blame] | 851 | "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 852 | "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 853 | "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c", |
| 854 | "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c", |
| 855 | "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c", |
| Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 856 | "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c", |
| 857 | "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c", |
| Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 858 | "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c", |
| 859 | "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c", |
| 860 | "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c", |
| 861 | "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c", |
| 862 | "src/f32-prelu/gen/wasmsimd-minmax-1x4.c", |
| 863 | "src/f32-prelu/gen/wasmsimd-minmax-1x8.c", |
| 864 | "src/f32-prelu/gen/wasmsimd-minmax-1x16.c", |
| Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 865 | "src/f32-prelu/gen/wasmsimd-minmax-2x4.c", |
| 866 | "src/f32-prelu/gen/wasmsimd-minmax-2x8.c", |
| Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 867 | "src/f32-prelu/gen/wasmsimd-minmax-2x16.c", |
| 868 | "src/f32-prelu/gen/wasmsimd-minmax-4x4.c", |
| 869 | "src/f32-prelu/gen/wasmsimd-minmax-4x8.c", |
| 870 | "src/f32-prelu/gen/wasmsimd-minmax-4x16.c", |
| Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 871 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c", |
| Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 872 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 873 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c", |
| Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 874 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c", |
| 875 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 876 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c", |
| Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 877 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c", |
| 878 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 879 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c", |
| Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 880 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c", |
| 881 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 882 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 883 | "src/f32-relu/gen/wasmsimd-x4.c", |
| 884 | "src/f32-relu/gen/wasmsimd-x8.c", |
| Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 885 | "src/f32-relu/gen/wasmsimd-x16.c", |
| Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 886 | "src/f32-rmax/wasmsimd-arm.c", |
| 887 | "src/f32-rmax/wasmsimd-x86.c", |
| Marat Dukhan | d187a5b | 2020-07-20 01:07:17 -0700 | [diff] [blame] | 888 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c", |
| 889 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c", |
| 890 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c", |
| 891 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c", |
| 892 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c", |
| 893 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 894 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c", |
| 895 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c", |
| 896 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c", |
| 897 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c", |
| 898 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c", |
| 899 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c", |
| 900 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll2.c", |
| 901 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll4.c", |
| 902 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c", |
| 903 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll2.c", |
| 904 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll4.c", |
| 905 | "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c", |
| 906 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll2.c", |
| 907 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll4.c", |
| 908 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c", |
| 909 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll2.c", |
| 910 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll4.c", |
| 911 | "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c", |
| 912 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll2.c", |
| 913 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll4.c", |
| 914 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c", |
| 915 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll2.c", |
| 916 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll4.c", |
| 917 | "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 918 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c", |
| 919 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 920 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 921 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c", |
| 922 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 923 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 924 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c", |
| 925 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 926 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 927 | "src/f32-vbinary/gen/vadd-wasmsimd-x4.c", |
| 928 | "src/f32-vbinary/gen/vadd-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 929 | "src/f32-vbinary/gen/vadd-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 930 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c", |
| 931 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 932 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 933 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c", |
| 934 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 935 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 936 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c", |
| 937 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 938 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 939 | "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c", |
| 940 | "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 941 | "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 942 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c", |
| 943 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 944 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 945 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c", |
| 946 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 947 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 948 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c", |
| 949 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 950 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 951 | "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c", |
| 952 | "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 953 | "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 954 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c", |
| 955 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 956 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 957 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c", |
| 958 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 959 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 960 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c", |
| 961 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 962 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 963 | "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c", |
| 964 | "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 965 | "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 966 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c", |
| 967 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 968 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 969 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c", |
| 970 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 971 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 972 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c", |
| 973 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 974 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 975 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c", |
| 976 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 977 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 978 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c", |
| 979 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 980 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 981 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c", |
| 982 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 983 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 984 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c", |
| 985 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 986 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 987 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c", |
| 988 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 989 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 990 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c", |
| 991 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 992 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 993 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c", |
| 994 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 995 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 996 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c", |
| 997 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 998 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 999 | "src/f32-vbinary/gen/vmul-wasmsimd-x4.c", |
| 1000 | "src/f32-vbinary/gen/vmul-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1001 | "src/f32-vbinary/gen/vmul-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1002 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c", |
| 1003 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1004 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1005 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c", |
| 1006 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1007 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1008 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c", |
| 1009 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1010 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1011 | "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c", |
| 1012 | "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1013 | "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1014 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c", |
| 1015 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1016 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1017 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c", |
| 1018 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1019 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1020 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c", |
| 1021 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1022 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1023 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c", |
| 1024 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1025 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1026 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c", |
| 1027 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1028 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1029 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c", |
| 1030 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1031 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1032 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c", |
| 1033 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1034 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1035 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c", |
| 1036 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1037 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1038 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c", |
| 1039 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1040 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1041 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c", |
| 1042 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1043 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1044 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c", |
| 1045 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1046 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 1047 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c", |
| 1048 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1049 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1050 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c", |
| 1051 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1052 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1053 | "src/f32-vbinary/gen/vsub-wasmsimd-x4.c", |
| 1054 | "src/f32-vbinary/gen/vsub-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1055 | "src/f32-vbinary/gen/vsub-wasmsimd-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1056 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c", |
| 1057 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1058 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c", |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1059 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c", |
| 1060 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1061 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c", |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1062 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c", |
| 1063 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1064 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c", |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1065 | "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c", |
| 1066 | "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c", |
| Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1067 | "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1068 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c", |
| 1069 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c", |
| 1070 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c", |
| 1071 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c", |
| Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1072 | "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c", |
| Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1073 | "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1074 | "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c", |
| Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1075 | "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c", |
| Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 1076 | "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c", |
| 1077 | "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c", |
| 1078 | "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c", |
| 1079 | "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1080 | "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c", |
| 1081 | "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c", |
| 1082 | "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c", |
| 1083 | "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c", |
| 1084 | "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c", |
| 1085 | "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c", |
| 1086 | "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c", |
| 1087 | "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c", |
| 1088 | "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c", |
| 1089 | "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1090 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c", |
| 1091 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c", |
| Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 1092 | "src/f32-vunary/gen/vabs-wasmsimd-x4.c", |
| 1093 | "src/f32-vunary/gen/vabs-wasmsimd-x8.c", |
| 1094 | "src/f32-vunary/gen/vneg-wasmsimd-x4.c", |
| 1095 | "src/f32-vunary/gen/vneg-wasmsimd-x8.c", |
| 1096 | "src/f32-vunary/gen/vsqr-wasmsimd-x4.c", |
| 1097 | "src/f32-vunary/gen/vsqr-wasmsimd-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1098 | "src/math/roundd-wasmsimd-addsub.c", |
| 1099 | "src/math/roundd-wasmsimd-cvt.c", |
| 1100 | "src/math/roundne-wasmsimd-addsub.c", |
| 1101 | "src/math/roundu-wasmsimd-addsub.c", |
| 1102 | "src/math/roundu-wasmsimd-cvt.c", |
| 1103 | "src/math/roundz-wasmsimd-addsub.c", |
| 1104 | "src/math/roundz-wasmsimd-cvt.c", |
| 1105 | "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c", |
| 1106 | "src/math/sigmoid-wasmsimd-rr2-p5-div.c", |
| Marat Dukhan | cc8f34c | 2020-08-05 16:36:38 -0700 | [diff] [blame] | 1107 | "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c", |
| 1108 | "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c", |
| 1109 | "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c", |
| Marat Dukhan | b5e3d17 | 2020-08-06 13:29:53 -0700 | [diff] [blame] | 1110 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c", |
| 1111 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c", |
| 1112 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1113 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c", |
| 1114 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c", |
| 1115 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1116 | "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1117 | "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1118 | "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1119 | "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c", |
| 1120 | "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1121 | "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1122 | "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c", |
| 1123 | "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1124 | "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c", |
| 1125 | "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1126 | "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1127 | "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1128 | "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1129 | "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c", |
| Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1130 | "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c", |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1131 | "src/qs8-requantization/fp32-wasmsimd.c", |
| Marat Dukhan | 138560c | 2020-08-03 18:57:34 -0700 | [diff] [blame] | 1132 | "src/qs8-requantization/q31-wasmsimd.c", |
| Marat Dukhan | 5df27f8 | 2020-09-02 23:59:21 -0700 | [diff] [blame] | 1133 | "src/qs8-vadd/gen/minmax-wasmsimd-x8.c", |
| 1134 | "src/qs8-vadd/gen/minmax-wasmsimd-x16.c", |
| 1135 | "src/qs8-vadd/gen/minmax-wasmsimd-x24.c", |
| 1136 | "src/qs8-vadd/gen/minmax-wasmsimd-x32.c", |
| 1137 | "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c", |
| 1138 | "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c", |
| 1139 | "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c", |
| 1140 | "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1141 | "src/qu8-requantization/fp32-wasmsimd.c", |
| Marat Dukhan | 22de5e7 | 2020-08-03 21:53:18 -0700 | [diff] [blame] | 1142 | "src/qu8-requantization/q31-wasmsimd.c", |
| Marat Dukhan | 8ee3701 | 2020-07-16 13:17:13 -0700 | [diff] [blame] | 1143 | "src/x32-fill/wasmsimd.c", |
| Marat Dukhan | 66d99e9 | 2020-07-16 12:56:21 -0700 | [diff] [blame] | 1144 | "src/x32-packx/x4-wasmsimd.c", |
| Marat Dukhan | 9306ae0 | 2020-07-16 15:51:13 -0700 | [diff] [blame] | 1145 | "src/x32-pad/wasmsimd.c", |
| Marat Dukhan | 9d4bfa2 | 2020-07-16 19:07:04 -0700 | [diff] [blame] | 1146 | "src/x32-unpool/wasmsimd.c", |
| Marat Dukhan | e3b7876 | 2020-07-16 20:02:58 -0700 | [diff] [blame] | 1147 | "src/x32-zip/x2-wasmsimd.c", |
| 1148 | "src/x32-zip/x3-wasmsimd.c", |
| 1149 | "src/x32-zip/x4-wasmsimd.c", |
| 1150 | "src/x32-zip/xm-wasmsimd.c", |
| Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 1151 | ] |
| 1152 | |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1153 | PSIMD_FASTMATH_UKERNELS = [ |
| Erich Elsen | 0a1970e | 2020-06-10 09:24:59 -0700 | [diff] [blame] | 1154 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c", |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1155 | "src/f32-dwconv2d-chw/3x3p1-psimd-1x4-acc3.c", |
| 1156 | "src/f32-dwconv2d-chw/3x3s2p1-psimd-1x4-acc3.c", |
| 1157 | "src/f32-dwconv2d-chw/5x5p2-psimd-3x4.c", |
| 1158 | "src/f32-dwconv2d-chw/5x5s2p2-psimd-1x4-acc2.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1159 | ] |
| 1160 | |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1161 | PSIMD_ACCMATH_UKERNELS = [ |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1162 | "src/qs8-requantization/fp32-psimd.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1163 | "src/qs8-requantization/precise-psimd.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1164 | "src/qu8-requantization/fp32-psimd.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1165 | "src/qu8-requantization/precise-psimd.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1166 | ] |
| 1167 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1168 | # ISA-specific micro-kernels |
| 1169 | NEON_UKERNELS = [ |
| Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 1170 | "src/f32-argmaxpool/4x-neon-c4.c", |
| 1171 | "src/f32-argmaxpool/9p8x-neon-c4.c", |
| 1172 | "src/f32-argmaxpool/9x-neon-c4.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1173 | "src/f32-avgpool/9p8x-minmax-neon-c4.c", |
| 1174 | "src/f32-avgpool/9x-minmax-neon-c4.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1175 | "src/f32-clamp/gen/neon-x4.c", |
| 1176 | "src/f32-clamp/gen/neon-x8.c", |
| Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1177 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1178 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1179 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1180 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c", |
| Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1181 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1182 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1183 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1184 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c", |
| Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1185 | "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1186 | "src/f32-dwconv/gen/up4x4-minmax-neon.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1187 | "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1188 | "src/f32-dwconv/gen/up4x9-minmax-neon.c", |
| Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1189 | "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1190 | "src/f32-dwconv/gen/up4x25-minmax-neon.c", |
| 1191 | "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", |
| 1192 | "src/f32-dwconv/gen/up8x4-minmax-neon.c", |
| 1193 | "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", |
| 1194 | "src/f32-dwconv/gen/up8x9-minmax-neon.c", |
| Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1195 | "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1196 | "src/f32-dwconv/gen/up8x25-minmax-neon.c", |
| Marat Dukhan | c581e48 | 2020-10-24 01:28:11 -0700 | [diff] [blame] | 1197 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c", |
| 1198 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c", |
| 1199 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c", |
| 1200 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c", |
| 1201 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c", |
| 1202 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c", |
| 1203 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c", |
| 1204 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c", |
| 1205 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c", |
| 1206 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c", |
| Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 1207 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c", |
| 1208 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c", |
| 1209 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c", |
| 1210 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c", |
| 1211 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c", |
| 1212 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c", |
| 1213 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c", |
| 1214 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c", |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1215 | "src/f32-gavgpool-cw/neon-x4.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1216 | "src/f32-gavgpool/7p7x-minmax-neon-c4.c", |
| 1217 | "src/f32-gavgpool/7x-minmax-neon-c4.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1218 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1219 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c", |
| 1220 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1221 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1222 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c", |
| 1223 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c", |
| 1224 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c", |
| 1225 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c", |
| 1226 | "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1227 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c", |
| 1228 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1229 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c", |
| 1230 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1231 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c", |
| 1232 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1233 | "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 1234 | "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 1235 | "src/f32-gemm/gen/1x8s4-minmax-neon.c", |
| 1236 | "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 1237 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 1238 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 1239 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 1240 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 1241 | "src/f32-gemm/gen/4x8s4-minmax-neon.c", |
| 1242 | "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c", |
| 1243 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 1244 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 1245 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 1246 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 1247 | "src/f32-gemm/gen/6x8s4-minmax-neon.c", |
| 1248 | "src/f32-gemm/gen/8x8s4-minmax-neon.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1249 | "src/f32-hswish/gen/hswish-neon-x4.c", |
| 1250 | "src/f32-hswish/gen/hswish-neon-x8.c", |
| 1251 | "src/f32-hswish/gen/hswish-neon-x16.c", |
| Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1252 | "src/f32-ibilinear/gen/neon-c4.c", |
| 1253 | "src/f32-ibilinear/gen/neon-c8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1254 | "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1255 | "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1256 | "src/f32-igemm/gen/1x8s4-minmax-neon.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1257 | "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 1258 | "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1259 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1260 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 1261 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 1262 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 1263 | "src/f32-igemm/gen/4x8s4-minmax-neon.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1264 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 1265 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1266 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 1267 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1268 | "src/f32-igemm/gen/6x8s4-minmax-neon.c", |
| 1269 | "src/f32-igemm/gen/8x8s4-minmax-neon.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1270 | "src/f32-maxpool/9p8x-minmax-neon-c4.c", |
| 1271 | "src/f32-pavgpool/9p8x-minmax-neon-c4.c", |
| 1272 | "src/f32-pavgpool/9x-minmax-neon-c4.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1273 | "src/f32-ppmm/gen/4x8-minmax-neon.c", |
| 1274 | "src/f32-ppmm/gen/8x8-minmax-neon.c", |
| Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1275 | "src/f32-prelu/gen/neon-1x4.c", |
| 1276 | "src/f32-prelu/gen/neon-1x8.c", |
| 1277 | "src/f32-prelu/gen/neon-1x16.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1278 | "src/f32-prelu/gen/neon-2x4.c", |
| 1279 | "src/f32-prelu/gen/neon-2x8.c", |
| Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1280 | "src/f32-prelu/gen/neon-2x16.c", |
| 1281 | "src/f32-prelu/gen/neon-4x4.c", |
| 1282 | "src/f32-prelu/gen/neon-4x8.c", |
| 1283 | "src/f32-prelu/gen/neon-4x16.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1284 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1285 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1286 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1287 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 1288 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1289 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1290 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 1291 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1292 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1293 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 1294 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1295 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 1296 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 1297 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 1298 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 1299 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 1300 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 1301 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 1302 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 1303 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 1304 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 1305 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 1306 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 1307 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1308 | "src/f32-relu/gen/neon-x4.c", |
| 1309 | "src/f32-relu/gen/neon-x8.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1310 | "src/f32-rmax/neon.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1311 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
| Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1312 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 1313 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 1314 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 1315 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 1316 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 1317 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 1318 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 1319 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 1320 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 1321 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 1322 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 1323 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1324 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 1325 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 1326 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 1327 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 1328 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 1329 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1330 | "src/f32-vbinary/gen/vadd-minmax-neon-x4.c", |
| 1331 | "src/f32-vbinary/gen/vadd-minmax-neon-x8.c", |
| 1332 | "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c", |
| 1333 | "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1334 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 1335 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 1336 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 1337 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 1338 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 1339 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 1340 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 1341 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1342 | "src/f32-vbinary/gen/vmul-minmax-neon-x4.c", |
| 1343 | "src/f32-vbinary/gen/vmul-minmax-neon-x8.c", |
| 1344 | "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c", |
| 1345 | "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c", |
| 1346 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c", |
| 1347 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c", |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1348 | "src/f32-vbinary/gen/vsqrdiff-neon-x4.c", |
| 1349 | "src/f32-vbinary/gen/vsqrdiff-neon-x8.c", |
| 1350 | "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c", |
| 1351 | "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1352 | "src/f32-vbinary/gen/vsub-minmax-neon-x4.c", |
| 1353 | "src/f32-vbinary/gen/vsub-minmax-neon-x8.c", |
| 1354 | "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c", |
| 1355 | "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1356 | "src/f32-vlrelu/gen/vlrelu-neon-x4.c", |
| 1357 | "src/f32-vlrelu/gen/vlrelu-neon-x8.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1358 | "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c", |
| 1359 | "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c", |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1360 | "src/f32-vrnd/gen/vrndd-neon-x4.c", |
| 1361 | "src/f32-vrnd/gen/vrndd-neon-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1362 | "src/f32-vrnd/gen/vrndne-neon-x4.c", |
| 1363 | "src/f32-vrnd/gen/vrndne-neon-x8.c", |
| 1364 | "src/f32-vrnd/gen/vrndu-neon-x4.c", |
| 1365 | "src/f32-vrnd/gen/vrndu-neon-x8.c", |
| 1366 | "src/f32-vrnd/gen/vrndz-neon-x4.c", |
| 1367 | "src/f32-vrnd/gen/vrndz-neon-x8.c", |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1368 | "src/f32-vunary/gen/vabs-neon-x4.c", |
| 1369 | "src/f32-vunary/gen/vabs-neon-x8.c", |
| 1370 | "src/f32-vunary/gen/vneg-neon-x4.c", |
| 1371 | "src/f32-vunary/gen/vneg-neon-x8.c", |
| 1372 | "src/f32-vunary/gen/vsqr-neon-x4.c", |
| 1373 | "src/f32-vunary/gen/vsqr-neon-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1374 | "src/math/roundd-neon-addsub.c", |
| 1375 | "src/math/roundd-neon-cvt.c", |
| 1376 | "src/math/roundne-neon-addsub.c", |
| 1377 | "src/math/roundu-neon-addsub.c", |
| 1378 | "src/math/roundu-neon-cvt.c", |
| 1379 | "src/math/roundz-neon-addsub.c", |
| 1380 | "src/math/roundz-neon-cvt.c", |
| 1381 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
| 1382 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 1383 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 1384 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 1385 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 1386 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 1387 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
| 1388 | "src/math/sqrt-neon-nr1rsqrts.c", |
| 1389 | "src/math/sqrt-neon-nr2rsqrts.c", |
| 1390 | "src/math/sqrt-neon-nr3rsqrts.c", |
| Marat Dukhan | 023bcf9 | 2020-08-10 12:40:50 -0700 | [diff] [blame] | 1391 | "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c", |
| 1392 | "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c", |
| 1393 | "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c", |
| 1394 | "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c", |
| Marat Dukhan | 281262d | 2020-08-10 13:23:21 -0700 | [diff] [blame] | 1395 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c", |
| 1396 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c", |
| 1397 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c", |
| 1398 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1399 | "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c", |
| 1400 | "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c", |
| 1401 | "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c", |
| 1402 | "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c", |
| Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1403 | "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c", |
| Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1404 | "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1405 | "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c", |
| Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1406 | "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c", |
| 1407 | "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c", |
| Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1408 | "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1409 | "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c", |
| Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1410 | "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c", |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1411 | "src/qs8-requantization/fp32-neon.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1412 | "src/qs8-requantization/precise-neon.c", |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1413 | "src/qs8-requantization/q31-neon.c", |
| Marat Dukhan | ba7b279 | 2020-09-02 14:26:45 -0700 | [diff] [blame] | 1414 | "src/qs8-vadd/gen/minmax-neon-ld64-x8.c", |
| 1415 | "src/qs8-vadd/gen/minmax-neon-ld64-x16.c", |
| 1416 | "src/qs8-vadd/gen/minmax-neon-ld64-x24.c", |
| 1417 | "src/qs8-vadd/gen/minmax-neon-ld64-x32.c", |
| 1418 | "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c", |
| 1419 | "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c", |
| 1420 | "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c", |
| 1421 | "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1422 | "src/qu8-avgpool/9p8x-minmax-neon-c8.c", |
| 1423 | "src/qu8-avgpool/9x-minmax-neon-c8.c", |
| 1424 | "src/qu8-dwconv/up8x9-minmax-neon.c", |
| 1425 | "src/qu8-gavgpool/7p7x-minmax-neon-c8.c", |
| 1426 | "src/qu8-gavgpool/7x-minmax-neon-c8.c", |
| 1427 | "src/qu8-gemm/4x8-minmax-neon.c", |
| 1428 | "src/qu8-gemm/8x8-minmax-neon.c", |
| 1429 | "src/qu8-igemm/4x8-minmax-neon.c", |
| 1430 | "src/qu8-igemm/8x8-minmax-neon.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1431 | "src/qu8-requantization/fp32-neon.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1432 | "src/qu8-requantization/precise-neon.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1433 | "src/qu8-requantization/q31-neon.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1434 | "src/qu8-vadd/minmax-neon.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1435 | "src/u8-clamp/neon-x64.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1436 | "src/u8-maxpool/9p8x-minmax-neon-c16.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1437 | "src/u8-rmax/neon.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1438 | "src/x8-zip/x2-neon.c", |
| 1439 | "src/x8-zip/x3-neon.c", |
| 1440 | "src/x8-zip/x4-neon.c", |
| 1441 | "src/x8-zip/xm-neon.c", |
| Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1442 | "src/x32-fill/neon.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1443 | "src/x32-packx/x4-neon-st4.c", |
| Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1444 | "src/x32-pad/neon.c", |
| Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1445 | "src/x32-unpool/neon.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1446 | "src/x32-zip/x2-neon.c", |
| 1447 | "src/x32-zip/x3-neon.c", |
| 1448 | "src/x32-zip/x4-neon.c", |
| 1449 | "src/x32-zip/xm-neon.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1450 | ] |
| 1451 | |
| 1452 | NEONFMA_UKERNELS = [ |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1453 | "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", |
| 1454 | "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", |
| 1455 | "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", |
| 1456 | "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", |
| 1457 | "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", |
| 1458 | "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", |
| 1459 | "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", |
| 1460 | "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", |
| 1461 | "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", |
| 1462 | "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", |
| 1463 | "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", |
| 1464 | "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", |
| 1465 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c", |
| 1466 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c", |
| 1467 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c", |
| 1468 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c", |
| 1469 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c", |
| 1470 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c", |
| 1471 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c", |
| 1472 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c", |
| 1473 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c", |
| 1474 | "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1475 | "src/f32-gemm/gen/1x8s4-minmax-neonfma.c", |
| 1476 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1477 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1478 | "src/f32-gemm/gen/4x8s4-minmax-neonfma.c", |
| 1479 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1480 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1481 | "src/f32-gemm/gen/6x8s4-minmax-neonfma.c", |
| 1482 | "src/f32-gemm/gen/8x8s4-minmax-neonfma.c", |
| Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1483 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 1484 | "src/f32-ibilinear/gen/neonfma-c8.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1485 | "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1486 | "src/f32-igemm/gen/1x8s4-minmax-neonfma.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1487 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1488 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1489 | "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1490 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1491 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1492 | "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", |
| 1493 | "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1494 | "src/f32-ppmm/gen/4x8-minmax-neonfma.c", |
| 1495 | "src/f32-ppmm/gen/8x8-minmax-neonfma.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1496 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1497 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1498 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1499 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 1500 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1501 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1502 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 1503 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1504 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1505 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 1506 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1507 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 1508 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 1509 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 1510 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 1511 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 1512 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 1513 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 1514 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 1515 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 1516 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 1517 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 1518 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 1519 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1520 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 1521 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 1522 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 1523 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 1524 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 1525 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1526 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 1527 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 1528 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 1529 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 1530 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 1531 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1532 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 1533 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 1534 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 1535 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 1536 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 1537 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1538 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 1539 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 1540 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 1541 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 1542 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 1543 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1544 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 1545 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 1546 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 1547 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 1548 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 1549 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1550 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 1551 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 1552 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 1553 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 1554 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 1555 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1556 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 1557 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 1558 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 1559 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 1560 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 1561 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 1562 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 1563 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 1564 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 1565 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 1566 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 1567 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 1568 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 1569 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 1570 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 1571 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 1572 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 1573 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1574 | "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c", |
| 1575 | "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1576 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c", |
| 1577 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c", |
| 1578 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c", |
| 1579 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c", |
| 1580 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c", |
| 1581 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c", |
| 1582 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c", |
| 1583 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c", |
| 1584 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c", |
| 1585 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c", |
| 1586 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c", |
| 1587 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c", |
| 1588 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c", |
| 1589 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c", |
| 1590 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c", |
| 1591 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c", |
| 1592 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c", |
| 1593 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c", |
| 1594 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c", |
| 1595 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c", |
| Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 1596 | "src/math/exp-neonfma-lut64-p2.c", |
| 1597 | "src/math/exp-neonfma-p5.c", |
| Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 1598 | "src/math/expminus-neonfma-lut64-p2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1599 | "src/math/expminus-neonfma-lut2048-p1.c", |
| Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1600 | "src/math/expminus-neonfma-p5.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1601 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 1602 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 1603 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1604 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 1605 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 1606 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1607 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 1608 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 1609 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1610 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 1611 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 1612 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1613 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 1614 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 1615 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1616 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 1617 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 1618 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1619 | "src/math/sqrt-neonfma-nr1fma.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1620 | "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1621 | "src/math/sqrt-neonfma-nr2fma.c", |
| 1622 | "src/math/sqrt-neonfma-nr2fma1adj.c", |
| 1623 | "src/math/sqrt-neonfma-nr3fma.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1624 | ] |
| 1625 | |
| 1626 | AARCH64_NEONFMA_UKERNELS = [ |
| Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1627 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1628 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1629 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1630 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c", |
| Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1631 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1632 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1633 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c", |
| Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1634 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c", |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1635 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c", |
| Marat Dukhan | 1268a24 | 2020-10-24 00:36:32 -0700 | [diff] [blame] | 1636 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c", |
| 1637 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c", |
| 1638 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c", |
| 1639 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c", |
| 1640 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c", |
| 1641 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c", |
| 1642 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c", |
| 1643 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c", |
| 1644 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c", |
| 1645 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c", |
| Marat Dukhan | 82f0c32 | 2020-10-25 19:17:35 -0700 | [diff] [blame] | 1646 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c", |
| 1647 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c", |
| 1648 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c", |
| 1649 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c", |
| 1650 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c", |
| 1651 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c", |
| 1652 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c", |
| 1653 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c", |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 1654 | "src/f32-dwconv2d-chw/5x5p2-neonfma-3x4.c", |
| 1655 | "src/f32-dwconv2d-chw/5x5s2p2-neonfma-1x4-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1656 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c", |
| 1657 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c", |
| 1658 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c", |
| 1659 | "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c", |
| 1660 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c", |
| 1661 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c", |
| 1662 | "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1663 | "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1664 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1665 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1666 | "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c", |
| 1667 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1668 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
| 1669 | "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1670 | "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1671 | "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c", |
| 1672 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1673 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1674 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1675 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
| Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1676 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 1677 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 1678 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 1679 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 1680 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 1681 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 1682 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 1683 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 1684 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 1685 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 1686 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 1687 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1688 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 1689 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 1690 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 1691 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 1692 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 1693 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1694 | "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c", |
| 1695 | "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c", |
| 1696 | "src/f32-spmm/gen/4x1-minmax-neonfma.c", |
| 1697 | "src/f32-spmm/gen/4x2-minmax-neonfma.c", |
| 1698 | "src/f32-spmm/gen/4x4-minmax-neonfma.c", |
| 1699 | "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c", |
| 1700 | "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c", |
| 1701 | "src/f32-spmm/gen/8x1-minmax-neonfma.c", |
| 1702 | "src/f32-spmm/gen/8x2-minmax-neonfma.c", |
| 1703 | "src/f32-spmm/gen/8x4-minmax-neonfma.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1704 | "src/f32-spmm/gen/12x1-minmax-neonfma.c", |
| 1705 | "src/f32-spmm/gen/12x2-minmax-neonfma.c", |
| 1706 | "src/f32-spmm/gen/12x4-minmax-neonfma.c", |
| 1707 | "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c", |
| 1708 | "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c", |
| 1709 | "src/f32-spmm/gen/16x1-minmax-neonfma.c", |
| 1710 | "src/f32-spmm/gen/16x2-minmax-neonfma.c", |
| 1711 | "src/f32-spmm/gen/16x4-minmax-neonfma.c", |
| 1712 | "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c", |
| 1713 | "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c", |
| 1714 | "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c", |
| 1715 | "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c", |
| 1716 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c", |
| 1717 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1718 | "src/f32-vsqrt/gen/neon-sqrt-x4.c", |
| 1719 | "src/f32-vsqrt/gen/neon-sqrt-x8.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1720 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1721 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1722 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1723 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1724 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1725 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1726 | ] |
| 1727 | |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1728 | NEONV8_UKERNELS = [ |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1729 | "src/f32-vrnd/gen/vrndd-neonv8-x4.c", |
| 1730 | "src/f32-vrnd/gen/vrndd-neonv8-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1731 | "src/f32-vrnd/gen/vrndne-neonv8-x4.c", |
| 1732 | "src/f32-vrnd/gen/vrndne-neonv8-x8.c", |
| 1733 | "src/f32-vrnd/gen/vrndu-neonv8-x4.c", |
| 1734 | "src/f32-vrnd/gen/vrndu-neonv8-x8.c", |
| 1735 | "src/f32-vrnd/gen/vrndz-neonv8-x4.c", |
| 1736 | "src/f32-vrnd/gen/vrndz-neonv8-x8.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1737 | "src/math/roundd-neonv8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1738 | "src/math/roundne-neonv8.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1739 | "src/math/roundu-neonv8.c", |
| Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1740 | "src/math/roundz-neonv8.c", |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1741 | ] |
| 1742 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1743 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
| Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1744 | "src/f16-clamp/gen/neonfp16arith-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1745 | "src/f16-clamp/gen/neonfp16arith-x16.c", |
| Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 1746 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c", |
| 1747 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c", |
| 1748 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c", |
| 1749 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1750 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c", |
| 1751 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c", |
| 1752 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c", |
| 1753 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c", |
| 1754 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c", |
| 1755 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c", |
| 1756 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c", |
| 1757 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c", |
| Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 1758 | "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c", |
| 1759 | "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1760 | "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c", |
| 1761 | "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c", |
| 1762 | "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c", |
| 1763 | "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c", |
| 1764 | "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c", |
| 1765 | "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c", |
| 1766 | "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c", |
| 1767 | "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c", |
| 1768 | "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1769 | "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1770 | "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1771 | "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1772 | "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1773 | "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1774 | "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1775 | "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1776 | "src/f16-hswish/gen/hswish-neonfp16arith-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1777 | "src/f16-hswish/gen/hswish-neonfp16arith-x16.c", |
| 1778 | "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1779 | "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1780 | "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1781 | "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1782 | "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1783 | "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1784 | "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1785 | "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
| Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1786 | "src/f16-prelu/gen/neonfp16arith-2x8.c", |
| Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1787 | "src/f16-prelu/gen/neonfp16arith-2x16.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1788 | "src/f16-relu/gen/neonfp16arith-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1789 | "src/f16-relu/gen/neonfp16arith-x16.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1790 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1791 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1792 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1793 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1794 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1795 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1796 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1797 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c", |
| 1798 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c", |
| 1799 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c", |
| 1800 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c", |
| 1801 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c", |
| 1802 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c", |
| 1803 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c", |
| 1804 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c", |
| 1805 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c", |
| 1806 | "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c", |
| 1807 | "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c", |
| 1808 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c", |
| 1809 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c", |
| 1810 | "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c", |
| 1811 | "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c", |
| 1812 | "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c", |
| 1813 | "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c", |
| 1814 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c", |
| 1815 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c", |
| 1816 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c", |
| 1817 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c", |
| 1818 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c", |
| 1819 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c", |
| 1820 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c", |
| 1821 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c", |
| 1822 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c", |
| 1823 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c", |
| 1824 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c", |
| 1825 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c", |
| 1826 | "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c", |
| 1827 | "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1828 | ] |
| 1829 | |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1830 | NEONDOT_UKERNELS = [ |
| Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1831 | "src/qs8-gemm/gen/1x8c4-minmax-neondot.c", |
| Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1832 | "src/qs8-gemm/gen/1x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1833 | "src/qs8-gemm/gen/4x8c4-minmax-neondot.c", |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1834 | "src/qs8-gemm/gen/4x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1835 | "src/qs8-gemm/gen/6x8c4-minmax-neondot.c", |
| Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1836 | "src/qs8-gemm/gen/6x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1837 | "src/qs8-gemm/gen/8x8c4-minmax-neondot.c", |
| Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1838 | "src/qs8-gemm/gen/8x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1839 | "src/qs8-gemm/gen/12x8c4-minmax-neondot.c", |
| Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1840 | "src/qs8-igemm/gen/1x8c4-minmax-neondot.c", |
| Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1841 | "src/qs8-igemm/gen/1x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1842 | "src/qs8-igemm/gen/4x8c4-minmax-neondot.c", |
| Benoit Jacob | a05487f | 2020-08-14 13:37:21 -0700 | [diff] [blame] | 1843 | "src/qs8-igemm/gen/4x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1844 | "src/qs8-igemm/gen/6x8c4-minmax-neondot.c", |
| Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1845 | "src/qs8-igemm/gen/6x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1846 | "src/qs8-igemm/gen/8x8c4-minmax-neondot.c", |
| Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1847 | "src/qs8-igemm/gen/8x16c4-minmax-neondot.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1848 | "src/qs8-igemm/gen/12x8c4-minmax-neondot.c", |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1849 | ] |
| 1850 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1851 | SSE_UKERNELS = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1852 | "src/f32-avgpool/9p8x-minmax-sse-c4.c", |
| 1853 | "src/f32-avgpool/9x-minmax-sse-c4.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1854 | "src/f32-clamp/gen/sse-x4.c", |
| 1855 | "src/f32-clamp/gen/sse-x8.c", |
| Erich Elsen | b123340 | 2020-06-08 15:53:15 -0700 | [diff] [blame] | 1856 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c", |
| 1857 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1858 | "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c", |
| 1859 | "src/f32-dwconv/gen/up4x4-minmax-sse.c", |
| 1860 | "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c", |
| 1861 | "src/f32-dwconv/gen/up4x9-minmax-sse.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1862 | "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c", |
| 1863 | "src/f32-dwconv/gen/up4x25-minmax-sse.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1864 | "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c", |
| 1865 | "src/f32-dwconv/gen/up8x4-minmax-sse.c", |
| 1866 | "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c", |
| 1867 | "src/f32-dwconv/gen/up8x9-minmax-sse.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1868 | "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c", |
| 1869 | "src/f32-dwconv/gen/up8x25-minmax-sse.c", |
| Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 1870 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c", |
| 1871 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c", |
| 1872 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c", |
| 1873 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c", |
| 1874 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c", |
| 1875 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c", |
| 1876 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c", |
| 1877 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c", |
| 1878 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c", |
| 1879 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c", |
| Marat Dukhan | 0ff9718 | 2020-10-25 19:14:03 -0700 | [diff] [blame] | 1880 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c", |
| 1881 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c", |
| 1882 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c", |
| 1883 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c", |
| 1884 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c", |
| 1885 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c", |
| 1886 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c", |
| 1887 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c", |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1888 | "src/f32-gavgpool-cw/sse-x4.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1889 | "src/f32-gavgpool/7p7x-minmax-sse-c4.c", |
| 1890 | "src/f32-gavgpool/7x-minmax-sse-c4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1891 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c", |
| 1892 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c", |
| 1893 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c", |
| 1894 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c", |
| 1895 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c", |
| 1896 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1897 | "src/f32-gemm/gen/1x8-minmax-sse-dup.c", |
| 1898 | "src/f32-gemm/gen/1x8-minmax-sse-load1.c", |
| 1899 | "src/f32-gemm/gen/1x8s4-minmax-sse.c", |
| 1900 | "src/f32-gemm/gen/4x2c4-minmax-sse.c", |
| 1901 | "src/f32-gemm/gen/4x8-minmax-sse-dup.c", |
| 1902 | "src/f32-gemm/gen/4x8-minmax-sse-load1.c", |
| 1903 | "src/f32-gemm/gen/4x8s4-minmax-sse.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1904 | "src/f32-hswish/gen/hswish-sse-x4.c", |
| 1905 | "src/f32-hswish/gen/hswish-sse-x8.c", |
| Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1906 | "src/f32-ibilinear/gen/sse-c4.c", |
| 1907 | "src/f32-ibilinear/gen/sse-c8.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1908 | "src/f32-igemm/gen/1x8-minmax-sse-dup.c", |
| 1909 | "src/f32-igemm/gen/1x8-minmax-sse-load1.c", |
| 1910 | "src/f32-igemm/gen/1x8s4-minmax-sse.c", |
| 1911 | "src/f32-igemm/gen/4x2c4-minmax-sse.c", |
| 1912 | "src/f32-igemm/gen/4x8-minmax-sse-dup.c", |
| 1913 | "src/f32-igemm/gen/4x8-minmax-sse-load1.c", |
| 1914 | "src/f32-igemm/gen/4x8s4-minmax-sse.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1915 | "src/f32-maxpool/9p8x-minmax-sse-c4.c", |
| 1916 | "src/f32-pavgpool/9p8x-minmax-sse-c4.c", |
| 1917 | "src/f32-pavgpool/9x-minmax-sse-c4.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1918 | "src/f32-ppmm/gen/4x8-minmax-sse.c", |
| Marat Dukhan | 39b5e94 | 2020-06-24 15:03:48 -0700 | [diff] [blame] | 1919 | "src/f32-prelu/gen/sse-2x4.c", |
| 1920 | "src/f32-prelu/gen/sse-2x8.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1921 | "src/f32-relu/gen/sse-x4.c", |
| 1922 | "src/f32-relu/gen/sse-x8.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1923 | "src/f32-rmax/sse.c", |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1924 | "src/f32-spmm/gen/4x1-minmax-sse.c", |
| 1925 | "src/f32-spmm/gen/8x1-minmax-sse.c", |
| Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 1926 | "src/f32-spmm/gen/16x1-minmax-sse.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1927 | "src/f32-vbinary/gen/vadd-minmax-sse-x4.c", |
| 1928 | "src/f32-vbinary/gen/vadd-minmax-sse-x8.c", |
| 1929 | "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c", |
| 1930 | "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c", |
| 1931 | "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c", |
| 1932 | "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c", |
| 1933 | "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c", |
| 1934 | "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c", |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1935 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 1936 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 1937 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 1938 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 1939 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 1940 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 1941 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 1942 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1943 | "src/f32-vbinary/gen/vmul-minmax-sse-x4.c", |
| 1944 | "src/f32-vbinary/gen/vmul-minmax-sse-x8.c", |
| 1945 | "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c", |
| 1946 | "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c", |
| 1947 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c", |
| 1948 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c", |
| 1949 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c", |
| 1950 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c", |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1951 | "src/f32-vbinary/gen/vsqrdiff-sse-x4.c", |
| 1952 | "src/f32-vbinary/gen/vsqrdiff-sse-x8.c", |
| 1953 | "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c", |
| 1954 | "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1955 | "src/f32-vbinary/gen/vsub-minmax-sse-x4.c", |
| 1956 | "src/f32-vbinary/gen/vsub-minmax-sse-x8.c", |
| 1957 | "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c", |
| 1958 | "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1959 | "src/f32-vlrelu/gen/vlrelu-sse-x4.c", |
| 1960 | "src/f32-vlrelu/gen/vlrelu-sse-x8.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1961 | "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c", |
| 1962 | "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1963 | "src/f32-vsqrt/gen/sse-sqrt-x4.c", |
| 1964 | "src/f32-vsqrt/gen/sse-sqrt-x8.c", |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1965 | "src/f32-vunary/gen/vabs-sse-x4.c", |
| 1966 | "src/f32-vunary/gen/vabs-sse-x8.c", |
| 1967 | "src/f32-vunary/gen/vneg-sse-x4.c", |
| 1968 | "src/f32-vunary/gen/vneg-sse-x8.c", |
| 1969 | "src/f32-vunary/gen/vsqr-sse-x4.c", |
| 1970 | "src/f32-vunary/gen/vsqr-sse-x8.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1971 | "src/math/roundd-sse-addsub.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1972 | "src/math/roundne-sse-addsub.c", |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1973 | "src/math/roundu-sse-addsub.c", |
| Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1974 | "src/math/roundz-sse-addsub.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1975 | "src/math/sqrt-sse-hh1mac.c", |
| 1976 | "src/math/sqrt-sse-nr1mac.c", |
| 1977 | "src/math/sqrt-sse-nr2mac.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1978 | "src/x32-fill/sse.c", |
| 1979 | "src/x32-packx/x4-sse.c", |
| 1980 | "src/x32-pad/sse.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1981 | ] |
| 1982 | |
| 1983 | SSE2_UKERNELS = [ |
| Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1984 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1985 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1986 | "src/f32-argmaxpool/9x-sse2-c4.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1987 | "src/f32-prelu/gen/sse2-2x4.c", |
| 1988 | "src/f32-prelu/gen/sse2-2x8.c", |
| Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1989 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1990 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1991 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1992 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 1993 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1994 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1995 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 1996 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 1997 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1998 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 1999 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2000 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 2001 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c", |
| 2002 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c", |
| 2003 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c", |
| 2004 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c", |
| 2005 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c", |
| 2006 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2007 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2008 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2009 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2010 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2011 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 2012 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2013 | "src/f32-vlrelu/gen/vlrelu-sse2-x4.c", |
| 2014 | "src/f32-vlrelu/gen/vlrelu-sse2-x8.c", |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2015 | "src/f32-vrnd/gen/vrndd-sse2-x4.c", |
| 2016 | "src/f32-vrnd/gen/vrndd-sse2-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2017 | "src/f32-vrnd/gen/vrndne-sse2-x4.c", |
| 2018 | "src/f32-vrnd/gen/vrndne-sse2-x8.c", |
| 2019 | "src/f32-vrnd/gen/vrndu-sse2-x4.c", |
| 2020 | "src/f32-vrnd/gen/vrndu-sse2-x8.c", |
| 2021 | "src/f32-vrnd/gen/vrndz-sse2-x4.c", |
| 2022 | "src/f32-vrnd/gen/vrndz-sse2-x8.c", |
| 2023 | "src/math/exp-sse2-lut64-p2.c", |
| 2024 | "src/math/exp-sse2-p5.c", |
| 2025 | "src/math/expminus-sse2-p5.c", |
| 2026 | "src/math/roundd-sse2-cvt.c", |
| 2027 | "src/math/roundne-sse2-cvt.c", |
| 2028 | "src/math/roundu-sse2-cvt.c", |
| 2029 | "src/math/roundz-sse2-cvt.c", |
| 2030 | "src/math/sigmoid-sse2-rr2-lut64-p2-div.c", |
| 2031 | "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c", |
| 2032 | "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c", |
| 2033 | "src/math/sigmoid-sse2-rr2-p5-div.c", |
| 2034 | "src/math/sigmoid-sse2-rr2-p5-nr1.c", |
| 2035 | "src/math/sigmoid-sse2-rr2-p5-nr2.c", |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2036 | "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c", |
| 2037 | "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c", |
| 2038 | "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c", |
| Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2039 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c", |
| 2040 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c", |
| 2041 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2042 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c", |
| 2043 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c", |
| 2044 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2045 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2046 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2047 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2048 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2049 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2050 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2051 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c", |
| 2052 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2053 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2054 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c", |
| 2055 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2056 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2057 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c", |
| 2058 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c", |
| 2059 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2060 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2061 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2062 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2063 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2064 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2065 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2066 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c", |
| Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2067 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2068 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c", |
| 2069 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c", |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2070 | "src/qs8-requantization/fp32-sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2071 | "src/qs8-requantization/precise-sse2.c", |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2072 | "src/qs8-requantization/q31-sse2.c", |
| Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 2073 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c", |
| 2074 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c", |
| 2075 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c", |
| 2076 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c", |
| Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 2077 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c", |
| 2078 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c", |
| 2079 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c", |
| 2080 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2081 | "src/qu8-avgpool/9p8x-minmax-sse2-c8.c", |
| 2082 | "src/qu8-avgpool/9x-minmax-sse2-c8.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2083 | "src/qu8-dwconv/up8x9-minmax-sse2.c", |
| 2084 | "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c", |
| 2085 | "src/qu8-gavgpool/7x-minmax-sse2-c8.c", |
| 2086 | "src/qu8-gemm/2x4c8-minmax-sse2.c", |
| 2087 | "src/qu8-gemm/4x4c2-minmax-sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2088 | "src/qu8-igemm/4x4c2-minmax-sse2.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2089 | "src/qu8-requantization/fp32-sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2090 | "src/qu8-requantization/precise-sse2.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2091 | "src/qu8-requantization/q31-sse2.c", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2092 | "src/qu8-vadd/minmax-sse2.c", |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2093 | "src/u8-clamp/sse2-x64.c", |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2094 | "src/u8-maxpool/9p8x-minmax-sse2-c16.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2095 | "src/u8-rmax/sse2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2096 | "src/x8-zip/x2-sse2.c", |
| 2097 | "src/x8-zip/x3-sse2.c", |
| 2098 | "src/x8-zip/x4-sse2.c", |
| 2099 | "src/x8-zip/xm-sse2.c", |
| Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 2100 | "src/x32-unpool/sse2.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2101 | "src/x32-zip/x2-sse2.c", |
| 2102 | "src/x32-zip/x3-sse2.c", |
| 2103 | "src/x32-zip/x4-sse2.c", |
| 2104 | "src/x32-zip/xm-sse2.c", |
| Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2105 | ] |
| 2106 | |
| 2107 | SSSE3_UKERNELS = [ |
| Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame] | 2108 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c", |
| 2109 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c", |
| 2110 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c", |
| 2111 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c", |
| 2112 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c", |
| 2113 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c", |
| 2114 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c", |
| 2115 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c", |
| 2116 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c", |
| 2117 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c", |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2118 | "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c", |
| 2119 | "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c", |
| 2120 | "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c", |
| Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2121 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c", |
| 2122 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c", |
| 2123 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2124 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c", |
| 2125 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c", |
| 2126 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2127 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2128 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2129 | "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2130 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2131 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2132 | "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2133 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c", |
| 2134 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2135 | "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2136 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c", |
| 2137 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2138 | "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2139 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c", |
| 2140 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c", |
| 2141 | "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2142 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2143 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2144 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2145 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2146 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2147 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2148 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c", |
| Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2149 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2150 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c", |
| 2151 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c", |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 2152 | "src/qs8-requantization/precise-ssse3.c", |
| 2153 | "src/qs8-requantization/q31-ssse3.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2154 | "src/qu8-requantization/precise-ssse3.c", |
| 2155 | "src/qu8-requantization/q31-ssse3.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2156 | ] |
| 2157 | |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2158 | SSE41_UKERNELS = [ |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2159 | "src/f32-prelu/gen/sse41-2x4.c", |
| 2160 | "src/f32-prelu/gen/sse41-2x8.c", |
| Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 2161 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c", |
| 2162 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c", |
| 2163 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c", |
| 2164 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c", |
| 2165 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c", |
| 2166 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2167 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2168 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2169 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
| Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2170 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2171 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 2172 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2173 | "src/f32-vlrelu/gen/vlrelu-sse41-x4.c", |
| 2174 | "src/f32-vlrelu/gen/vlrelu-sse41-x8.c", |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2175 | "src/f32-vrnd/gen/vrndd-sse41-x4.c", |
| 2176 | "src/f32-vrnd/gen/vrndd-sse41-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2177 | "src/f32-vrnd/gen/vrndne-sse41-x4.c", |
| 2178 | "src/f32-vrnd/gen/vrndne-sse41-x8.c", |
| 2179 | "src/f32-vrnd/gen/vrndu-sse41-x4.c", |
| 2180 | "src/f32-vrnd/gen/vrndu-sse41-x8.c", |
| 2181 | "src/f32-vrnd/gen/vrndz-sse41-x4.c", |
| 2182 | "src/f32-vrnd/gen/vrndz-sse41-x8.c", |
| 2183 | "src/math/roundd-sse41.c", |
| 2184 | "src/math/roundne-sse41.c", |
| 2185 | "src/math/roundu-sse41.c", |
| 2186 | "src/math/roundz-sse41.c", |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2187 | "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c", |
| 2188 | "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c", |
| 2189 | "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c", |
| Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2190 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c", |
| 2191 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c", |
| 2192 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2193 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c", |
| 2194 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c", |
| 2195 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2196 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2197 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2198 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2199 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2200 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2201 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2202 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c", |
| 2203 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2204 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2205 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c", |
| 2206 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2207 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2208 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c", |
| 2209 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c", |
| 2210 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c", |
| Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2211 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c", |
| Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2212 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2213 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2214 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2215 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c", |
| Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2216 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2217 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c", |
| Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2218 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2219 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c", |
| 2220 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c", |
| Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 2221 | "src/qs8-requantization/fp32-sse4.c", |
| 2222 | "src/qs8-requantization/precise-sse4.c", |
| 2223 | "src/qs8-requantization/q31-sse4.c", |
| Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 2224 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c", |
| 2225 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c", |
| 2226 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c", |
| 2227 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c", |
| Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2228 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c", |
| 2229 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c", |
| 2230 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c", |
| 2231 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c", |
| Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 2232 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c", |
| 2233 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c", |
| 2234 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c", |
| 2235 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c", |
| Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2236 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c", |
| 2237 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c", |
| 2238 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c", |
| 2239 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2240 | "src/qu8-requantization/precise-sse4.c", |
| 2241 | "src/qu8-requantization/q31-sse4.c", |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2242 | ] |
| 2243 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2244 | AVX_UKERNELS = [ |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2245 | "src/f32-clamp/gen/avx-x8.c", |
| 2246 | "src/f32-clamp/gen/avx-x16.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2247 | "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c", |
| 2248 | "src/f32-dwconv/gen/up8x4-minmax-avx.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2249 | "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c", |
| 2250 | "src/f32-dwconv/gen/up8x9-minmax-avx.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2251 | "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c", |
| 2252 | "src/f32-dwconv/gen/up8x25-minmax-avx.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2253 | "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c", |
| 2254 | "src/f32-dwconv/gen/up16x4-minmax-avx.c", |
| 2255 | "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c", |
| 2256 | "src/f32-dwconv/gen/up16x9-minmax-avx.c", |
| 2257 | "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c", |
| 2258 | "src/f32-dwconv/gen/up16x25-minmax-avx.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2259 | "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2260 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c", |
| 2261 | "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2262 | "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2263 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2264 | "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2265 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2266 | "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c", |
| 2267 | "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c", |
| 2268 | "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c", |
| 2269 | "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c", |
| 2270 | "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c", |
| 2271 | "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c", |
| 2272 | "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c", |
| 2273 | "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c", |
| 2274 | "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c", |
| 2275 | "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c", |
| 2276 | "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2277 | "src/f32-hswish/gen/hswish-avx-x8.c", |
| 2278 | "src/f32-hswish/gen/hswish-avx-x16.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2279 | "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2280 | "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c", |
| 2281 | "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2282 | "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2283 | "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2284 | "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2285 | "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2286 | "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c", |
| 2287 | "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c", |
| Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2288 | "src/f32-prelu/gen/avx-2x8.c", |
| 2289 | "src/f32-prelu/gen/avx-2x16.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2290 | "src/f32-relu/gen/avx-x8.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2291 | "src/f32-relu/gen/avx-x16.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2292 | "src/f32-rmax/avx.c", |
| T.J. Alumbaugh | dc2b29c | 2020-10-14 13:56:08 -0700 | [diff] [blame] | 2293 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c", |
| 2294 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c", |
| 2295 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c", |
| 2296 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c", |
| 2297 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c", |
| 2298 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c", |
| 2299 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c", |
| 2300 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c", |
| 2301 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c", |
| 2302 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c", |
| 2303 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c", |
| 2304 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c", |
| 2305 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c", |
| 2306 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c", |
| 2307 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c", |
| 2308 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c", |
| 2309 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c", |
| 2310 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c", |
| 2311 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c", |
| 2312 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2313 | "src/f32-vbinary/gen/vadd-minmax-avx-x8.c", |
| 2314 | "src/f32-vbinary/gen/vadd-minmax-avx-x16.c", |
| 2315 | "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c", |
| 2316 | "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c", |
| 2317 | "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c", |
| 2318 | "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c", |
| 2319 | "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c", |
| 2320 | "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c", |
| Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2321 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 2322 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 2323 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 2324 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 2325 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 2326 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 2327 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 2328 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2329 | "src/f32-vbinary/gen/vmul-minmax-avx-x8.c", |
| 2330 | "src/f32-vbinary/gen/vmul-minmax-avx-x16.c", |
| 2331 | "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c", |
| 2332 | "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c", |
| 2333 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c", |
| 2334 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c", |
| 2335 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c", |
| 2336 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c", |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2337 | "src/f32-vbinary/gen/vsqrdiff-avx-x8.c", |
| 2338 | "src/f32-vbinary/gen/vsqrdiff-avx-x16.c", |
| 2339 | "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c", |
| 2340 | "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2341 | "src/f32-vbinary/gen/vsub-minmax-avx-x8.c", |
| 2342 | "src/f32-vbinary/gen/vsub-minmax-avx-x16.c", |
| 2343 | "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c", |
| 2344 | "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2345 | "src/f32-vlrelu/gen/vlrelu-avx-x8.c", |
| 2346 | "src/f32-vlrelu/gen/vlrelu-avx-x16.c", |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2347 | "src/f32-vrnd/gen/vrndd-avx-x8.c", |
| 2348 | "src/f32-vrnd/gen/vrndd-avx-x16.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2349 | "src/f32-vrnd/gen/vrndne-avx-x8.c", |
| 2350 | "src/f32-vrnd/gen/vrndne-avx-x16.c", |
| 2351 | "src/f32-vrnd/gen/vrndu-avx-x8.c", |
| 2352 | "src/f32-vrnd/gen/vrndu-avx-x16.c", |
| 2353 | "src/f32-vrnd/gen/vrndz-avx-x8.c", |
| 2354 | "src/f32-vrnd/gen/vrndz-avx-x16.c", |
| 2355 | "src/f32-vscale/avx-unroll32.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2356 | "src/f32-vsqrt/gen/avx-sqrt-x8.c", |
| 2357 | "src/f32-vsqrt/gen/avx-sqrt-x16.c", |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2358 | "src/f32-vunary/gen/vabs-avx-x8.c", |
| 2359 | "src/f32-vunary/gen/vabs-avx-x16.c", |
| 2360 | "src/f32-vunary/gen/vneg-avx-x8.c", |
| 2361 | "src/f32-vunary/gen/vneg-avx-x16.c", |
| 2362 | "src/f32-vunary/gen/vsqr-avx-x8.c", |
| 2363 | "src/f32-vunary/gen/vsqr-avx-x16.c", |
| Marat Dukhan | d08abcb | 2020-09-21 00:26:25 -0700 | [diff] [blame] | 2364 | "src/math/exp-avx-p5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2365 | "src/math/sigmoid-avx-rr2-lut64-p2-div.c", |
| Marat Dukhan | 2e4727b | 2020-09-24 00:48:39 -0700 | [diff] [blame] | 2366 | "src/math/sigmoid-avx-rr2-p5-div.c", |
| Marat Dukhan | 39ab16f | 2020-10-16 14:30:35 -0700 | [diff] [blame] | 2367 | "src/math/sigmoid-avx-rr2-p5-nr1.c", |
| 2368 | "src/math/sigmoid-avx-rr2-p5-nr2.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2369 | ] |
| 2370 | |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2371 | XOP_UKERNELS = [ |
| 2372 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2373 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2374 | "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2375 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2376 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2377 | "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2378 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c", |
| 2379 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2380 | "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2381 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c", |
| 2382 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2383 | "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2384 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c", |
| 2385 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c", |
| 2386 | "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2387 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2388 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2389 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2390 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2391 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2392 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2393 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c", |
| Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2394 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2395 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c", |
| 2396 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c", |
| Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2397 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c", |
| 2398 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c", |
| 2399 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c", |
| 2400 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c", |
| 2401 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c", |
| 2402 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c", |
| 2403 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c", |
| 2404 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2405 | ] |
| 2406 | |
| Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2407 | FMA3_UKERNELS = [ |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2408 | "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c", |
| 2409 | "src/f32-dwconv/gen/up8x4-minmax-fma3.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2410 | "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c", |
| 2411 | "src/f32-dwconv/gen/up8x9-minmax-fma3.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2412 | "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c", |
| 2413 | "src/f32-dwconv/gen/up8x25-minmax-fma3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2414 | "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c", |
| 2415 | "src/f32-dwconv/gen/up16x4-minmax-fma3.c", |
| 2416 | "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c", |
| 2417 | "src/f32-dwconv/gen/up16x9-minmax-fma3.c", |
| 2418 | "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c", |
| 2419 | "src/f32-dwconv/gen/up16x25-minmax-fma3.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2420 | "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2421 | "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c", |
| 2422 | "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c", |
| 2423 | "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c", |
| 2424 | "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2425 | "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2426 | "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c", |
| 2427 | "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2428 | "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2429 | "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c", |
| 2430 | "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2431 | "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c", |
| 2432 | "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c", |
| 2433 | "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2434 | "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c", |
| 2435 | "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c", |
| 2436 | "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 2437 | "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c", |
| 2438 | "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 2439 | "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c", |
| 2440 | "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c", |
| 2441 | "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 2442 | "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c", |
| 2443 | "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c", |
| 2444 | "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| 2445 | "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c", |
| 2446 | "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c", |
| 2447 | "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2448 | "src/f32-hswish/gen/hswish-fma3-x8.c", |
| 2449 | "src/f32-hswish/gen/hswish-fma3-x16.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2450 | "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2451 | "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c", |
| 2452 | "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 2453 | "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c", |
| 2454 | "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2455 | "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2456 | "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c", |
| 2457 | "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2458 | "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2459 | "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c", |
| 2460 | "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2461 | "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c", |
| 2462 | "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c", |
| 2463 | "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2464 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c", |
| 2465 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c", |
| 2466 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c", |
| 2467 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c", |
| 2468 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c", |
| 2469 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c", |
| 2470 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c", |
| 2471 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2472 | "src/math/sqrt-fma3-nr1fma.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2473 | "src/math/sqrt-fma3-nr1fma1adj.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2474 | "src/math/sqrt-fma3-nr2fma.c", |
| Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2475 | ] |
| 2476 | |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2477 | AVX2_UKERNELS = [ |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2478 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 2479 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2480 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2481 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2482 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2483 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 2484 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2485 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2486 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 2487 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 2488 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2489 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2490 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 2491 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2492 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2493 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2494 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2495 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 2496 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2497 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2498 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 2499 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 2500 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2501 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2502 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 2503 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2504 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2505 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2506 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2507 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 2508 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2509 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2510 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 2511 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 2512 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2513 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 2514 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 2515 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 2516 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 2517 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 2518 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 2519 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 2520 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 2521 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 2522 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 2523 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 2524 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 2525 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 2526 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 2527 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 2528 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 2529 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 2530 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 2531 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 2532 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 2533 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 2534 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 2535 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 2536 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 2537 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 2538 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 2539 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 2540 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 2541 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 2542 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 2543 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2544 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 2545 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 2546 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 2547 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 2548 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 2549 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 2550 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 2551 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 2552 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 2553 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 2554 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 2555 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 2556 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 2557 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 2558 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 2559 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 2560 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 2561 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 2562 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 2563 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 2564 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 2565 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 2566 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 2567 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2568 | "src/math/exp-avx2-lut8-p3-perm.c", |
| 2569 | "src/math/exp-avx2-lut8-p4-perm.c", |
| 2570 | "src/math/exp-avx2-p5.c", |
| 2571 | "src/math/expminus-avx2-p5.c", |
| 2572 | "src/math/extexp-avx2-p5.c", |
| 2573 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c", |
| 2574 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c", |
| 2575 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c", |
| 2576 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c", |
| 2577 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
| 2578 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
| 2579 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 2580 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c", |
| 2581 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c", |
| 2582 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c", |
| 2583 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c", |
| 2584 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 2585 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 2586 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2587 | "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2588 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c", |
| Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2589 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c", |
| 2590 | "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2591 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c", |
| Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2592 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c", |
| Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 2593 | "src/qs8-gemm/gen/1x8c8-minmax-avx2.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2594 | "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2595 | "src/qs8-gemm/gen/2x8c8-minmax-avx2.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2596 | "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2597 | "src/qs8-gemm/gen/3x8c8-minmax-avx2.c", |
| Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2598 | "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c", |
| Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 2599 | "src/qs8-igemm/gen/1x8c8-minmax-avx2.c", |
| 2600 | "src/qs8-igemm/gen/2x8c8-minmax-avx2.c", |
| 2601 | "src/qs8-igemm/gen/3x8c8-minmax-avx2.c", |
| Marat Dukhan | e6dc0b6 | 2020-09-08 23:57:14 -0700 | [diff] [blame] | 2602 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2603 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2604 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2605 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c", |
| 2606 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2607 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2608 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2609 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c", |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2610 | ] |
| 2611 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2612 | AVX512F_UKERNELS = [ |
| Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2613 | "src/f32-clamp/gen/avx512f-x16.c", |
| 2614 | "src/f32-clamp/gen/avx512f-x32.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2615 | "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c", |
| 2616 | "src/f32-dwconv/gen/up16x4-minmax-avx512f.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2617 | "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c", |
| 2618 | "src/f32-dwconv/gen/up16x9-minmax-avx512f.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2619 | "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c", |
| 2620 | "src/f32-dwconv/gen/up16x25-minmax-avx512f.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2621 | "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c", |
| 2622 | "src/f32-dwconv/gen/up32x4-minmax-avx512f.c", |
| 2623 | "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c", |
| 2624 | "src/f32-dwconv/gen/up32x9-minmax-avx512f.c", |
| 2625 | "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c", |
| 2626 | "src/f32-dwconv/gen/up32x25-minmax-avx512f.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2627 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c", |
| 2628 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c", |
| 2629 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c", |
| 2630 | "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c", |
| 2631 | "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c", |
| 2632 | "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2633 | "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2634 | "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2635 | "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2636 | "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2637 | "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2638 | "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c", |
| Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2639 | "src/f32-hswish/gen/hswish-avx512f-x16.c", |
| 2640 | "src/f32-hswish/gen/hswish-avx512f-x32.c", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2641 | "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2642 | "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2643 | "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2644 | "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2645 | "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2646 | "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c", |
| Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2647 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 2648 | "src/f32-prelu/gen/avx512f-2x32.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2649 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2650 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2651 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2652 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2653 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2654 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2655 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2656 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2657 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2658 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2659 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2660 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2661 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2662 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2663 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2664 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2665 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2666 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2667 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2668 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2669 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2670 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2671 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2672 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2673 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2674 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2675 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2676 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2677 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2678 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2679 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2680 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2681 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2682 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2683 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2684 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2685 | "src/f32-relu/gen/avx512f-x16.c", |
| 2686 | "src/f32-relu/gen/avx512f-x32.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2687 | "src/f32-rmax/avx512f.c", |
| Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 2688 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c", |
| 2689 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c", |
| 2690 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c", |
| 2691 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c", |
| 2692 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c", |
| 2693 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c", |
| 2694 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c", |
| 2695 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c", |
| 2696 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c", |
| 2697 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c", |
| 2698 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c", |
| 2699 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c", |
| 2700 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c", |
| 2701 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c", |
| 2702 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c", |
| 2703 | "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] | 2704 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c", |
| 2705 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c", |
| 2706 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c", |
| 2707 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c", |
| 2708 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c", |
| 2709 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c", |
| 2710 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c", |
| 2711 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c", |
| 2712 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c", |
| 2713 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c", |
| 2714 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c", |
| 2715 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c", |
| 2716 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c", |
| 2717 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c", |
| 2718 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c", |
| 2719 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c", |
| Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 2720 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c", |
| 2721 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c", |
| 2722 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c", |
| 2723 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c", |
| 2724 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c", |
| 2725 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c", |
| 2726 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c", |
| 2727 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c", |
| 2728 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c", |
| 2729 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c", |
| 2730 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c", |
| 2731 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c", |
| 2732 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c", |
| 2733 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c", |
| 2734 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c", |
| 2735 | "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] | 2736 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c", |
| 2737 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c", |
| 2738 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c", |
| 2739 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c", |
| 2740 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c", |
| 2741 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c", |
| 2742 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c", |
| 2743 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c", |
| Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2744 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 2745 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 2746 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 2747 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 2748 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 2749 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 2750 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 2751 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2752 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c", |
| 2753 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c", |
| 2754 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c", |
| 2755 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c", |
| 2756 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c", |
| 2757 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c", |
| 2758 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c", |
| 2759 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c", |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2760 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c", |
| 2761 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c", |
| 2762 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c", |
| 2763 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c", |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2764 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c", |
| 2765 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c", |
| 2766 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c", |
| 2767 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c", |
| Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2768 | "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c", |
| 2769 | "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2770 | "src/f32-vrnd/gen/vrndd-avx512f-x16.c", |
| 2771 | "src/f32-vrnd/gen/vrndd-avx512f-x32.c", |
| 2772 | "src/f32-vrnd/gen/vrndne-avx512f-x16.c", |
| 2773 | "src/f32-vrnd/gen/vrndne-avx512f-x32.c", |
| 2774 | "src/f32-vrnd/gen/vrndu-avx512f-x16.c", |
| 2775 | "src/f32-vrnd/gen/vrndu-avx512f-x32.c", |
| 2776 | "src/f32-vrnd/gen/vrndz-avx512f-x16.c", |
| 2777 | "src/f32-vrnd/gen/vrndz-avx512f-x32.c", |
| Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2778 | "src/f32-vscale/avx512f-unroll64.c", |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2779 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 2780 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 2781 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 2782 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 2783 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 2784 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 2785 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 2786 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2787 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2788 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2789 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 2790 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2791 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 2792 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 2793 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 2794 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 2795 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 2796 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 2797 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 2798 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 2799 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 2800 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 2801 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 2802 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2803 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c", |
| 2804 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c", |
| 2805 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c", |
| 2806 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c", |
| 2807 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c", |
| 2808 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c", |
| 2809 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c", |
| 2810 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c", |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2811 | "src/f32-vunary/gen/vabs-avx512f-x16.c", |
| 2812 | "src/f32-vunary/gen/vabs-avx512f-x32.c", |
| 2813 | "src/f32-vunary/gen/vneg-avx512f-x16.c", |
| 2814 | "src/f32-vunary/gen/vneg-avx512f-x32.c", |
| 2815 | "src/f32-vunary/gen/vsqr-avx512f-x16.c", |
| 2816 | "src/f32-vunary/gen/vsqr-avx512f-x32.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2817 | "src/math/exp-avx512f-lut16-p3-perm-scalef.c", |
| 2818 | "src/math/exp-avx512f-lut16-p3-perm.c", |
| 2819 | "src/math/exp-avx512f-lut32-p2-perm2-scalef.c", |
| 2820 | "src/math/exp-avx512f-lut32-p2-perm2.c", |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2821 | "src/math/exp-avx512f-p5-scalef.c", |
| 2822 | "src/math/exp-avx512f-p5.c", |
| Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2823 | "src/math/extexp-avx512f-p5.c", |
| Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2824 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c", |
| Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2825 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2826 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c", |
| Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2827 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c", |
| Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2828 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2829 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c", |
| Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2830 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c", |
| Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2831 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2832 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c", |
| 2833 | "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c", |
| 2834 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c", |
| 2835 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c", |
| 2836 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c", |
| 2837 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c", |
| 2838 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c", |
| 2839 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c", |
| 2840 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c", |
| 2841 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c", |
| Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2842 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c", |
| Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 2843 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2844 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c", |
| 2845 | "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c", |
| 2846 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c", |
| 2847 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2848 | "src/math/sqrt-avx512f-nr1fma.c", |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2849 | "src/math/sqrt-avx512f-nr1fma1adj.c", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2850 | "src/math/sqrt-avx512f-nr2fma.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2851 | ] |
| 2852 | |
| Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 2853 | AVX512SKX_UKERNELS = [ |
| Marat Dukhan | 2ffc5e6 | 2020-09-06 22:33:38 -0700 | [diff] [blame] | 2854 | "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c", |
| 2855 | "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c", |
| Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 2856 | "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c", |
| 2857 | "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c", |
| 2858 | "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c", |
| 2859 | "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c", |
| 2860 | "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c", |
| 2861 | "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c", |
| 2862 | "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c", |
| 2863 | "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c", |
| 2864 | ] |
| 2865 | |
| Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 2866 | WASM32_ASM_UKERNELS = [ |
| 2867 | "src/f32-relu/wasm_shr_x1.S", |
| 2868 | "src/f32-relu/wasm_shr_x2.S", |
| 2869 | "src/f32-relu/wasm_shr_x4.S", |
| 2870 | ] |
| 2871 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2872 | AARCH32_ASM_UKERNELS = [ |
| Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 2873 | "src/f32-gemm/4x4-aarch32-vfp-ld64.S", |
| Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 2874 | "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2875 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2876 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
| Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 2877 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2878 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 2879 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 2880 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2881 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2882 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
| Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 2883 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
| 2884 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 2885 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| 2886 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2887 | ] |
| 2888 | |
| 2889 | AARCH64_ASM_UKERNELS = [ |
| Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2890 | "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2891 | "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
| Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2892 | "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2893 | "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
| Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2894 | "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2895 | "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
| Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 2896 | "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2897 | "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2898 | "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2899 | "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2900 | "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2901 | "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2902 | "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2903 | "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2904 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S", |
| 2905 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2906 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2907 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2908 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2909 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2910 | "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2911 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2912 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2913 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2914 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2915 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2916 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2917 | "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2918 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2919 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2920 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2921 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2922 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2923 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2924 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2925 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2926 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2927 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2928 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2929 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2930 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S", |
| 2931 | "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2932 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2933 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2934 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2935 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2936 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S", |
| 2937 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S", |
| 2938 | "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2939 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2940 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2941 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2942 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2943 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2944 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2945 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2946 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S", |
| 2947 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2948 | "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2949 | "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2950 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2951 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2952 | "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2953 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2954 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2955 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2956 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2957 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2958 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2959 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2960 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2961 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2962 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2963 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
| Frank Barchard | f1fd89e | 2020-10-06 23:27:11 -0700 | [diff] [blame] | 2964 | "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S", |
| Frank Barchard | f2742c4 | 2020-10-12 21:16:46 -0700 | [diff] [blame] | 2965 | "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S", |
| Frank Barchard | a5237a5 | 2020-10-06 16:18:46 -0700 | [diff] [blame] | 2966 | "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2967 | ] |
| 2968 | |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2969 | INTERNAL_MICROKERNEL_HDRS = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2970 | "src/xnnpack/argmaxpool.h", |
| 2971 | "src/xnnpack/avgpool.h", |
| 2972 | "src/xnnpack/clamp.h", |
| 2973 | "src/xnnpack/common.h", |
| 2974 | "src/xnnpack/conv.h", |
| 2975 | "src/xnnpack/dwconv.h", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 2976 | "src/xnnpack/fill.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2977 | "src/xnnpack/gavgpool.h", |
| 2978 | "src/xnnpack/gemm.h", |
| 2979 | "src/xnnpack/hswish.h", |
| Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2980 | "src/xnnpack/ibilinear.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2981 | "src/xnnpack/igemm.h", |
| Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 2982 | "src/xnnpack/intrinsics-polyfill.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2983 | "src/xnnpack/lut.h", |
| 2984 | "src/xnnpack/math.h", |
| 2985 | "src/xnnpack/maxpool.h", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2986 | "src/xnnpack/memory.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2987 | "src/xnnpack/packx.h", |
| 2988 | "src/xnnpack/pad.h", |
| 2989 | "src/xnnpack/params.h", |
| 2990 | "src/xnnpack/pavgpool.h", |
| 2991 | "src/xnnpack/ppmm.h", |
| 2992 | "src/xnnpack/prelu.h", |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2993 | "src/xnnpack/raddexpminusmax.h", |
| Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2994 | "src/xnnpack/raddextexp.h", |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2995 | "src/xnnpack/raddstoreexpminusmax.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2996 | "src/xnnpack/rmax.h", |
| 2997 | "src/xnnpack/scalar-utils.h", |
| 2998 | "src/xnnpack/spmm.h", |
| 2999 | "src/xnnpack/unpool.h", |
| 3000 | "src/xnnpack/vadd.h", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 3001 | "src/xnnpack/vbinary.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3002 | "src/xnnpack/vmulcaddc.h", |
| Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 3003 | "src/xnnpack/vscale.h", |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 3004 | "src/xnnpack/vscaleexpminusmax.h", |
| Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 3005 | "src/xnnpack/vscaleextexp.h", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 3006 | "src/xnnpack/vunary.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3007 | "src/xnnpack/zip.h", |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3008 | ] |
| 3009 | |
| 3010 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3011 | "include/xnnpack.h", |
| 3012 | "src/xnnpack/allocator.h", |
| 3013 | "src/xnnpack/compute.h", |
| 3014 | "src/xnnpack/im2col.h", |
| 3015 | "src/xnnpack/indirection.h", |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3016 | "src/xnnpack/math-stubs.h", |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3017 | "src/xnnpack/memory-planner.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3018 | "src/xnnpack/operator.h", |
| 3019 | "src/xnnpack/pack.h", |
| Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 3020 | "src/xnnpack/params-init.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3021 | "src/xnnpack/requantization-stubs.h", |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3022 | "src/xnnpack/requantization.h", |
| Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 3023 | "src/xnnpack/subgraph.h", |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3024 | ] |
| 3025 | |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3026 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3027 | "src/xnnpack/math-stubs.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3028 | ] |
| 3029 | |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3030 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 3031 | "src/xnnpack/params-init.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3032 | "include/xnnpack.h", |
| 3033 | ] |
| 3034 | |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 3035 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3036 | "src/xnnpack/isa-checks.h", |
| Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 3037 | "src/xnnpack/params-init.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3038 | "src/xnnpack/requantization.h", |
| 3039 | "include/xnnpack.h", |
| 3040 | ] |
| 3041 | |
| 3042 | OPERATOR_TEST_PARAMS_HDRS = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3043 | "src/xnnpack/common.h", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3044 | "src/xnnpack/params.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3045 | ] |
| 3046 | |
| 3047 | WEIGHTS_PACK_HDRS = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3048 | "src/xnnpack/compute.h", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 3049 | "src/xnnpack/operator.h", |
| 3050 | "src/xnnpack/pack.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3051 | ] |
| 3052 | |
| Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 3053 | LOGGING_COPTS = select({ |
| 3054 | # No logging in optimized mode |
| 3055 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 3056 | # Full logging in debug mode |
| 3057 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 3058 | # Error-only logging in default (fastbuild) mode |
| 3059 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 3060 | }) |
| 3061 | |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3062 | LOGGING_SRCS = select({ |
| 3063 | # No logging in optimized mode |
| 3064 | ":optimized_build": [], |
| 3065 | "//conditions:default": [ |
| 3066 | "src/operator-strings.c", |
| 3067 | "src/subgraph-strings.c", |
| 3068 | ], |
| 3069 | }) |
| 3070 | |
| Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 3071 | LOGGING_HDRS = [ |
| 3072 | "src/xnnpack/log.h", |
| 3073 | ] |
| 3074 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3075 | xnnpack_cc_library( |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3076 | name = "tables", |
| 3077 | srcs = TABLE_SRCS, |
| 3078 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3079 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3080 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3081 | ) |
| 3082 | |
| 3083 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3084 | name = "scalar_ukernels", |
| 3085 | srcs = SCALAR_UKERNELS, |
| 3086 | hdrs = INTERNAL_HDRS, |
| 3087 | aarch32_copts = ["-marm"], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3088 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3089 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3090 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3091 | ":tables", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3092 | "@FP16", |
| 3093 | "@FXdiv", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3094 | "@pthreadpool", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3095 | ], |
| 3096 | ) |
| 3097 | |
| 3098 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3099 | name = "scalar_ukernels_test_mode", |
| 3100 | srcs = SCALAR_UKERNELS, |
| 3101 | hdrs = INTERNAL_HDRS, |
| 3102 | aarch32_copts = ["-marm"], |
| 3103 | copts = [ |
| 3104 | "-UNDEBUG", |
| 3105 | "-DXNN_TEST_MODE=1", |
| 3106 | ], |
| 3107 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3108 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3109 | deps = [ |
| 3110 | ":tables", |
| 3111 | "@FP16", |
| 3112 | "@FXdiv", |
| 3113 | "@pthreadpool", |
| 3114 | ], |
| 3115 | ) |
| 3116 | |
| 3117 | xnnpack_cc_library( |
| Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3118 | name = "wasm_ukernels", |
| 3119 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3120 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3121 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3122 | wasm_srcs = WASM_UKERNELS, |
| Marat Dukhan | 1c6e389 | 2020-06-25 23:56:10 -0700 | [diff] [blame] | 3123 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
| Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3124 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3125 | ":tables", |
| Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 3126 | "@FP16", |
| 3127 | "@FXdiv", |
| 3128 | "@pthreadpool", |
| 3129 | ], |
| 3130 | ) |
| 3131 | |
| 3132 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3133 | name = "wasm_ukernels_test_mode", |
| 3134 | hdrs = INTERNAL_HDRS, |
| 3135 | copts = [ |
| 3136 | "-UNDEBUG", |
| 3137 | "-DXNN_TEST_MODE=1", |
| 3138 | ], |
| 3139 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3140 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3141 | wasm_srcs = WASM_UKERNELS, |
| Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 3142 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3143 | deps = [ |
| 3144 | ":tables", |
| 3145 | "@FP16", |
| 3146 | "@FXdiv", |
| 3147 | "@pthreadpool", |
| 3148 | ], |
| 3149 | ) |
| 3150 | |
| 3151 | xnnpack_cc_library( |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3152 | name = "psimd_fastmath_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3153 | hdrs = INTERNAL_HDRS, |
| 3154 | aarch32_copts = [ |
| 3155 | "-marm", |
| 3156 | "-mfpu=neon", |
| 3157 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3158 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3159 | gcc_x86_copts = ["-msse2"], |
| 3160 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3161 | optimized_copts = [ |
| 3162 | "-O3", |
| 3163 | "-ffast-math", |
| 3164 | ], |
| Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3165 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3166 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3167 | ":tables", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3168 | "@FP16", |
| 3169 | "@psimd", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3170 | "@pthreadpool", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3171 | ], |
| 3172 | ) |
| 3173 | |
| 3174 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3175 | name = "psimd_fastmath_ukernels_test_mode", |
| 3176 | hdrs = INTERNAL_HDRS, |
| 3177 | aarch32_copts = [ |
| 3178 | "-marm", |
| 3179 | "-mfpu=neon", |
| 3180 | ], |
| 3181 | copts = [ |
| 3182 | "-UNDEBUG", |
| 3183 | "-DXNN_TEST_MODE=1", |
| 3184 | ], |
| 3185 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3186 | gcc_x86_copts = ["-msse2"], |
| 3187 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3188 | optimized_copts = [ |
| 3189 | "-O3", |
| 3190 | "-ffast-math", |
| 3191 | ], |
| 3192 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
| 3193 | deps = [ |
| 3194 | ":tables", |
| 3195 | "@FP16", |
| 3196 | "@psimd", |
| 3197 | "@pthreadpool", |
| 3198 | ], |
| 3199 | ) |
| 3200 | |
| 3201 | xnnpack_cc_library( |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3202 | name = "psimd_accmath_ukernels", |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3203 | hdrs = INTERNAL_HDRS, |
| 3204 | aarch32_copts = [ |
| 3205 | "-marm", |
| 3206 | "-mfpu=neon", |
| 3207 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3208 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3209 | gcc_x86_copts = ["-msse2"], |
| 3210 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3211 | optimized_copts = [ |
| 3212 | "-O3", |
| 3213 | ], |
| Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3214 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3215 | deps = [ |
| 3216 | ":tables", |
| 3217 | "@FP16", |
| 3218 | "@psimd", |
| 3219 | "@pthreadpool", |
| 3220 | ], |
| 3221 | ) |
| 3222 | |
| 3223 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3224 | name = "psimd_accmath_ukernels_test_mode", |
| 3225 | hdrs = INTERNAL_HDRS, |
| 3226 | aarch32_copts = [ |
| 3227 | "-marm", |
| 3228 | "-mfpu=neon", |
| 3229 | ], |
| 3230 | copts = [ |
| 3231 | "-UNDEBUG", |
| 3232 | "-DXNN_TEST_MODE=1", |
| 3233 | ], |
| 3234 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3235 | gcc_x86_copts = ["-msse2"], |
| 3236 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3237 | optimized_copts = [ |
| 3238 | "-O3", |
| 3239 | ], |
| 3240 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
| 3241 | deps = [ |
| 3242 | ":tables", |
| 3243 | "@FP16", |
| 3244 | "@psimd", |
| 3245 | "@pthreadpool", |
| 3246 | ], |
| 3247 | ) |
| 3248 | |
| 3249 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3250 | name = "neon_ukernels", |
| 3251 | hdrs = INTERNAL_HDRS, |
| 3252 | aarch32_copts = [ |
| 3253 | "-marm", |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3254 | "-march=armv7-a", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3255 | "-mfpu=neon", |
| 3256 | ], |
| 3257 | aarch32_srcs = NEON_UKERNELS, |
| 3258 | aarch64_srcs = NEON_UKERNELS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3259 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3260 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3261 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3262 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3263 | "@FP16", |
| 3264 | "@pthreadpool", |
| 3265 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3266 | ) |
| 3267 | |
| 3268 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3269 | name = "neon_ukernels_test_mode", |
| 3270 | hdrs = INTERNAL_HDRS, |
| 3271 | aarch32_copts = [ |
| 3272 | "-marm", |
| 3273 | "-march=armv7-a", |
| 3274 | "-mfpu=neon", |
| 3275 | ], |
| 3276 | aarch32_srcs = NEON_UKERNELS, |
| 3277 | aarch64_srcs = NEON_UKERNELS, |
| 3278 | copts = [ |
| 3279 | "-UNDEBUG", |
| 3280 | "-DXNN_TEST_MODE=1", |
| 3281 | ], |
| 3282 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3283 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3284 | deps = [ |
| 3285 | ":tables", |
| 3286 | "@FP16", |
| 3287 | "@pthreadpool", |
| 3288 | ], |
| 3289 | ) |
| 3290 | |
| 3291 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3292 | name = "neonfma_ukernels", |
| 3293 | hdrs = INTERNAL_HDRS, |
| 3294 | aarch32_copts = [ |
| 3295 | "-marm", |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3296 | "-march=armv7-a", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3297 | "-mfpu=neon-vfpv4", |
| 3298 | ], |
| 3299 | aarch32_srcs = NEONFMA_UKERNELS, |
| 3300 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3301 | apple_aarch32_copts = [ |
| 3302 | "-mcpu=swift", |
| 3303 | "-mtune=generic", |
| 3304 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3305 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3306 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3307 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3308 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3309 | "@FP16", |
| 3310 | "@pthreadpool", |
| 3311 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3312 | ) |
| 3313 | |
| 3314 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3315 | name = "neonfma_ukernels_test_mode", |
| 3316 | hdrs = INTERNAL_HDRS, |
| 3317 | aarch32_copts = [ |
| 3318 | "-marm", |
| 3319 | "-march=armv7-a", |
| 3320 | "-mfpu=neon-vfpv4", |
| 3321 | ], |
| 3322 | aarch32_srcs = NEONFMA_UKERNELS, |
| 3323 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3324 | apple_aarch32_copts = [ |
| 3325 | "-mcpu=swift", |
| 3326 | "-mtune=generic", |
| 3327 | ], |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3328 | copts = [ |
| 3329 | "-UNDEBUG", |
| 3330 | "-DXNN_TEST_MODE=1", |
| 3331 | ], |
| 3332 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3333 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3334 | deps = [ |
| 3335 | ":tables", |
| 3336 | "@FP16", |
| 3337 | "@pthreadpool", |
| 3338 | ], |
| 3339 | ) |
| 3340 | |
| 3341 | xnnpack_cc_library( |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3342 | name = "neonv8_ukernels", |
| 3343 | hdrs = INTERNAL_HDRS, |
| 3344 | aarch32_copts = [ |
| 3345 | "-marm", |
| 3346 | "-march=armv8-a", |
| 3347 | "-mfpu=neon-fp-armv8", |
| 3348 | ], |
| 3349 | aarch32_srcs = NEONV8_UKERNELS, |
| 3350 | aarch64_srcs = NEONV8_UKERNELS, |
| Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3351 | apple_aarch32_copts = [ |
| 3352 | "-mcpu=cyclone", |
| 3353 | "-mtune=generic", |
| 3354 | ], |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3355 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3356 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3357 | deps = [ |
| 3358 | ":tables", |
| 3359 | "@FP16", |
| 3360 | "@pthreadpool", |
| 3361 | ], |
| 3362 | ) |
| 3363 | |
| 3364 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3365 | name = "neonv8_ukernels_test_mode", |
| 3366 | hdrs = INTERNAL_HDRS, |
| 3367 | aarch32_copts = [ |
| 3368 | "-marm", |
| 3369 | "-march=armv8-a", |
| 3370 | "-mfpu=neon-fp-armv8", |
| 3371 | ], |
| 3372 | aarch32_srcs = NEONV8_UKERNELS, |
| 3373 | aarch64_srcs = NEONV8_UKERNELS, |
| Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3374 | apple_aarch32_copts = [ |
| 3375 | "-mcpu=cyclone", |
| 3376 | "-mtune=generic", |
| 3377 | ], |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3378 | copts = [ |
| 3379 | "-UNDEBUG", |
| 3380 | "-DXNN_TEST_MODE=1", |
| 3381 | ], |
| 3382 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3383 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3384 | deps = [ |
| 3385 | ":tables", |
| 3386 | "@FP16", |
| 3387 | "@pthreadpool", |
| 3388 | ], |
| 3389 | ) |
| 3390 | |
| 3391 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3392 | name = "neonfp16arith_ukernels", |
| 3393 | hdrs = INTERNAL_HDRS, |
| 3394 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 3395 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3396 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3397 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3398 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3399 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3400 | "@FP16", |
| 3401 | "@pthreadpool", |
| 3402 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3403 | ) |
| 3404 | |
| 3405 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3406 | name = "neonfp16arith_ukernels_test_mode", |
| 3407 | hdrs = INTERNAL_HDRS, |
| 3408 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 3409 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 3410 | copts = [ |
| 3411 | "-UNDEBUG", |
| 3412 | "-DXNN_TEST_MODE=1", |
| 3413 | ], |
| 3414 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3415 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3416 | deps = [ |
| 3417 | ":tables", |
| 3418 | "@FP16", |
| 3419 | "@pthreadpool", |
| 3420 | ], |
| 3421 | ) |
| 3422 | |
| 3423 | xnnpack_cc_library( |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3424 | name = "neondot_ukernels", |
| 3425 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 799ac75 | 2020-08-13 16:08:03 -0700 | [diff] [blame] | 3426 | aarch32_copts = [ |
| 3427 | "-marm", |
| 3428 | "-march=armv8.2-a+dotprod", |
| 3429 | "-mfpu=neon-fp-armv8", |
| 3430 | ], |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3431 | aarch32_srcs = NEONDOT_UKERNELS, |
| 3432 | aarch64_copts = ["-march=armv8.2-a+dotprod"], |
| 3433 | aarch64_srcs = NEONDOT_UKERNELS, |
| 3434 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3435 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3436 | deps = [ |
| 3437 | ":tables", |
| 3438 | "@FP16", |
| 3439 | "@pthreadpool", |
| 3440 | ], |
| 3441 | ) |
| 3442 | |
| 3443 | xnnpack_cc_library( |
| 3444 | name = "neondot_ukernels_test_mode", |
| 3445 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 799ac75 | 2020-08-13 16:08:03 -0700 | [diff] [blame] | 3446 | aarch32_copts = [ |
| 3447 | "-marm", |
| 3448 | "-march=armv8.2-a+dotprod", |
| 3449 | "-mfpu=neon-fp-armv8", |
| 3450 | ], |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3451 | aarch32_srcs = NEONDOT_UKERNELS, |
| 3452 | aarch64_copts = ["-march=armv8.2-a+dotprod"], |
| 3453 | aarch64_srcs = NEONDOT_UKERNELS, |
| 3454 | copts = [ |
| 3455 | "-UNDEBUG", |
| 3456 | "-DXNN_TEST_MODE=1", |
| 3457 | ], |
| 3458 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3459 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3460 | deps = [ |
| 3461 | ":tables", |
| 3462 | "@FP16", |
| 3463 | "@pthreadpool", |
| 3464 | ], |
| 3465 | ) |
| 3466 | |
| 3467 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3468 | name = "sse2_ukernels", |
| 3469 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3470 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3471 | gcc_x86_copts = ["-msse2"], |
| 3472 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3473 | msvc_x86_32_copts = ["/arch:SSE2"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3474 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3475 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3476 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3477 | "@FP16", |
| 3478 | "@pthreadpool", |
| 3479 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3480 | ) |
| 3481 | |
| 3482 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3483 | name = "sse2_ukernels_test_mode", |
| 3484 | hdrs = INTERNAL_HDRS, |
| 3485 | copts = [ |
| 3486 | "-UNDEBUG", |
| 3487 | "-DXNN_TEST_MODE=1", |
| 3488 | ], |
| 3489 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3490 | gcc_x86_copts = ["-msse2"], |
| 3491 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3492 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3493 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| 3494 | deps = [ |
| 3495 | ":tables", |
| 3496 | "@FP16", |
| 3497 | "@pthreadpool", |
| 3498 | ], |
| 3499 | ) |
| 3500 | |
| 3501 | xnnpack_cc_library( |
| Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3502 | name = "ssse3_ukernels", |
| 3503 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3504 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3505 | gcc_x86_copts = ["-mssse3"], |
| 3506 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3507 | msvc_x86_32_copts = ["/arch:SSE2"], |
| Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3508 | x86_srcs = SSSE3_UKERNELS, |
| 3509 | deps = [ |
| 3510 | ":tables", |
| 3511 | "@FP16", |
| 3512 | "@pthreadpool", |
| 3513 | ], |
| 3514 | ) |
| 3515 | |
| 3516 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3517 | name = "ssse3_ukernels_test_mode", |
| 3518 | hdrs = INTERNAL_HDRS, |
| 3519 | copts = [ |
| 3520 | "-UNDEBUG", |
| 3521 | "-DXNN_TEST_MODE=1", |
| 3522 | ], |
| 3523 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3524 | gcc_x86_copts = ["-mssse3"], |
| 3525 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3526 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3527 | x86_srcs = SSSE3_UKERNELS, |
| 3528 | deps = [ |
| 3529 | ":tables", |
| 3530 | "@FP16", |
| 3531 | "@pthreadpool", |
| 3532 | ], |
| 3533 | ) |
| 3534 | |
| 3535 | xnnpack_cc_library( |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3536 | name = "sse41_ukernels", |
| 3537 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3538 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3539 | gcc_x86_copts = ["-msse4.1"], |
| 3540 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3541 | msvc_x86_32_copts = ["/arch:SSE2"], |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3542 | x86_srcs = SSE41_UKERNELS, |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3543 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3544 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3545 | "@FP16", |
| 3546 | "@pthreadpool", |
| 3547 | ], |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3548 | ) |
| 3549 | |
| 3550 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3551 | name = "sse41_ukernels_test_mode", |
| 3552 | hdrs = INTERNAL_HDRS, |
| 3553 | copts = [ |
| 3554 | "-UNDEBUG", |
| 3555 | "-DXNN_TEST_MODE=1", |
| 3556 | ], |
| 3557 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3558 | gcc_x86_copts = ["-msse4.1"], |
| 3559 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3560 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3561 | x86_srcs = SSE41_UKERNELS, |
| 3562 | deps = [ |
| 3563 | ":tables", |
| 3564 | "@FP16", |
| 3565 | "@pthreadpool", |
| 3566 | ], |
| 3567 | ) |
| 3568 | |
| 3569 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3570 | name = "avx_ukernels", |
| 3571 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3572 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3573 | gcc_x86_copts = ["-mavx"], |
| 3574 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3575 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3576 | msvc_x86_64_copts = ["/arch:AVX"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3577 | x86_srcs = AVX_UKERNELS, |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3578 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3579 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3580 | "@FP16", |
| 3581 | "@pthreadpool", |
| 3582 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3583 | ) |
| 3584 | |
| 3585 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3586 | name = "avx_ukernels_test_mode", |
| 3587 | hdrs = INTERNAL_HDRS, |
| 3588 | copts = [ |
| 3589 | "-UNDEBUG", |
| 3590 | "-DXNN_TEST_MODE=1", |
| 3591 | ], |
| 3592 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3593 | gcc_x86_copts = ["-mavx"], |
| 3594 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3595 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3596 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3597 | x86_srcs = AVX_UKERNELS, |
| 3598 | deps = [ |
| 3599 | ":tables", |
| 3600 | "@FP16", |
| 3601 | "@pthreadpool", |
| 3602 | ], |
| 3603 | ) |
| 3604 | |
| 3605 | xnnpack_cc_library( |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3606 | name = "xop_ukernels", |
| 3607 | hdrs = INTERNAL_HDRS, |
| 3608 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3609 | gcc_x86_copts = ["-mxop"], |
| 3610 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3611 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3612 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3613 | x86_srcs = XOP_UKERNELS, |
| 3614 | deps = [ |
| 3615 | ":tables", |
| 3616 | "@FP16", |
| 3617 | "@pthreadpool", |
| 3618 | ], |
| 3619 | ) |
| 3620 | |
| 3621 | xnnpack_cc_library( |
| 3622 | name = "xop_ukernels_test_mode", |
| 3623 | hdrs = INTERNAL_HDRS, |
| 3624 | copts = [ |
| 3625 | "-UNDEBUG", |
| 3626 | "-DXNN_TEST_MODE=1", |
| 3627 | ], |
| 3628 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3629 | gcc_x86_copts = ["-mxop"], |
| 3630 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3631 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3632 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3633 | x86_srcs = XOP_UKERNELS, |
| 3634 | deps = [ |
| 3635 | ":tables", |
| 3636 | "@FP16", |
| 3637 | "@pthreadpool", |
| 3638 | ], |
| 3639 | ) |
| 3640 | |
| 3641 | xnnpack_cc_library( |
| Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3642 | name = "fma3_ukernels", |
| 3643 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3644 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3645 | gcc_x86_copts = ["-mfma"], |
| 3646 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3647 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3648 | msvc_x86_64_copts = ["/arch:AVX"], |
| Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3649 | x86_srcs = FMA3_UKERNELS, |
| 3650 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3651 | ":tables", |
| Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3652 | "@FP16", |
| 3653 | "@pthreadpool", |
| 3654 | ], |
| 3655 | ) |
| 3656 | |
| 3657 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3658 | name = "fma3_ukernels_test_mode", |
| 3659 | hdrs = INTERNAL_HDRS, |
| 3660 | copts = [ |
| 3661 | "-UNDEBUG", |
| 3662 | "-DXNN_TEST_MODE=1", |
| 3663 | ], |
| 3664 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3665 | gcc_x86_copts = ["-mfma"], |
| 3666 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3667 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3668 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3669 | x86_srcs = FMA3_UKERNELS, |
| 3670 | deps = [ |
| 3671 | ":tables", |
| 3672 | "@FP16", |
| 3673 | "@pthreadpool", |
| 3674 | ], |
| 3675 | ) |
| 3676 | |
| 3677 | xnnpack_cc_library( |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3678 | name = "avx2_ukernels", |
| 3679 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3680 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3681 | gcc_x86_copts = [ |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3682 | "-mfma", |
| 3683 | "-mavx2", |
| 3684 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3685 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3686 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 3687 | msvc_x86_64_copts = ["/arch:AVX2"], |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3688 | x86_srcs = AVX2_UKERNELS, |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3689 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3690 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3691 | "@FP16", |
| 3692 | "@pthreadpool", |
| 3693 | ], |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3694 | ) |
| 3695 | |
| 3696 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3697 | name = "avx2_ukernels_test_mode", |
| 3698 | hdrs = INTERNAL_HDRS, |
| 3699 | copts = [ |
| 3700 | "-UNDEBUG", |
| 3701 | "-DXNN_TEST_MODE=1", |
| 3702 | ], |
| 3703 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3704 | gcc_x86_copts = [ |
| 3705 | "-mfma", |
| 3706 | "-mavx2", |
| 3707 | ], |
| 3708 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3709 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 3710 | msvc_x86_64_copts = ["/arch:AVX2"], |
| 3711 | x86_srcs = AVX2_UKERNELS, |
| 3712 | deps = [ |
| 3713 | ":tables", |
| 3714 | "@FP16", |
| 3715 | "@pthreadpool", |
| 3716 | ], |
| 3717 | ) |
| 3718 | |
| 3719 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3720 | name = "avx512f_ukernels", |
| 3721 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3722 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3723 | gcc_x86_copts = ["-mavx512f"], |
| 3724 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3725 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3726 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3727 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3728 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3729 | x86_srcs = AVX512F_UKERNELS, |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3730 | deps = [ |
| Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3731 | ":tables", |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3732 | "@FP16", |
| 3733 | "@pthreadpool", |
| 3734 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3735 | ) |
| 3736 | |
| 3737 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3738 | name = "avx512f_ukernels_test_mode", |
| 3739 | hdrs = INTERNAL_HDRS, |
| 3740 | copts = [ |
| 3741 | "-UNDEBUG", |
| 3742 | "-DXNN_TEST_MODE=1", |
| 3743 | ], |
| 3744 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3745 | gcc_x86_copts = ["-mavx512f"], |
| 3746 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3747 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3748 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3749 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3750 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3751 | x86_srcs = AVX512F_UKERNELS, |
| 3752 | deps = [ |
| 3753 | ":tables", |
| 3754 | "@FP16", |
| 3755 | "@pthreadpool", |
| 3756 | ], |
| 3757 | ) |
| 3758 | |
| 3759 | xnnpack_cc_library( |
| Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3760 | name = "avx512skx_ukernels", |
| 3761 | hdrs = INTERNAL_HDRS, |
| 3762 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3763 | gcc_x86_copts = [ |
| 3764 | "-mavx512f", |
| 3765 | "-mavx512cd", |
| 3766 | "-mavx512bw", |
| 3767 | "-mavx512dq", |
| 3768 | "-mavx512vl", |
| 3769 | ], |
| 3770 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3771 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3772 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3773 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3774 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3775 | x86_srcs = AVX512SKX_UKERNELS, |
| 3776 | deps = [ |
| 3777 | ":tables", |
| 3778 | "@FP16", |
| 3779 | "@pthreadpool", |
| 3780 | ], |
| 3781 | ) |
| 3782 | |
| 3783 | xnnpack_cc_library( |
| 3784 | name = "avx512skx_ukernels_test_mode", |
| 3785 | hdrs = INTERNAL_HDRS, |
| 3786 | copts = [ |
| 3787 | "-UNDEBUG", |
| 3788 | "-DXNN_TEST_MODE=1", |
| 3789 | ], |
| 3790 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3791 | gcc_x86_copts = [ |
| 3792 | "-mavx512f", |
| 3793 | "-mavx512cd", |
| 3794 | "-mavx512bw", |
| 3795 | "-mavx512dq", |
| 3796 | "-mavx512vl", |
| 3797 | ], |
| 3798 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3799 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3800 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3801 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3802 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3803 | x86_srcs = AVX512SKX_UKERNELS, |
| 3804 | deps = [ |
| 3805 | ":tables", |
| 3806 | "@FP16", |
| 3807 | "@pthreadpool", |
| 3808 | ], |
| 3809 | ) |
| 3810 | |
| 3811 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3812 | name = "asm_ukernels", |
| 3813 | hdrs = ["src/xnnpack/assembly.h"], |
| 3814 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
| Frank Barchard | 31bb45b | 2020-10-06 00:26:33 -0700 | [diff] [blame] | 3815 | aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3816 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
| Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3817 | wasm_srcs = WASM32_ASM_UKERNELS, |
| 3818 | wasmsimd_srcs = WASM32_ASM_UKERNELS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3819 | ) |
| 3820 | |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3821 | xnnpack_cc_library( |
| 3822 | name = "logging_utils", |
| 3823 | srcs = LOGGING_SRCS, |
| 3824 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3825 | copts = LOGGING_COPTS + [ |
| 3826 | "-Isrc", |
| 3827 | "-Iinclude", |
| 3828 | ] + select({ |
| 3829 | ":debug_build": [], |
| 3830 | "//conditions:default": xnnpack_min_size_copts(), |
| 3831 | }), |
| 3832 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3833 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3834 | visibility = xnnpack_visibility(), |
| 3835 | deps = [ |
| 3836 | "@FP16", |
| 3837 | "@clog", |
| 3838 | "@pthreadpool", |
| 3839 | ], |
| 3840 | ) |
| 3841 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3842 | xnnpack_aggregate_library( |
| 3843 | name = "ukernels", |
| 3844 | aarch32_deps = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3845 | ":neon_ukernels", |
| 3846 | ":neonfma_ukernels", |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3847 | ":neonv8_ukernels", |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3848 | ":neondot_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3849 | ":asm_ukernels", |
| 3850 | ], |
| 3851 | aarch64_deps = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3852 | ":neon_ukernels", |
| 3853 | ":neonfma_ukernels", |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3854 | ":neonv8_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3855 | ":neonfp16arith_ukernels", |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3856 | ":neondot_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3857 | ":asm_ukernels", |
| 3858 | ], |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3859 | generic_deps = [ |
| 3860 | ":scalar_ukernels", |
| 3861 | ], |
| Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3862 | psimd_deps = [ |
| Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3863 | ":psimd_fastmath_ukernels", |
| 3864 | ":psimd_accmath_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3865 | ], |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3866 | wasm_deps = [ |
| 3867 | ":wasm_ukernels", |
| Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3868 | ":asm_ukernels", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3869 | ], |
| 3870 | wasmsimd_deps = [ |
| 3871 | ":wasm_ukernels", |
| Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3872 | ":asm_ukernels", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3873 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3874 | x86_deps = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3875 | ":sse2_ukernels", |
| Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3876 | ":ssse3_ukernels", |
| Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3877 | ":sse41_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3878 | ":avx_ukernels", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3879 | ":xop_ukernels", |
| Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3880 | ":fma3_ukernels", |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3881 | ":avx2_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3882 | ":avx512f_ukernels", |
| Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3883 | ":avx512skx_ukernels", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3884 | ], |
| 3885 | ) |
| 3886 | |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3887 | xnnpack_aggregate_library( |
| 3888 | name = "ukernels_test_mode", |
| 3889 | aarch32_deps = [ |
| 3890 | ":neon_ukernels_test_mode", |
| 3891 | ":neonfma_ukernels_test_mode", |
| 3892 | ":neonv8_ukernels_test_mode", |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3893 | ":neondot_ukernels_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3894 | ":asm_ukernels", |
| 3895 | ], |
| 3896 | aarch64_deps = [ |
| 3897 | ":neon_ukernels_test_mode", |
| 3898 | ":neonfma_ukernels_test_mode", |
| 3899 | ":neonv8_ukernels_test_mode", |
| 3900 | ":neonfp16arith_ukernels_test_mode", |
| Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3901 | ":neondot_ukernels_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3902 | ":asm_ukernels", |
| 3903 | ], |
| 3904 | generic_deps = [ |
| 3905 | ":scalar_ukernels_test_mode", |
| 3906 | ], |
| 3907 | psimd_deps = [ |
| 3908 | ":psimd_fastmath_ukernels_test_mode", |
| 3909 | ":psimd_accmath_ukernels_test_mode", |
| 3910 | ], |
| 3911 | wasm_deps = [ |
| 3912 | ":wasm_ukernels_test_mode", |
| Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3913 | ":asm_ukernels", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3914 | ], |
| 3915 | wasmsimd_deps = [ |
| 3916 | ":wasm_ukernels_test_mode", |
| Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3917 | ":asm_ukernels", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3918 | ], |
| 3919 | x86_deps = [ |
| 3920 | ":sse2_ukernels_test_mode", |
| 3921 | ":ssse3_ukernels_test_mode", |
| 3922 | ":sse41_ukernels_test_mode", |
| 3923 | ":avx_ukernels_test_mode", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3924 | ":xop_ukernels_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3925 | ":fma3_ukernels_test_mode", |
| 3926 | ":avx2_ukernels_test_mode", |
| 3927 | ":avx512f_ukernels_test_mode", |
| Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3928 | ":avx512skx_ukernels_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3929 | ], |
| 3930 | ) |
| 3931 | |
| Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3932 | # This target must be compatible with Android NDK r17 and compile with Clang 6. |
| 3933 | # It is similar to :ukernels target, but excludes microkernels which require |
| 3934 | # NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics. |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3935 | xnnpack_aggregate_library( |
| Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3936 | name = "ukernels_max_compatibility", |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3937 | aarch32_deps = [ |
| 3938 | ":neon_ukernels", |
| 3939 | ":neonfma_ukernels", |
| 3940 | ":neonv8_ukernels", |
| 3941 | ":asm_ukernels", |
| 3942 | ], |
| 3943 | aarch64_deps = [ |
| 3944 | ":neon_ukernels", |
| 3945 | ":neonfma_ukernels", |
| 3946 | ":neonv8_ukernels", |
| 3947 | ":asm_ukernels", |
| 3948 | ], |
| 3949 | generic_deps = [ |
| 3950 | ":scalar_ukernels", |
| 3951 | ], |
| 3952 | psimd_deps = [ |
| 3953 | ":psimd_fastmath_ukernels", |
| 3954 | ":psimd_accmath_ukernels", |
| 3955 | ], |
| 3956 | wasm_deps = [ |
| 3957 | ":wasm_ukernels", |
| 3958 | ], |
| 3959 | wasmsimd_deps = [ |
| 3960 | ":wasm_ukernels", |
| 3961 | ], |
| 3962 | x86_deps = [ |
| 3963 | ":sse2_ukernels", |
| 3964 | ":ssse3_ukernels", |
| 3965 | ":sse41_ukernels", |
| 3966 | ":avx_ukernels", |
| Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3967 | ":xop_ukernels", |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3968 | ":fma3_ukernels", |
| 3969 | ":avx2_ukernels", |
| 3970 | ":avx512f_ukernels", |
| Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3971 | ":avx512skx_ukernels", |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3972 | ], |
| 3973 | ) |
| 3974 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3975 | xnnpack_cc_library( |
| 3976 | name = "im2col", |
| 3977 | srcs = ["src/im2col.c"], |
| 3978 | hdrs = [ |
| 3979 | "src/xnnpack/common.h", |
| 3980 | "src/xnnpack/im2col.h", |
| 3981 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3982 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3983 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3984 | ) |
| 3985 | |
| 3986 | xnnpack_cc_library( |
| 3987 | name = "indirection", |
| 3988 | srcs = ["src/indirection.c"], |
| 3989 | hdrs = INTERNAL_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3990 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3991 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3992 | deps = [ |
| 3993 | "@FP16", |
| 3994 | "@FXdiv", |
| 3995 | "@pthreadpool", |
| 3996 | ], |
| 3997 | ) |
| 3998 | |
| 3999 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4000 | name = "indirection_test_mode", |
| 4001 | srcs = ["src/indirection.c"], |
| 4002 | hdrs = INTERNAL_HDRS, |
| 4003 | copts = [ |
| 4004 | "-UNDEBUG", |
| 4005 | "-DXNN_TEST_MODE=1", |
| 4006 | ], |
| 4007 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4008 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4009 | deps = [ |
| 4010 | "@FP16", |
| 4011 | "@FXdiv", |
| 4012 | "@pthreadpool", |
| 4013 | ], |
| 4014 | ) |
| 4015 | |
| 4016 | xnnpack_cc_library( |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4017 | name = "packing", |
| 4018 | srcs = ["src/packing.c"], |
| 4019 | hdrs = INTERNAL_HDRS, |
| 4020 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4021 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4022 | deps = [ |
| 4023 | "@FP16", |
| 4024 | "@FXdiv", |
| 4025 | "@pthreadpool", |
| 4026 | ], |
| 4027 | ) |
| 4028 | |
| 4029 | xnnpack_cc_library( |
| 4030 | name = "packing_test_mode", |
| 4031 | srcs = ["src/packing.c"], |
| 4032 | hdrs = INTERNAL_HDRS, |
| 4033 | copts = [ |
| 4034 | "-UNDEBUG", |
| 4035 | "-DXNN_TEST_MODE=1", |
| 4036 | ], |
| 4037 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4038 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4039 | deps = [ |
| 4040 | "@FP16", |
| 4041 | "@FXdiv", |
| 4042 | "@pthreadpool", |
| 4043 | ], |
| 4044 | ) |
| 4045 | |
| 4046 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4047 | name = "operator_run", |
| 4048 | srcs = ["src/operator-run.c"], |
| Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 4049 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4050 | copts = LOGGING_COPTS + select({ |
| Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4051 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4052 | "//conditions:default": [], |
| 4053 | }), |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4054 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4055 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4056 | deps = [ |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4057 | ":logging_utils", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4058 | "@FP16", |
| 4059 | "@FXdiv", |
| 4060 | "@clog", |
| 4061 | "@pthreadpool", |
| 4062 | ], |
| 4063 | ) |
| 4064 | |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4065 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4066 | name = "operator_run_test_mode", |
| 4067 | srcs = ["src/operator-run.c"], |
| 4068 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 4069 | copts = LOGGING_COPTS + [ |
| 4070 | "-UNDEBUG", |
| 4071 | "-DXNN_TEST_MODE=1", |
| 4072 | ] + select({ |
| 4073 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4074 | "//conditions:default": [], |
| 4075 | }), |
| 4076 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4077 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4078 | deps = [ |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4079 | ":logging_utils", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4080 | "@FP16", |
| 4081 | "@FXdiv", |
| 4082 | "@clog", |
| 4083 | "@pthreadpool", |
| 4084 | ], |
| 4085 | ) |
| 4086 | |
| 4087 | xnnpack_cc_library( |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4088 | name = "memory_planner", |
| 4089 | srcs = ["src/memory-planner.c"], |
| 4090 | hdrs = INTERNAL_HDRS, |
| 4091 | defines = select({ |
| 4092 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 4093 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 4094 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 4095 | }), |
| 4096 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4097 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4098 | deps = [ |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4099 | ":logging_utils", |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4100 | "@pthreadpool", |
| 4101 | ], |
| 4102 | ) |
| 4103 | |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4104 | xnnpack_cc_library( |
| 4105 | name = "memory_planner_test_mode", |
| 4106 | srcs = ["src/memory-planner.c"], |
| 4107 | hdrs = INTERNAL_HDRS, |
| 4108 | copts = [ |
| 4109 | "-UNDEBUG", |
| 4110 | "-DXNN_TEST_MODE=1", |
| 4111 | ], |
| 4112 | defines = select({ |
| 4113 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 4114 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 4115 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 4116 | }), |
| 4117 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4118 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4119 | deps = [ |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4120 | ":logging_utils", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4121 | "@pthreadpool", |
| 4122 | ], |
| 4123 | ) |
| 4124 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4125 | cc_library( |
| 4126 | name = "enable_assembly", |
| 4127 | defines = select({ |
| 4128 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 4129 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
| Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 4130 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4131 | }), |
| 4132 | ) |
| 4133 | |
| Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4134 | cc_library( |
| 4135 | name = "enable_sparse", |
| 4136 | defines = select({ |
| 4137 | ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"], |
| 4138 | ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"], |
| Marat Dukhan | 8b283aa | 2020-06-18 16:48:57 -0700 | [diff] [blame] | 4139 | "//conditions:default": ["XNN_ENABLE_SPARSE=0"], |
| Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4140 | }), |
| 4141 | ) |
| 4142 | |
| Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4143 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4144 | name = "operators", |
| 4145 | srcs = OPERATOR_SRCS + [ |
| Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 4146 | "src/memory.c", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4147 | "src/operator-delete.c", |
| Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4148 | ], |
| 4149 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4150 | copts = LOGGING_COPTS + [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4151 | "-Isrc", |
| 4152 | "-Iinclude", |
| 4153 | ] + select({ |
| 4154 | ":debug_build": [], |
| 4155 | "//conditions:default": xnnpack_min_size_copts(), |
| Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4156 | }) + select({ |
| 4157 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4158 | "//conditions:default": [], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4159 | }), |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4160 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4161 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4162 | deps = [ |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4163 | ":indirection", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4164 | ":logging_utils", |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4165 | ":packing", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4166 | "@FP16", |
| 4167 | "@FXdiv", |
| 4168 | "@clog", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4169 | "@pthreadpool", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4170 | ], |
| 4171 | ) |
| 4172 | |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4173 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4174 | name = "operators_test_mode", |
| 4175 | srcs = OPERATOR_SRCS + [ |
| 4176 | "src/memory.c", |
| 4177 | "src/operator-delete.c", |
| 4178 | ], |
| 4179 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 4180 | copts = LOGGING_COPTS + [ |
| 4181 | "-Isrc", |
| 4182 | "-Iinclude", |
| 4183 | "-UNDEBUG", |
| 4184 | "-DXNN_TEST_MODE=1", |
| 4185 | ] + select({ |
| 4186 | ":debug_build": [], |
| 4187 | "//conditions:default": xnnpack_min_size_copts(), |
| 4188 | }) + select({ |
| 4189 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4190 | "//conditions:default": [], |
| 4191 | }), |
| 4192 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4193 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4194 | deps = [ |
| 4195 | ":indirection_test_mode", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4196 | ":logging_utils", |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4197 | ":packing_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4198 | "@FP16", |
| 4199 | "@FXdiv", |
| 4200 | "@clog", |
| 4201 | "@pthreadpool", |
| 4202 | ], |
| 4203 | ) |
| 4204 | |
| 4205 | xnnpack_cc_library( |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4206 | name = "XNNPACK", |
| 4207 | srcs = [ |
| 4208 | "src/init.c", |
| Marat Dukhan | ccfdbd1 | 2020-02-03 14:27:45 -0800 | [diff] [blame] | 4209 | "src/runtime.c", |
| 4210 | "src/subgraph.c", |
| 4211 | "src/tensor.c", |
| Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 4212 | ] + SUBGRAPH_SRCS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4213 | hdrs = ["include/xnnpack.h"], |
| 4214 | copts = LOGGING_COPTS + [ |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4215 | "-Isrc", |
| 4216 | "-Iinclude", |
| 4217 | ] + select({ |
| 4218 | ":debug_build": [], |
| 4219 | "//conditions:default": xnnpack_min_size_copts(), |
| Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4220 | }) + select({ |
| 4221 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4222 | "//conditions:default": [], |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4223 | }), |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4224 | gcc_copts = xnnpack_gcc_std_copts(), |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4225 | includes = ["include"], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4226 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4227 | visibility = xnnpack_visibility(), |
| 4228 | deps = [ |
| 4229 | ":enable_assembly", |
| Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4230 | ":enable_sparse", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4231 | ":logging_utils", |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4232 | ":memory_planner", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4233 | ":operator_run", |
| 4234 | ":operators", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4235 | ":ukernels", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4236 | "@clog", |
| Marat Dukhan | ab2946c | 2020-05-21 20:04:13 -0700 | [diff] [blame] | 4237 | "@FP16", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4238 | "@pthreadpool", |
| Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 4239 | ] + select({ |
| 4240 | ":emscripten": [], |
| 4241 | "//conditions:default": ["@cpuinfo"], |
| 4242 | }), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4243 | ) |
| 4244 | |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4245 | xnnpack_cc_library( |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4246 | name = "XNNPACK_test_mode", |
| 4247 | srcs = [ |
| 4248 | "src/init.c", |
| 4249 | "src/runtime.c", |
| 4250 | "src/subgraph.c", |
| 4251 | "src/tensor.c", |
| Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 4252 | ] + SUBGRAPH_SRCS, |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4253 | hdrs = ["include/xnnpack.h"], |
| 4254 | copts = LOGGING_COPTS + [ |
| 4255 | "-Isrc", |
| 4256 | "-Iinclude", |
| 4257 | "-UNDEBUG", |
| 4258 | "-DXNN_TEST_MODE=1", |
| 4259 | ] + select({ |
| 4260 | ":debug_build": [], |
| 4261 | "//conditions:default": xnnpack_min_size_copts(), |
| 4262 | }) + select({ |
| 4263 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4264 | "//conditions:default": [], |
| 4265 | }), |
| 4266 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4267 | includes = ["include"], |
| 4268 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4269 | visibility = xnnpack_visibility(), |
| 4270 | deps = [ |
| 4271 | ":enable_assembly", |
| Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4272 | ":enable_sparse", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4273 | ":logging_utils", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4274 | ":memory_planner_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4275 | ":operator_run_test_mode", |
| 4276 | ":operators_test_mode", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4277 | ":ukernels_test_mode", |
| Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4278 | "@clog", |
| 4279 | "@FP16", |
| 4280 | "@pthreadpool", |
| 4281 | ] + select({ |
| 4282 | ":emscripten": [], |
| 4283 | "//conditions:default": ["@cpuinfo"], |
| 4284 | }), |
| 4285 | ) |
| 4286 | |
| 4287 | xnnpack_cc_library( |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4288 | name = "xnnpack_f32", |
| 4289 | srcs = [ |
| 4290 | "src/init.c", |
| 4291 | "src/runtime.c", |
| 4292 | "src/subgraph.c", |
| 4293 | "src/tensor.c", |
| 4294 | ] + SUBGRAPH_SRCS, |
| 4295 | hdrs = ["include/xnnpack.h"], |
| 4296 | copts = LOGGING_COPTS + [ |
| 4297 | "-Isrc", |
| 4298 | "-Iinclude", |
| 4299 | ] + select({ |
| 4300 | ":debug_build": [], |
| 4301 | "//conditions:default": xnnpack_min_size_copts(), |
| 4302 | }) + select({ |
| 4303 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4304 | "//conditions:default": [], |
| 4305 | }), |
| 4306 | defines = [ |
| Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 4307 | "XNN_NO_QS8_OPERATORS", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4308 | "XNN_NO_QU8_OPERATORS", |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4309 | "XNN_NO_U8_OPERATORS", |
| 4310 | "XNN_NO_X8_OPERATORS", |
| 4311 | "XNN_NO_F16_OPERATORS", |
| 4312 | "XNN_NO_X16_OPERATORS", |
| 4313 | ], |
| 4314 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4315 | includes = ["include"], |
| 4316 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4317 | visibility = xnnpack_visibility(), |
| 4318 | deps = [ |
| 4319 | ":enable_assembly", |
| Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4320 | ":enable_sparse", |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4321 | ":logging_utils", |
| 4322 | ":memory_planner", |
| 4323 | ":operator_run", |
| 4324 | ":operators", |
| Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 4325 | ":ukernels_max_compatibility", |
| Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4326 | "@clog", |
| 4327 | "@FP16", |
| 4328 | "@pthreadpool", |
| 4329 | ] + select({ |
| 4330 | ":emscripten": [], |
| 4331 | "//conditions:default": ["@cpuinfo"], |
| 4332 | }), |
| 4333 | ) |
| 4334 | |
| 4335 | xnnpack_cc_library( |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4336 | name = "xnnpack_operators_nhwc_f32", |
| 4337 | srcs = [ |
| 4338 | "src/init.c", |
| 4339 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4340 | hdrs = ["include/xnnpack.h"], |
| 4341 | copts = LOGGING_COPTS + [ |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4342 | "-Isrc", |
| 4343 | "-Iinclude", |
| 4344 | ] + select({ |
| 4345 | ":debug_build": [], |
| 4346 | "//conditions:default": xnnpack_min_size_copts(), |
| Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4347 | }) + select({ |
| 4348 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4349 | "//conditions:default": [], |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4350 | }), |
| 4351 | defines = [ |
| Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 4352 | "XNN_NO_QS8_OPERATORS", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4353 | "XNN_NO_QU8_OPERATORS", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4354 | "XNN_NO_U8_OPERATORS", |
| 4355 | "XNN_NO_X8_OPERATORS", |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4356 | "XNN_NO_NCHW_OPERATORS", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4357 | ], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4358 | gcc_copts = xnnpack_gcc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4359 | includes = ["include"], |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4360 | msvc_copts = xnnpack_msvc_std_copts(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4361 | visibility = xnnpack_visibility(), |
| 4362 | deps = [ |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4363 | ":enable_assembly", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4364 | ":logging_utils", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4365 | ":operator_run", |
| 4366 | ":operators", |
| Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4367 | ":ukernels", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4368 | "@clog", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4369 | "@pthreadpool", |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4370 | ] + select({ |
| 4371 | ":emscripten": [], |
| 4372 | "//conditions:default": ["@cpuinfo"], |
| 4373 | }), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4374 | ) |
| 4375 | |
| Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4376 | xnnpack_cc_library( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4377 | name = "bench_utils", |
| 4378 | srcs = ["bench/utils.cc"], |
| 4379 | hdrs = ["bench/utils.h"], |
| Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 4380 | deps = [ |
| 4381 | "@com_google_benchmark//:benchmark", |
| 4382 | "@cpuinfo", |
| 4383 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4384 | ) |
| 4385 | |
| Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 4386 | ######################### Benchmarks for micro-kernels ######################### |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4387 | |
| 4388 | xnnpack_benchmark( |
| Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 4389 | name = "qs8_gemm_bench", |
| 4390 | srcs = [ |
| 4391 | "bench/gemm.h", |
| 4392 | "bench/qs8-gemm.cc", |
| 4393 | "src/xnnpack/AlignedAllocator.h", |
| 4394 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Frank Barchard | 31328cb | 2020-10-12 11:55:18 -0700 | [diff] [blame] | 4395 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| 4396 | 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] | 4397 | ) |
| 4398 | |
| 4399 | xnnpack_benchmark( |
| Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4400 | name = "qs8_requantization_bench", |
| 4401 | srcs = [ |
| 4402 | "bench/qs8-requantization.cc", |
| Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4403 | "src/xnnpack/AlignedAllocator.h", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 4404 | "src/xnnpack/requantization-stubs.h", |
| Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4405 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4406 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4407 | ) |
| 4408 | |
| 4409 | xnnpack_benchmark( |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4410 | name = "qu8_gemm_bench", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4411 | srcs = [ |
| 4412 | "bench/gemm.h", |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4413 | "bench/qu8-gemm.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4414 | "src/xnnpack/AlignedAllocator.h", |
| 4415 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4416 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4417 | 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] | 4418 | ) |
| 4419 | |
| 4420 | xnnpack_benchmark( |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4421 | name = "qu8_requantization_bench", |
| 4422 | srcs = [ |
| 4423 | "bench/qu8-requantization.cc", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4424 | "src/xnnpack/AlignedAllocator.h", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 4425 | "src/xnnpack/requantization-stubs.h", |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4426 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4427 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4428 | ) |
| 4429 | |
| 4430 | xnnpack_benchmark( |
| Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 4431 | name = "f16_igemm_bench", |
| 4432 | srcs = [ |
| 4433 | "bench/f16-igemm.cc", |
| 4434 | "bench/conv.h", |
| 4435 | "bench/google/conv.h", |
| 4436 | "src/xnnpack/AlignedAllocator.h", |
| 4437 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4438 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4439 | ":indirection", |
| 4440 | ":packing", |
| 4441 | ], |
| Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 4442 | ) |
| 4443 | |
| 4444 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4445 | name = "f16_gemm_bench", |
| 4446 | srcs = [ |
| 4447 | "bench/f16-gemm.cc", |
| 4448 | "bench/gemm.h", |
| 4449 | "src/xnnpack/AlignedAllocator.h", |
| 4450 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4451 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4452 | ":packing", |
| 4453 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4454 | ) |
| 4455 | |
| 4456 | xnnpack_benchmark( |
| Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4457 | name = "f16_spmm_bench", |
| 4458 | srcs = [ |
| 4459 | "bench/f16-spmm.cc", |
| 4460 | "bench/gemm.h", |
| Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4461 | "src/xnnpack/AlignedAllocator.h", |
| 4462 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4463 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4464 | ) |
| 4465 | |
| 4466 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4467 | name = "f32_igemm_bench", |
| 4468 | srcs = [ |
| 4469 | "bench/f32-igemm.cc", |
| 4470 | "bench/conv.h", |
| 4471 | "src/xnnpack/AlignedAllocator.h", |
| 4472 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4473 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4474 | ":indirection", |
| 4475 | ":packing", |
| 4476 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4477 | ) |
| 4478 | |
| 4479 | xnnpack_benchmark( |
| Frank Barchard | c201006 | 2020-07-19 10:04:12 -0700 | [diff] [blame] | 4480 | name = "f16_relu_bench", |
| 4481 | srcs = [ |
| 4482 | "bench/f16-relu.cc", |
| 4483 | "src/xnnpack/AlignedAllocator.h", |
| 4484 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4485 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4486 | ) |
| 4487 | |
| 4488 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4489 | name = "f32_conv_hwc_bench", |
| 4490 | srcs = [ |
| 4491 | "bench/f32-conv-hwc.cc", |
| 4492 | "bench/dconv.h", |
| 4493 | "src/xnnpack/AlignedAllocator.h", |
| 4494 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4495 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4496 | ":packing", |
| 4497 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4498 | ) |
| 4499 | |
| 4500 | xnnpack_benchmark( |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4501 | name = "f32_conv_hwc2chw_bench", |
| Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4502 | srcs = [ |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4503 | "bench/f32-conv-hwc2chw.cc", |
| Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4504 | "bench/dconv.h", |
| 4505 | "src/xnnpack/AlignedAllocator.h", |
| 4506 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4507 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4508 | ":packing", |
| 4509 | ], |
| Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4510 | ) |
| 4511 | |
| 4512 | xnnpack_benchmark( |
| Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4513 | name = "f16_dwconv_bench", |
| 4514 | srcs = [ |
| 4515 | "bench/f16-dwconv.cc", |
| 4516 | "bench/dwconv.h", |
| 4517 | "bench/google/dwconv.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 | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4524 | ) |
| 4525 | |
| 4526 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4527 | name = "f32_dwconv_bench", |
| 4528 | srcs = [ |
| 4529 | "bench/f32-dwconv.cc", |
| 4530 | "bench/dwconv.h", |
| 4531 | "src/xnnpack/AlignedAllocator.h", |
| 4532 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4533 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4534 | ":indirection", |
| 4535 | ":packing", |
| 4536 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4537 | ) |
| 4538 | |
| 4539 | xnnpack_benchmark( |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 4540 | name = "f32_dwconv2d_chw_bench", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4541 | srcs = [ |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 4542 | "bench/f32-dwconv2d-chw.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4543 | "bench/dwconv.h", |
| 4544 | "src/xnnpack/AlignedAllocator.h", |
| 4545 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4546 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4547 | ":indirection", |
| 4548 | ":packing", |
| 4549 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4550 | ) |
| 4551 | |
| 4552 | xnnpack_benchmark( |
| 4553 | name = "f32_gemm_bench", |
| 4554 | srcs = [ |
| 4555 | "bench/f32-gemm.cc", |
| 4556 | "bench/gemm.h", |
| 4557 | "src/xnnpack/AlignedAllocator.h", |
| 4558 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4559 | copts = xnnpack_optional_ruy_copts(), |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4560 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4561 | ) |
| 4562 | |
| 4563 | xnnpack_benchmark( |
| Marat Dukhan | a11ca34 | 2020-06-25 23:45:07 -0700 | [diff] [blame] | 4564 | name = "f32_hswish_bench", |
| 4565 | srcs = [ |
| 4566 | "bench/f32-hswish.cc", |
| 4567 | "src/xnnpack/AlignedAllocator.h", |
| 4568 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4569 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4570 | ) |
| 4571 | |
| 4572 | xnnpack_benchmark( |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 4573 | name = "f32_raddexpminusmax_bench", |
| 4574 | srcs = [ |
| 4575 | "bench/f32-raddexpminusmax.cc", |
| 4576 | "src/xnnpack/AlignedAllocator.h", |
| 4577 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4578 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4579 | ) |
| 4580 | |
| 4581 | xnnpack_benchmark( |
| 4582 | name = "f32_raddextexp_bench", |
| 4583 | srcs = [ |
| 4584 | "bench/f32-raddextexp.cc", |
| 4585 | "src/xnnpack/AlignedAllocator.h", |
| 4586 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4587 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4588 | ) |
| 4589 | |
| 4590 | xnnpack_benchmark( |
| 4591 | name = "f32_raddstoreexpminusmax_bench", |
| 4592 | srcs = [ |
| 4593 | "bench/f32-raddstoreexpminusmax.cc", |
| 4594 | "src/xnnpack/AlignedAllocator.h", |
| 4595 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4596 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4597 | ) |
| 4598 | |
| 4599 | xnnpack_benchmark( |
| Frank Barchard | 3c4a952 | 2020-07-18 16:11:22 -0700 | [diff] [blame] | 4600 | name = "f32_relu_bench", |
| 4601 | srcs = [ |
| 4602 | "bench/f32-relu.cc", |
| 4603 | "src/xnnpack/AlignedAllocator.h", |
| 4604 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4605 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4606 | ) |
| 4607 | |
| 4608 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4609 | name = "f32_rmax_bench", |
| 4610 | srcs = [ |
| 4611 | "bench/f32-rmax.cc", |
| 4612 | "src/xnnpack/AlignedAllocator.h", |
| 4613 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4614 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4615 | ) |
| 4616 | |
| 4617 | xnnpack_benchmark( |
| Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 4618 | name = "f32_sigmoid_bench", |
| 4619 | srcs = [ |
| 4620 | "bench/f32-sigmoid.cc", |
| 4621 | "src/xnnpack/AlignedAllocator.h", |
| 4622 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 4623 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4624 | ) |
| 4625 | |
| 4626 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4627 | name = "f32_spmm_bench", |
| 4628 | srcs = [ |
| 4629 | "bench/f32-spmm.cc", |
| 4630 | "bench/gemm.h", |
| 4631 | "src/xnnpack/AlignedAllocator.h", |
| 4632 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4633 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4634 | ) |
| 4635 | |
| 4636 | xnnpack_benchmark( |
| Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4637 | name = "f32_softmax_bench", |
| Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4638 | srcs = [ |
| Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4639 | "bench/f32-softmax.cc", |
| Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4640 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4641 | copts = xnnpack_optional_dnnl_copts(), |
| Marat Dukhan | 8d3c693 | 2020-03-06 20:27:27 -0800 | [diff] [blame] | 4642 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(), |
| Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4643 | ) |
| 4644 | |
| 4645 | xnnpack_benchmark( |
| Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 4646 | name = "f32_vscaleexpminusmax_bench", |
| 4647 | srcs = [ |
| 4648 | "bench/f32-vscaleexpminusmax.cc", |
| 4649 | "src/xnnpack/AlignedAllocator.h", |
| 4650 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4651 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4652 | ) |
| 4653 | |
| 4654 | xnnpack_benchmark( |
| 4655 | name = "f32_vscaleextexp_bench", |
| 4656 | srcs = [ |
| 4657 | "bench/f32-vscaleextexp.cc", |
| 4658 | "src/xnnpack/AlignedAllocator.h", |
| 4659 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4660 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4661 | ) |
| 4662 | |
| 4663 | xnnpack_benchmark( |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 4664 | name = "f32_vsqrt_bench", |
| 4665 | srcs = [ |
| 4666 | "bench/f32-vsqrt.cc", |
| 4667 | "src/xnnpack/AlignedAllocator.h", |
| 4668 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4669 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4670 | ) |
| 4671 | |
| 4672 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4673 | name = "f32_im2col_gemm_bench", |
| 4674 | srcs = [ |
| 4675 | "bench/f32-im2col-gemm.cc", |
| 4676 | "bench/conv.h", |
| 4677 | "src/xnnpack/AlignedAllocator.h", |
| 4678 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4679 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4680 | ":im2col", |
| 4681 | ":packing", |
| 4682 | ], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4683 | ) |
| 4684 | |
| Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 4685 | xnnpack_benchmark( |
| Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4686 | name = "rounding_bench", |
| 4687 | srcs = [ |
| 4688 | "bench/rounding.cc", |
| Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4689 | "src/xnnpack/AlignedAllocator.h", |
| Frank Barchard | 04336c1 | 2020-10-22 16:48:55 -0700 | [diff] [blame] | 4690 | "src/xnnpack/math-stubs.h", |
| Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4691 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4692 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4693 | ) |
| 4694 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4695 | ########################### Benchmarks for operators ########################### |
| 4696 | |
| 4697 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4698 | name = "average_pooling_bench", |
| 4699 | srcs = ["bench/average-pooling.cc"], |
| Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 4700 | copts = xnnpack_optional_tflite_copts(), |
| Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4701 | tags = ["nowin32"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4702 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4703 | ) |
| 4704 | |
| 4705 | xnnpack_benchmark( |
| Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4706 | name = "bankers_rounding_bench", |
| 4707 | srcs = ["bench/bankers-rounding.cc"], |
| 4708 | copts = xnnpack_optional_tflite_copts(), |
| 4709 | tags = ["nowin32"], |
| 4710 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4711 | ) |
| 4712 | |
| 4713 | xnnpack_benchmark( |
| 4714 | name = "ceiling_bench", |
| 4715 | srcs = ["bench/ceiling.cc"], |
| 4716 | copts = xnnpack_optional_tflite_copts(), |
| 4717 | tags = ["nowin32"], |
| 4718 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4719 | ) |
| 4720 | |
| 4721 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4722 | name = "channel_shuffle_bench", |
| 4723 | srcs = ["bench/channel-shuffle.cc"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4724 | deps = OPERATOR_BENCHMARK_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4725 | ) |
| 4726 | |
| 4727 | xnnpack_benchmark( |
| 4728 | name = "convolution_bench", |
| 4729 | srcs = ["bench/convolution.cc"], |
| 4730 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
| Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4731 | tags = ["nowin32"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4732 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4733 | ) |
| 4734 | |
| 4735 | xnnpack_benchmark( |
| 4736 | name = "deconvolution_bench", |
| 4737 | srcs = ["bench/deconvolution.cc"], |
| 4738 | copts = xnnpack_optional_tflite_copts(), |
| Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4739 | tags = ["nowin32"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4740 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4741 | ) |
| 4742 | |
| 4743 | xnnpack_benchmark( |
| Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4744 | name = "floor_bench", |
| 4745 | srcs = ["bench/floor.cc"], |
| 4746 | copts = xnnpack_optional_tflite_copts(), |
| 4747 | tags = ["nowin32"], |
| 4748 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4749 | ) |
| 4750 | |
| 4751 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4752 | name = "global_average_pooling_bench", |
| 4753 | srcs = ["bench/global-average-pooling.cc"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4754 | deps = OPERATOR_BENCHMARK_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4755 | ) |
| 4756 | |
| 4757 | xnnpack_benchmark( |
| Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 4758 | name = "hardswish_bench", |
| 4759 | srcs = ["bench/hardswish.cc"], |
| 4760 | copts = xnnpack_optional_tflite_copts(), |
| 4761 | tags = ["nowin32"], |
| 4762 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4763 | ) |
| 4764 | |
| 4765 | xnnpack_benchmark( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4766 | name = "max_pooling_bench", |
| 4767 | srcs = ["bench/max-pooling.cc"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4768 | deps = OPERATOR_BENCHMARK_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4769 | ) |
| 4770 | |
| 4771 | xnnpack_benchmark( |
| 4772 | name = "sigmoid_bench", |
| 4773 | srcs = ["bench/sigmoid.cc"], |
| Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 4774 | copts = xnnpack_optional_tflite_copts(), |
| Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 4775 | tags = ["nowin32"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4776 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4777 | ) |
| 4778 | |
| 4779 | xnnpack_benchmark( |
| Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 4780 | name = "prelu_bench", |
| 4781 | srcs = ["bench/prelu.cc"], |
| 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 | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 4785 | ) |
| 4786 | |
| 4787 | xnnpack_benchmark( |
| Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4788 | name = "softmax_bench", |
| 4789 | srcs = ["bench/softmax.cc"], |
| Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame] | 4790 | copts = xnnpack_optional_tflite_copts(), |
| Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 4791 | tags = ["nowin32"], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4792 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4793 | ) |
| 4794 | |
| Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4795 | xnnpack_benchmark( |
| Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 4796 | name = "square_root_bench", |
| 4797 | srcs = ["bench/square-root.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 | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4804 | name = "truncation_bench", |
| 4805 | srcs = ["bench/truncation.cc"], |
| 4806 | deps = OPERATOR_BENCHMARK_DEPS, |
| 4807 | ) |
| 4808 | |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4809 | ############################# End-to-end benchmarks ############################ |
| 4810 | |
| 4811 | cc_library( |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4812 | name = "fp32_mobilenet_v1", |
| 4813 | srcs = ["models/fp32-mobilenet-v1.cc"], |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4814 | hdrs = ["models/models.h"], |
| Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4815 | copts = xnnpack_std_cxxopts(), |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4816 | linkstatic = True, |
| 4817 | deps = [ |
| 4818 | ":XNNPACK", |
| 4819 | "@pthreadpool", |
| 4820 | ], |
| 4821 | ) |
| 4822 | |
| 4823 | cc_library( |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4824 | name = "fp16_mobilenet_v1", |
| 4825 | srcs = ["models/fp16-mobilenet-v1.cc"], |
| 4826 | hdrs = ["models/models.h"], |
| 4827 | copts = xnnpack_std_cxxopts(), |
| 4828 | linkstatic = True, |
| 4829 | deps = [ |
| 4830 | ":XNNPACK", |
| 4831 | "@FP16", |
| 4832 | "@pthreadpool", |
| 4833 | ], |
| 4834 | ) |
| 4835 | |
| 4836 | cc_library( |
| Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 4837 | name = "qs8_mobilenet_v1", |
| 4838 | srcs = ["models/qs8-mobilenet-v1.cc"], |
| 4839 | hdrs = ["models/models.h"], |
| 4840 | copts = xnnpack_std_cxxopts(), |
| 4841 | linkstatic = True, |
| 4842 | deps = [ |
| 4843 | ":XNNPACK", |
| 4844 | "@pthreadpool", |
| 4845 | ], |
| 4846 | ) |
| 4847 | |
| 4848 | cc_library( |
| Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 4849 | name = "qs8_mobilenet_v2", |
| 4850 | srcs = ["models/qs8-mobilenet-v2.cc"], |
| 4851 | hdrs = ["models/models.h"], |
| 4852 | copts = xnnpack_std_cxxopts(), |
| 4853 | linkstatic = True, |
| 4854 | deps = [ |
| 4855 | ":XNNPACK", |
| 4856 | "@pthreadpool", |
| 4857 | ], |
| 4858 | ) |
| 4859 | |
| 4860 | cc_library( |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4861 | name = "fp32_mobilenet_v2", |
| 4862 | srcs = ["models/fp32-mobilenet-v2.cc"], |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4863 | hdrs = ["models/models.h"], |
| Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4864 | copts = xnnpack_std_cxxopts(), |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4865 | linkstatic = True, |
| 4866 | deps = [ |
| 4867 | ":XNNPACK", |
| 4868 | "@pthreadpool", |
| 4869 | ], |
| 4870 | ) |
| 4871 | |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4872 | cc_library( |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4873 | name = "fp16_mobilenet_v2", |
| 4874 | srcs = ["models/fp16-mobilenet-v2.cc"], |
| 4875 | hdrs = ["models/models.h"], |
| 4876 | copts = xnnpack_std_cxxopts(), |
| 4877 | linkstatic = True, |
| 4878 | deps = [ |
| 4879 | ":XNNPACK", |
| 4880 | "@FP16", |
| 4881 | "@pthreadpool", |
| 4882 | ], |
| 4883 | ) |
| 4884 | |
| 4885 | cc_library( |
| 4886 | name = "fp32_mobilenet_v3_large", |
| 4887 | srcs = ["models/fp32-mobilenet-v3-large.cc"], |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4888 | hdrs = ["models/models.h"], |
| Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4889 | copts = xnnpack_std_cxxopts(), |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4890 | linkstatic = True, |
| 4891 | deps = [ |
| 4892 | ":XNNPACK", |
| 4893 | "@pthreadpool", |
| 4894 | ], |
| 4895 | ) |
| 4896 | |
| 4897 | cc_library( |
| Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4898 | name = "fp16_mobilenet_v3_large", |
| 4899 | srcs = ["models/fp16-mobilenet-v3-large.cc"], |
| 4900 | hdrs = ["models/models.h"], |
| 4901 | copts = xnnpack_std_cxxopts(), |
| 4902 | linkstatic = True, |
| 4903 | deps = [ |
| 4904 | ":XNNPACK", |
| 4905 | "@FP16", |
| 4906 | "@pthreadpool", |
| 4907 | ], |
| 4908 | ) |
| 4909 | |
| 4910 | cc_library( |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4911 | name = "fp32_mobilenet_v3_small", |
| 4912 | srcs = ["models/fp32-mobilenet-v3-small.cc"], |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4913 | hdrs = ["models/models.h"], |
| Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4914 | copts = xnnpack_std_cxxopts(), |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4915 | linkstatic = True, |
| 4916 | deps = [ |
| 4917 | ":XNNPACK", |
| 4918 | "@pthreadpool", |
| 4919 | ], |
| 4920 | ) |
| 4921 | |
| Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4922 | cc_library( |
| 4923 | name = "fp16_mobilenet_v3_small", |
| 4924 | srcs = ["models/fp16-mobilenet-v3-small.cc"], |
| 4925 | hdrs = ["models/models.h"], |
| 4926 | copts = xnnpack_std_cxxopts(), |
| 4927 | linkstatic = True, |
| 4928 | deps = [ |
| 4929 | ":XNNPACK", |
| 4930 | "@FP16", |
| 4931 | "@pthreadpool", |
| 4932 | ], |
| 4933 | ) |
| 4934 | |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4935 | xnnpack_benchmark( |
| Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 4936 | name = "f32_dwconv_e2e_bench", |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4937 | srcs = [ |
| 4938 | "bench/f32-dwconv-e2e.cc", |
| 4939 | "bench/end2end.h", |
| 4940 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 4941 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4942 | ":XNNPACK", |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4943 | ":fp32_mobilenet_v1", |
| 4944 | ":fp32_mobilenet_v2", |
| 4945 | ":fp32_mobilenet_v3_large", |
| 4946 | ":fp32_mobilenet_v3_small", |
| Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 4947 | ], |
| 4948 | ) |
| 4949 | |
| 4950 | xnnpack_benchmark( |
| Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4951 | name = "f32_gemm_e2e_bench", |
| Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4952 | srcs = [ |
| 4953 | "bench/f32-gemm-e2e.cc", |
| 4954 | "bench/end2end.h", |
| 4955 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4956 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4957 | ":XNNPACK", |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4958 | ":fp32_mobilenet_v1", |
| 4959 | ":fp32_mobilenet_v2", |
| 4960 | ":fp32_mobilenet_v3_large", |
| 4961 | ":fp32_mobilenet_v3_small", |
| Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4962 | ], |
| 4963 | ) |
| 4964 | |
| 4965 | xnnpack_benchmark( |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4966 | name = "end2end_bench", |
| 4967 | srcs = ["bench/end2end.cc"], |
| 4968 | deps = [ |
| 4969 | ":XNNPACK", |
| Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 4970 | ":bench_utils", |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4971 | ":fp16_mobilenet_v1", |
| 4972 | ":fp16_mobilenet_v2", |
| Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4973 | ":fp16_mobilenet_v3_large", |
| 4974 | ":fp16_mobilenet_v3_small", |
| Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4975 | ":fp32_mobilenet_v1", |
| 4976 | ":fp32_mobilenet_v2", |
| 4977 | ":fp32_mobilenet_v3_large", |
| 4978 | ":fp32_mobilenet_v3_small", |
| Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 4979 | ":qs8_mobilenet_v1", |
| Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 4980 | ":qs8_mobilenet_v2", |
| Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4981 | "@pthreadpool", |
| 4982 | ], |
| 4983 | ) |
| 4984 | |
| Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 4985 | #################### Accuracy evaluation for math functions #################### |
| 4986 | |
| 4987 | xnnpack_benchmark( |
| 4988 | name = "f32_exp_eval", |
| 4989 | srcs = [ |
| 4990 | "eval/f32-exp.cc", |
| 4991 | "src/xnnpack/AlignedAllocator.h", |
| 4992 | ] + ACCURACY_EVAL_HDRS, |
| 4993 | deps = ACCURACY_EVAL_DEPS, |
| 4994 | ) |
| 4995 | |
| Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 4996 | xnnpack_benchmark( |
| 4997 | name = "f32_expminus_eval", |
| 4998 | srcs = [ |
| 4999 | "eval/f32-expminus.cc", |
| 5000 | "src/xnnpack/AlignedAllocator.h", |
| 5001 | ] + ACCURACY_EVAL_HDRS, |
| 5002 | deps = ACCURACY_EVAL_DEPS, |
| 5003 | ) |
| 5004 | |
| Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 5005 | xnnpack_benchmark( |
| 5006 | name = "f32_extexp_eval", |
| 5007 | srcs = [ |
| 5008 | "eval/f32-extexp.cc", |
| 5009 | "src/xnnpack/AlignedAllocator.h", |
| 5010 | ] + ACCURACY_EVAL_HDRS, |
| 5011 | deps = ACCURACY_EVAL_DEPS, |
| 5012 | ) |
| 5013 | |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 5014 | xnnpack_unit_test( |
| 5015 | name = "f32_roundne_eval", |
| 5016 | srcs = [ |
| 5017 | "eval/f32-roundne.cc", |
| 5018 | "src/xnnpack/AlignedAllocator.h", |
| 5019 | "src/xnnpack/math-stubs.h", |
| 5020 | ] + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | 22eed3d | 2020-05-11 20:13:37 -0700 | [diff] [blame] | 5021 | automatic = False, |
| Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 5022 | deps = MICROKERNEL_TEST_DEPS, |
| 5023 | ) |
| 5024 | |
| Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 5025 | xnnpack_unit_test( |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 5026 | name = "f32_roundd_eval", |
| 5027 | srcs = [ |
| 5028 | "eval/f32-roundd.cc", |
| 5029 | "src/xnnpack/AlignedAllocator.h", |
| 5030 | "src/xnnpack/math-stubs.h", |
| 5031 | ] + MICROKERNEL_TEST_HDRS, |
| 5032 | automatic = False, |
| 5033 | deps = MICROKERNEL_TEST_DEPS, |
| 5034 | ) |
| 5035 | |
| 5036 | xnnpack_unit_test( |
| 5037 | name = "f32_roundu_eval", |
| 5038 | srcs = [ |
| 5039 | "eval/f32-roundu.cc", |
| 5040 | "src/xnnpack/AlignedAllocator.h", |
| 5041 | "src/xnnpack/math-stubs.h", |
| 5042 | ] + MICROKERNEL_TEST_HDRS, |
| 5043 | automatic = False, |
| 5044 | deps = MICROKERNEL_TEST_DEPS, |
| 5045 | ) |
| 5046 | |
| 5047 | xnnpack_unit_test( |
| Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 5048 | name = "f32_roundz_eval", |
| 5049 | srcs = [ |
| 5050 | "eval/f32-roundz.cc", |
| 5051 | "src/xnnpack/AlignedAllocator.h", |
| 5052 | "src/xnnpack/math-stubs.h", |
| 5053 | ] + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 5054 | automatic = False, |
| Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 5055 | deps = MICROKERNEL_TEST_DEPS, |
| 5056 | ) |
| 5057 | |
| Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5058 | xnnpack_benchmark( |
| 5059 | name = "f32_sigmoid_eval", |
| 5060 | srcs = [ |
| 5061 | "eval/f32-sigmoid.cc", |
| 5062 | "src/xnnpack/AlignedAllocator.h", |
| 5063 | ] + ACCURACY_EVAL_HDRS, |
| 5064 | deps = ACCURACY_EVAL_DEPS, |
| 5065 | ) |
| 5066 | |
| Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 5067 | xnnpack_benchmark( |
| 5068 | name = "f32_sqrt_eval", |
| 5069 | srcs = [ |
| 5070 | "eval/f32-sqrt.cc", |
| 5071 | "src/xnnpack/AlignedAllocator.h", |
| 5072 | ] + ACCURACY_EVAL_HDRS, |
| 5073 | deps = ACCURACY_EVAL_DEPS + [ |
| 5074 | ":bench_utils", |
| 5075 | ], |
| 5076 | ) |
| 5077 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5078 | ######################### Unit tests for micro-kernels ######################### |
| 5079 | |
| 5080 | xnnpack_unit_test( |
| Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 5081 | name = "f16_gemm_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5082 | srcs = [ |
| Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 5083 | "test/f16-gemm-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5084 | "test/gemm-microkernel-tester.h", |
| 5085 | "src/xnnpack/AlignedAllocator.h", |
| 5086 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5087 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5088 | ) |
| 5089 | |
| 5090 | xnnpack_unit_test( |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5091 | name = "f16_spmm_minmax_test", |
| Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 5092 | srcs = [ |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5093 | "test/f16-spmm-minmax.cc", |
| Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 5094 | "test/spmm-microkernel-tester.h", |
| 5095 | "src/xnnpack/AlignedAllocator.h", |
| 5096 | ] + MICROKERNEL_TEST_HDRS, |
| 5097 | deps = MICROKERNEL_TEST_DEPS, |
| 5098 | ) |
| 5099 | |
| 5100 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5101 | name = "f32_argmaxpool_test", |
| 5102 | srcs = [ |
| 5103 | "test/f32-argmaxpool.cc", |
| 5104 | "test/argmaxpool-microkernel-tester.h", |
| 5105 | "src/xnnpack/AlignedAllocator.h", |
| 5106 | ] + MICROKERNEL_TEST_HDRS, |
| 5107 | deps = MICROKERNEL_TEST_DEPS, |
| 5108 | ) |
| 5109 | |
| 5110 | xnnpack_unit_test( |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5111 | name = "f32_avgpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5112 | srcs = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5113 | "test/f32-avgpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5114 | "test/avgpool-microkernel-tester.h", |
| 5115 | "src/xnnpack/AlignedAllocator.h", |
| 5116 | ] + MICROKERNEL_TEST_HDRS, |
| 5117 | deps = MICROKERNEL_TEST_DEPS, |
| 5118 | ) |
| 5119 | |
| 5120 | xnnpack_unit_test( |
| Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 5121 | name = "f32_ibilinear_test", |
| Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 5122 | srcs = [ |
| Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 5123 | "test/f32-ibilinear.cc", |
| 5124 | "test/ibilinear-microkernel-tester.h", |
| Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 5125 | "src/xnnpack/AlignedAllocator.h", |
| 5126 | ] + MICROKERNEL_TEST_HDRS, |
| 5127 | deps = MICROKERNEL_TEST_DEPS, |
| 5128 | ) |
| 5129 | |
| 5130 | xnnpack_unit_test( |
| XNNPACK Team | 2143267 | 2020-10-19 19:58:48 -0700 | [diff] [blame] | 5131 | name = "f32_ibilinear_chw_test", |
| 5132 | srcs = [ |
| 5133 | "test/f32-ibilinear-chw.cc", |
| XNNPACK Team | 6be46b2 | 2020-10-22 23:34:54 -0700 | [diff] [blame] | 5134 | "test/ibilinear-microkernel-tester.h", |
| XNNPACK Team | 2143267 | 2020-10-19 19:58:48 -0700 | [diff] [blame] | 5135 | "src/xnnpack/AlignedAllocator.h", |
| 5136 | ] + MICROKERNEL_TEST_HDRS, |
| 5137 | deps = MICROKERNEL_TEST_DEPS, |
| 5138 | ) |
| 5139 | |
| 5140 | xnnpack_unit_test( |
| Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5141 | name = "f16_clamp_test", |
| 5142 | srcs = [ |
| 5143 | "test/f16-clamp.cc", |
| 5144 | "test/clamp-microkernel-tester.h", |
| 5145 | ] + MICROKERNEL_TEST_HDRS, |
| 5146 | deps = MICROKERNEL_TEST_DEPS, |
| 5147 | ) |
| 5148 | |
| 5149 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5150 | name = "f32_clamp_test", |
| 5151 | srcs = [ |
| 5152 | "test/f32-clamp.cc", |
| 5153 | "test/clamp-microkernel-tester.h", |
| 5154 | ] + MICROKERNEL_TEST_HDRS, |
| 5155 | deps = MICROKERNEL_TEST_DEPS, |
| 5156 | ) |
| 5157 | |
| 5158 | xnnpack_unit_test( |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5159 | name = "f32_igemm_test", |
| 5160 | srcs = [ |
| 5161 | "test/f32-igemm.cc", |
| 5162 | "test/gemm-microkernel-tester.h", |
| 5163 | "src/xnnpack/AlignedAllocator.h", |
| 5164 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5165 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5166 | ) |
| 5167 | |
| 5168 | xnnpack_unit_test( |
| Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5169 | name = "f32_igemm_relu_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5170 | srcs = [ |
| Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5171 | "test/f32-igemm-relu.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5172 | "test/gemm-microkernel-tester.h", |
| 5173 | "src/xnnpack/AlignedAllocator.h", |
| 5174 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5175 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5176 | ) |
| 5177 | |
| 5178 | xnnpack_unit_test( |
| Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 5179 | name = "f32_igemm_minmax_test", |
| 5180 | srcs = [ |
| 5181 | "test/f32-igemm-minmax.cc", |
| 5182 | "test/gemm-microkernel-tester.h", |
| 5183 | "src/xnnpack/AlignedAllocator.h", |
| 5184 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5185 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 5186 | ) |
| 5187 | |
| 5188 | xnnpack_unit_test( |
| Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 5189 | name = "f16_igemm_minmax_test", |
| 5190 | srcs = [ |
| 5191 | "test/f16-igemm-minmax.cc", |
| 5192 | "test/gemm-microkernel-tester.h", |
| 5193 | "src/xnnpack/AlignedAllocator.h", |
| 5194 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5195 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 5196 | ) |
| 5197 | |
| 5198 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5199 | name = "f32_conv_hwc_test", |
| 5200 | srcs = [ |
| 5201 | "test/f32-conv-hwc.cc", |
| 5202 | "test/conv-hwc-microkernel-tester.h", |
| 5203 | "src/xnnpack/AlignedAllocator.h", |
| 5204 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5205 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5206 | ) |
| 5207 | |
| 5208 | xnnpack_unit_test( |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5209 | name = "f32_conv_hwc2chw_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5210 | srcs = [ |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5211 | "test/f32-conv-hwc2chw.cc", |
| 5212 | "test/conv-hwc2chw-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5213 | "src/xnnpack/AlignedAllocator.h", |
| 5214 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5215 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5216 | ) |
| 5217 | |
| 5218 | xnnpack_unit_test( |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5219 | name = "f32_dwconv_test", |
| 5220 | srcs = [ |
| 5221 | "test/f32-dwconv.cc", |
| 5222 | "test/dwconv-microkernel-tester.h", |
| 5223 | "src/xnnpack/AlignedAllocator.h", |
| 5224 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5225 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5226 | ) |
| 5227 | |
| 5228 | xnnpack_unit_test( |
| Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 5229 | name = "f16_dwconv_minmax_test", |
| 5230 | srcs = [ |
| 5231 | "test/f16-dwconv-minmax.cc", |
| 5232 | "test/dwconv-microkernel-tester.h", |
| 5233 | "src/xnnpack/AlignedAllocator.h", |
| 5234 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5235 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 5236 | ) |
| 5237 | |
| 5238 | xnnpack_unit_test( |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5239 | name = "f32_dwconv_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5240 | srcs = [ |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5241 | "test/f32-dwconv-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5242 | "test/dwconv-microkernel-tester.h", |
| 5243 | "src/xnnpack/AlignedAllocator.h", |
| 5244 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5245 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5246 | ) |
| 5247 | |
| 5248 | xnnpack_unit_test( |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 5249 | name = "f32_dwconv2d_chw_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5250 | srcs = [ |
| Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 5251 | "test/f32-dwconv2d-chw.cc", |
| 5252 | "test/dwconv2d-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5253 | "src/xnnpack/AlignedAllocator.h", |
| 5254 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5255 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5256 | ) |
| 5257 | |
| 5258 | xnnpack_unit_test( |
| Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 5259 | name = "f16_gavgpool_minmax_test", |
| 5260 | srcs = [ |
| 5261 | "test/f16-gavgpool-minmax.cc", |
| 5262 | "test/gavgpool-microkernel-tester.h", |
| 5263 | "src/xnnpack/AlignedAllocator.h", |
| 5264 | ] + MICROKERNEL_TEST_HDRS, |
| 5265 | deps = MICROKERNEL_TEST_DEPS, |
| 5266 | ) |
| 5267 | |
| 5268 | xnnpack_unit_test( |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5269 | name = "f32_gavgpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5270 | srcs = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5271 | "test/f32-gavgpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5272 | "test/gavgpool-microkernel-tester.h", |
| 5273 | "src/xnnpack/AlignedAllocator.h", |
| 5274 | ] + MICROKERNEL_TEST_HDRS, |
| 5275 | deps = MICROKERNEL_TEST_DEPS, |
| 5276 | ) |
| 5277 | |
| 5278 | xnnpack_unit_test( |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5279 | name = "f32_gavgpool_cw_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5280 | srcs = [ |
| Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5281 | "test/f32-gavgpool-cw.cc", |
| 5282 | "test/gavgpool-cw-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5283 | "src/xnnpack/AlignedAllocator.h", |
| 5284 | ] + MICROKERNEL_TEST_HDRS, |
| 5285 | deps = MICROKERNEL_TEST_DEPS, |
| 5286 | ) |
| 5287 | |
| 5288 | xnnpack_unit_test( |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5289 | name = "f32_gemm_test", |
| 5290 | srcs = [ |
| 5291 | "test/f32-gemm.cc", |
| 5292 | "test/gemm-microkernel-tester.h", |
| 5293 | "src/xnnpack/AlignedAllocator.h", |
| 5294 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5295 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5296 | ) |
| 5297 | |
| 5298 | xnnpack_unit_test( |
| Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5299 | name = "f32_gemm_relu_test", |
| 5300 | srcs = [ |
| 5301 | "test/f32-gemm-relu.cc", |
| 5302 | "test/gemm-microkernel-tester.h", |
| 5303 | "src/xnnpack/AlignedAllocator.h", |
| 5304 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5305 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5306 | ) |
| 5307 | |
| 5308 | xnnpack_unit_test( |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5309 | name = "f32_gemm_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5310 | srcs = [ |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5311 | "test/f32-gemm-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5312 | "test/gemm-microkernel-tester.h", |
| 5313 | "src/xnnpack/AlignedAllocator.h", |
| 5314 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5315 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5316 | ) |
| 5317 | |
| 5318 | xnnpack_unit_test( |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5319 | name = "f32_gemminc_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5320 | srcs = [ |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5321 | "test/f32-gemminc-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5322 | "test/gemm-microkernel-tester.h", |
| 5323 | "src/xnnpack/AlignedAllocator.h", |
| 5324 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5325 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5326 | ) |
| 5327 | |
| 5328 | xnnpack_unit_test( |
| Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5329 | name = "f16_hswish_test", |
| 5330 | srcs = [ |
| 5331 | "test/f16-hswish.cc", |
| 5332 | "test/hswish-microkernel-tester.h", |
| 5333 | ] + MICROKERNEL_TEST_HDRS, |
| 5334 | deps = MICROKERNEL_TEST_DEPS, |
| 5335 | ) |
| 5336 | |
| 5337 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5338 | name = "f32_hswish_test", |
| 5339 | srcs = [ |
| 5340 | "test/f32-hswish.cc", |
| 5341 | "test/hswish-microkernel-tester.h", |
| 5342 | ] + MICROKERNEL_TEST_HDRS, |
| 5343 | deps = MICROKERNEL_TEST_DEPS, |
| 5344 | ) |
| 5345 | |
| 5346 | xnnpack_unit_test( |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5347 | name = "f32_maxpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5348 | srcs = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5349 | "test/f32-maxpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5350 | "test/maxpool-microkernel-tester.h", |
| 5351 | ] + MICROKERNEL_TEST_HDRS, |
| 5352 | deps = MICROKERNEL_TEST_DEPS, |
| 5353 | ) |
| 5354 | |
| 5355 | xnnpack_unit_test( |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5356 | name = "f32_pavgpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5357 | srcs = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5358 | "test/f32-pavgpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5359 | "test/avgpool-microkernel-tester.h", |
| 5360 | "src/xnnpack/AlignedAllocator.h", |
| 5361 | ] + MICROKERNEL_TEST_HDRS, |
| 5362 | deps = MICROKERNEL_TEST_DEPS, |
| 5363 | ) |
| 5364 | |
| 5365 | xnnpack_unit_test( |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5366 | name = "f32_ppmm_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5367 | srcs = [ |
| Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5368 | "test/f32-ppmm-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5369 | "test/gemm-microkernel-tester.h", |
| 5370 | "src/xnnpack/AlignedAllocator.h", |
| 5371 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5372 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5373 | ) |
| 5374 | |
| 5375 | xnnpack_unit_test( |
| Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5376 | name = "f16_prelu_test", |
| 5377 | srcs = [ |
| 5378 | "test/f16-prelu.cc", |
| 5379 | "test/prelu-microkernel-tester.h", |
| 5380 | "src/xnnpack/AlignedAllocator.h", |
| 5381 | ] + MICROKERNEL_TEST_HDRS, |
| 5382 | deps = MICROKERNEL_TEST_DEPS, |
| 5383 | ) |
| 5384 | |
| 5385 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5386 | name = "f32_prelu_test", |
| 5387 | srcs = [ |
| 5388 | "test/f32-prelu.cc", |
| 5389 | "test/prelu-microkernel-tester.h", |
| 5390 | "src/xnnpack/AlignedAllocator.h", |
| 5391 | ] + MICROKERNEL_TEST_HDRS, |
| 5392 | deps = MICROKERNEL_TEST_DEPS, |
| 5393 | ) |
| 5394 | |
| 5395 | xnnpack_unit_test( |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5396 | name = "f32_raddexpminusmax_test", |
| 5397 | srcs = [ |
| 5398 | "test/f32-raddexpminusmax.cc", |
| 5399 | "test/raddexpminusmax-microkernel-tester.h", |
| 5400 | ] + MICROKERNEL_TEST_HDRS, |
| 5401 | deps = MICROKERNEL_TEST_DEPS, |
| 5402 | ) |
| 5403 | |
| 5404 | xnnpack_unit_test( |
| Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 5405 | name = "f32_raddextexp_test", |
| 5406 | srcs = [ |
| 5407 | "test/f32-raddextexp.cc", |
| 5408 | "test/raddextexp-microkernel-tester.h", |
| 5409 | ] + MICROKERNEL_TEST_HDRS, |
| 5410 | deps = MICROKERNEL_TEST_DEPS, |
| 5411 | ) |
| 5412 | |
| 5413 | xnnpack_unit_test( |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5414 | name = "f32_raddstoreexpminusmax_test", |
| 5415 | srcs = [ |
| 5416 | "test/f32-raddstoreexpminusmax.cc", |
| 5417 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 5418 | ] + MICROKERNEL_TEST_HDRS, |
| 5419 | deps = MICROKERNEL_TEST_DEPS, |
| 5420 | ) |
| 5421 | |
| 5422 | xnnpack_unit_test( |
| Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 5423 | name = "f32_relu_test", |
| 5424 | srcs = [ |
| 5425 | "test/f32-relu.cc", |
| 5426 | "test/vunary-microkernel-tester.h", |
| 5427 | ] + MICROKERNEL_TEST_HDRS, |
| 5428 | deps = MICROKERNEL_TEST_DEPS, |
| 5429 | ) |
| 5430 | |
| 5431 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5432 | name = "f32_rmax_test", |
| 5433 | srcs = [ |
| 5434 | "test/f32-rmax.cc", |
| 5435 | "test/rmax-microkernel-tester.h", |
| 5436 | ] + MICROKERNEL_TEST_HDRS, |
| 5437 | deps = MICROKERNEL_TEST_DEPS, |
| 5438 | ) |
| 5439 | |
| 5440 | xnnpack_unit_test( |
| Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5441 | name = "f32_sigmoid_test", |
| 5442 | srcs = [ |
| 5443 | "test/f32-sigmoid.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5444 | "test/vunary-microkernel-tester.h", |
| Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5445 | ] + MICROKERNEL_TEST_HDRS, |
| 5446 | deps = MICROKERNEL_TEST_DEPS, |
| 5447 | ) |
| 5448 | |
| 5449 | xnnpack_unit_test( |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5450 | name = "f32_spmm_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5451 | srcs = [ |
| Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5452 | "test/f32-spmm-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5453 | "test/spmm-microkernel-tester.h", |
| 5454 | "src/xnnpack/AlignedAllocator.h", |
| 5455 | ] + MICROKERNEL_TEST_HDRS, |
| 5456 | deps = MICROKERNEL_TEST_DEPS, |
| 5457 | ) |
| 5458 | |
| 5459 | xnnpack_unit_test( |
| Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 5460 | name = "f16_vadd_minmax_test", |
| 5461 | srcs = [ |
| 5462 | "test/f16-vadd-minmax.cc", |
| 5463 | "test/vbinary-microkernel-tester.h", |
| 5464 | ] + MICROKERNEL_TEST_HDRS, |
| 5465 | deps = MICROKERNEL_TEST_DEPS, |
| 5466 | ) |
| 5467 | |
| 5468 | xnnpack_unit_test( |
| 5469 | name = "f16_vaddc_minmax_test", |
| 5470 | srcs = [ |
| 5471 | "test/f16-vaddc-minmax.cc", |
| 5472 | "test/vbinaryc-microkernel-tester.h", |
| 5473 | ] + MICROKERNEL_TEST_HDRS, |
| 5474 | deps = MICROKERNEL_TEST_DEPS, |
| 5475 | ) |
| 5476 | |
| 5477 | xnnpack_unit_test( |
| 5478 | name = "f16_vdiv_minmax_test", |
| 5479 | srcs = [ |
| 5480 | "test/f16-vdiv-minmax.cc", |
| 5481 | "test/vbinary-microkernel-tester.h", |
| 5482 | ] + MICROKERNEL_TEST_HDRS, |
| 5483 | deps = MICROKERNEL_TEST_DEPS, |
| 5484 | ) |
| 5485 | |
| 5486 | xnnpack_unit_test( |
| 5487 | name = "f16_vdivc_minmax_test", |
| 5488 | srcs = [ |
| 5489 | "test/f16-vdivc-minmax.cc", |
| 5490 | "test/vbinaryc-microkernel-tester.h", |
| 5491 | ] + MICROKERNEL_TEST_HDRS, |
| 5492 | deps = MICROKERNEL_TEST_DEPS, |
| 5493 | ) |
| 5494 | |
| 5495 | xnnpack_unit_test( |
| 5496 | name = "f16_vrdivc_minmax_test", |
| 5497 | srcs = [ |
| 5498 | "test/f16-vrdivc-minmax.cc", |
| 5499 | "test/vbinaryc-microkernel-tester.h", |
| 5500 | ] + MICROKERNEL_TEST_HDRS, |
| 5501 | deps = MICROKERNEL_TEST_DEPS, |
| 5502 | ) |
| 5503 | |
| 5504 | xnnpack_unit_test( |
| 5505 | name = "f16_vmax_test", |
| 5506 | srcs = [ |
| 5507 | "test/f16-vmax.cc", |
| 5508 | "test/vbinary-microkernel-tester.h", |
| 5509 | ] + MICROKERNEL_TEST_HDRS, |
| 5510 | deps = MICROKERNEL_TEST_DEPS, |
| 5511 | ) |
| 5512 | |
| 5513 | xnnpack_unit_test( |
| 5514 | name = "f16_vmaxc_test", |
| 5515 | srcs = [ |
| 5516 | "test/f16-vmaxc.cc", |
| 5517 | "test/vbinaryc-microkernel-tester.h", |
| 5518 | ] + MICROKERNEL_TEST_HDRS, |
| 5519 | deps = MICROKERNEL_TEST_DEPS, |
| 5520 | ) |
| 5521 | |
| 5522 | xnnpack_unit_test( |
| 5523 | name = "f16_vmin_test", |
| 5524 | srcs = [ |
| 5525 | "test/f16-vmin.cc", |
| 5526 | "test/vbinary-microkernel-tester.h", |
| 5527 | ] + MICROKERNEL_TEST_HDRS, |
| 5528 | deps = MICROKERNEL_TEST_DEPS, |
| 5529 | ) |
| 5530 | |
| 5531 | xnnpack_unit_test( |
| 5532 | name = "f16_vminc_test", |
| 5533 | srcs = [ |
| 5534 | "test/f16-vminc.cc", |
| 5535 | "test/vbinaryc-microkernel-tester.h", |
| 5536 | ] + MICROKERNEL_TEST_HDRS, |
| 5537 | deps = MICROKERNEL_TEST_DEPS, |
| 5538 | ) |
| 5539 | |
| 5540 | xnnpack_unit_test( |
| 5541 | name = "f16_vmul_minmax_test", |
| 5542 | srcs = [ |
| 5543 | "test/f16-vmul-minmax.cc", |
| 5544 | "test/vbinary-microkernel-tester.h", |
| 5545 | ] + MICROKERNEL_TEST_HDRS, |
| 5546 | deps = MICROKERNEL_TEST_DEPS, |
| 5547 | ) |
| 5548 | |
| 5549 | xnnpack_unit_test( |
| 5550 | name = "f16_vmulc_minmax_test", |
| 5551 | srcs = [ |
| 5552 | "test/f16-vmulc-minmax.cc", |
| 5553 | "test/vbinaryc-microkernel-tester.h", |
| 5554 | ] + MICROKERNEL_TEST_HDRS, |
| 5555 | deps = MICROKERNEL_TEST_DEPS, |
| 5556 | ) |
| 5557 | |
| 5558 | xnnpack_unit_test( |
| 5559 | name = "f16_vsub_minmax_test", |
| 5560 | srcs = [ |
| 5561 | "test/f16-vsub-minmax.cc", |
| 5562 | "test/vbinary-microkernel-tester.h", |
| 5563 | ] + MICROKERNEL_TEST_HDRS, |
| 5564 | deps = MICROKERNEL_TEST_DEPS, |
| 5565 | ) |
| 5566 | |
| 5567 | xnnpack_unit_test( |
| 5568 | name = "f16_vsubc_minmax_test", |
| 5569 | srcs = [ |
| 5570 | "test/f16-vsubc-minmax.cc", |
| 5571 | "test/vbinaryc-microkernel-tester.h", |
| 5572 | ] + MICROKERNEL_TEST_HDRS, |
| 5573 | deps = MICROKERNEL_TEST_DEPS, |
| 5574 | ) |
| 5575 | |
| 5576 | xnnpack_unit_test( |
| 5577 | name = "f16_vrsubc_minmax_test", |
| 5578 | srcs = [ |
| 5579 | "test/f16-vrsubc-minmax.cc", |
| 5580 | "test/vbinaryc-microkernel-tester.h", |
| 5581 | ] + MICROKERNEL_TEST_HDRS, |
| 5582 | deps = MICROKERNEL_TEST_DEPS, |
| 5583 | ) |
| 5584 | |
| 5585 | xnnpack_unit_test( |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5586 | name = "f32_vabs_test", |
| 5587 | srcs = [ |
| 5588 | "test/f32-vabs.cc", |
| 5589 | "test/vunary-microkernel-tester.h", |
| 5590 | ] + MICROKERNEL_TEST_HDRS, |
| 5591 | deps = MICROKERNEL_TEST_DEPS, |
| 5592 | ) |
| 5593 | |
| 5594 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5595 | name = "f32_vadd_test", |
| 5596 | srcs = [ |
| 5597 | "test/f32-vadd.cc", |
| 5598 | "test/vbinary-microkernel-tester.h", |
| 5599 | ] + MICROKERNEL_TEST_HDRS, |
| 5600 | deps = MICROKERNEL_TEST_DEPS, |
| 5601 | ) |
| 5602 | |
| 5603 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5604 | name = "f32_vadd_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5605 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5606 | "test/f32-vadd-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5607 | "test/vbinary-microkernel-tester.h", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5608 | ] + MICROKERNEL_TEST_HDRS, |
| 5609 | deps = MICROKERNEL_TEST_DEPS, |
| 5610 | ) |
| 5611 | |
| 5612 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5613 | name = "f32_vadd_relu_test", |
| 5614 | srcs = [ |
| 5615 | "test/f32-vadd-relu.cc", |
| 5616 | "test/vbinary-microkernel-tester.h", |
| 5617 | ] + MICROKERNEL_TEST_HDRS, |
| 5618 | deps = MICROKERNEL_TEST_DEPS, |
| 5619 | ) |
| 5620 | |
| 5621 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5622 | name = "f32_vaddc_test", |
| 5623 | srcs = [ |
| 5624 | "test/f32-vaddc.cc", |
| 5625 | "test/vbinaryc-microkernel-tester.h", |
| 5626 | ] + MICROKERNEL_TEST_HDRS, |
| 5627 | deps = MICROKERNEL_TEST_DEPS, |
| 5628 | ) |
| 5629 | |
| 5630 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5631 | name = "f32_vaddc_minmax_test", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5632 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5633 | "test/f32-vaddc-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5634 | "test/vbinaryc-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5635 | ] + MICROKERNEL_TEST_HDRS, |
| 5636 | deps = MICROKERNEL_TEST_DEPS, |
| 5637 | ) |
| 5638 | |
| 5639 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5640 | name = "f32_vaddc_relu_test", |
| 5641 | srcs = [ |
| 5642 | "test/f32-vaddc-relu.cc", |
| 5643 | "test/vbinaryc-microkernel-tester.h", |
| 5644 | ] + MICROKERNEL_TEST_HDRS, |
| 5645 | deps = MICROKERNEL_TEST_DEPS, |
| 5646 | ) |
| 5647 | |
| 5648 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5649 | name = "f32_vdiv_test", |
| 5650 | srcs = [ |
| 5651 | "test/f32-vdiv.cc", |
| 5652 | "test/vbinary-microkernel-tester.h", |
| 5653 | ] + MICROKERNEL_TEST_HDRS, |
| 5654 | deps = MICROKERNEL_TEST_DEPS, |
| 5655 | ) |
| 5656 | |
| 5657 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5658 | name = "f32_vdiv_minmax_test", |
| Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5659 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5660 | "test/f32-vdiv-minmax.cc", |
| Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5661 | "test/vbinary-microkernel-tester.h", |
| 5662 | ] + MICROKERNEL_TEST_HDRS, |
| 5663 | deps = MICROKERNEL_TEST_DEPS, |
| 5664 | ) |
| 5665 | |
| 5666 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5667 | name = "f32_vdiv_relu_test", |
| 5668 | srcs = [ |
| 5669 | "test/f32-vdiv-relu.cc", |
| 5670 | "test/vbinary-microkernel-tester.h", |
| 5671 | ] + MICROKERNEL_TEST_HDRS, |
| 5672 | deps = MICROKERNEL_TEST_DEPS, |
| 5673 | ) |
| 5674 | |
| 5675 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5676 | name = "f32_vdivc_test", |
| 5677 | srcs = [ |
| 5678 | "test/f32-vdivc.cc", |
| 5679 | "test/vbinaryc-microkernel-tester.h", |
| 5680 | ] + MICROKERNEL_TEST_HDRS, |
| 5681 | deps = MICROKERNEL_TEST_DEPS, |
| 5682 | ) |
| 5683 | |
| 5684 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5685 | name = "f32_vdivc_minmax_test", |
| Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5686 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5687 | "test/f32-vdivc-minmax.cc", |
| Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5688 | "test/vbinaryc-microkernel-tester.h", |
| 5689 | ] + MICROKERNEL_TEST_HDRS, |
| 5690 | deps = MICROKERNEL_TEST_DEPS, |
| 5691 | ) |
| 5692 | |
| 5693 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5694 | name = "f32_vdivc_relu_test", |
| 5695 | srcs = [ |
| 5696 | "test/f32-vdivc-relu.cc", |
| 5697 | "test/vbinaryc-microkernel-tester.h", |
| 5698 | ] + MICROKERNEL_TEST_HDRS, |
| 5699 | deps = MICROKERNEL_TEST_DEPS, |
| 5700 | ) |
| 5701 | |
| 5702 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5703 | name = "f32_vrdivc_test", |
| 5704 | srcs = [ |
| 5705 | "test/f32-vrdivc.cc", |
| 5706 | "test/vbinaryc-microkernel-tester.h", |
| 5707 | ] + MICROKERNEL_TEST_HDRS, |
| 5708 | deps = MICROKERNEL_TEST_DEPS, |
| 5709 | ) |
| 5710 | |
| 5711 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5712 | name = "f32_vrdivc_minmax_test", |
| Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5713 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5714 | "test/f32-vrdivc-minmax.cc", |
| Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5715 | "test/vbinaryc-microkernel-tester.h", |
| 5716 | ] + MICROKERNEL_TEST_HDRS, |
| 5717 | deps = MICROKERNEL_TEST_DEPS, |
| 5718 | ) |
| 5719 | |
| 5720 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5721 | name = "f32_vrdivc_relu_test", |
| 5722 | srcs = [ |
| 5723 | "test/f32-vrdivc-relu.cc", |
| 5724 | "test/vbinaryc-microkernel-tester.h", |
| 5725 | ] + MICROKERNEL_TEST_HDRS, |
| 5726 | deps = MICROKERNEL_TEST_DEPS, |
| 5727 | ) |
| 5728 | |
| 5729 | xnnpack_unit_test( |
| Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 5730 | name = "f32_vmax_test", |
| 5731 | srcs = [ |
| 5732 | "test/f32-vmax.cc", |
| 5733 | "test/vbinary-microkernel-tester.h", |
| 5734 | ] + MICROKERNEL_TEST_HDRS, |
| 5735 | deps = MICROKERNEL_TEST_DEPS, |
| 5736 | ) |
| 5737 | |
| 5738 | xnnpack_unit_test( |
| 5739 | name = "f32_vmaxc_test", |
| 5740 | srcs = [ |
| 5741 | "test/f32-vmaxc.cc", |
| 5742 | "test/vbinaryc-microkernel-tester.h", |
| 5743 | ] + MICROKERNEL_TEST_HDRS, |
| 5744 | deps = MICROKERNEL_TEST_DEPS, |
| 5745 | ) |
| 5746 | |
| 5747 | xnnpack_unit_test( |
| 5748 | name = "f32_vmin_test", |
| 5749 | srcs = [ |
| 5750 | "test/f32-vmin.cc", |
| 5751 | "test/vbinary-microkernel-tester.h", |
| 5752 | ] + MICROKERNEL_TEST_HDRS, |
| 5753 | deps = MICROKERNEL_TEST_DEPS, |
| 5754 | ) |
| 5755 | |
| 5756 | xnnpack_unit_test( |
| 5757 | name = "f32_vminc_test", |
| 5758 | srcs = [ |
| 5759 | "test/f32-vminc.cc", |
| 5760 | "test/vbinaryc-microkernel-tester.h", |
| 5761 | ] + MICROKERNEL_TEST_HDRS, |
| 5762 | deps = MICROKERNEL_TEST_DEPS, |
| 5763 | ) |
| 5764 | |
| 5765 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5766 | name = "f32_vmul_test", |
| 5767 | srcs = [ |
| 5768 | "test/f32-vmul.cc", |
| 5769 | "test/vbinary-microkernel-tester.h", |
| 5770 | ] + MICROKERNEL_TEST_HDRS, |
| 5771 | deps = MICROKERNEL_TEST_DEPS, |
| 5772 | ) |
| 5773 | |
| 5774 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5775 | name = "f32_vmul_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5776 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5777 | "test/f32-vmul-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5778 | "test/vbinary-microkernel-tester.h", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5779 | ] + MICROKERNEL_TEST_HDRS, |
| 5780 | deps = MICROKERNEL_TEST_DEPS, |
| 5781 | ) |
| 5782 | |
| 5783 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5784 | name = "f32_vmul_relu_test", |
| 5785 | srcs = [ |
| 5786 | "test/f32-vmul-relu.cc", |
| 5787 | "test/vbinary-microkernel-tester.h", |
| 5788 | ] + MICROKERNEL_TEST_HDRS, |
| 5789 | deps = MICROKERNEL_TEST_DEPS, |
| 5790 | ) |
| 5791 | |
| 5792 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5793 | name = "f32_vmulc_test", |
| 5794 | srcs = [ |
| 5795 | "test/f32-vmulc.cc", |
| 5796 | "test/vbinaryc-microkernel-tester.h", |
| 5797 | ] + MICROKERNEL_TEST_HDRS, |
| 5798 | deps = MICROKERNEL_TEST_DEPS, |
| 5799 | ) |
| 5800 | |
| 5801 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5802 | name = "f32_vmulc_minmax_test", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5803 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5804 | "test/f32-vmulc-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5805 | "test/vbinaryc-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5806 | ] + MICROKERNEL_TEST_HDRS, |
| 5807 | deps = MICROKERNEL_TEST_DEPS, |
| 5808 | ) |
| 5809 | |
| 5810 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5811 | name = "f32_vmulc_relu_test", |
| 5812 | srcs = [ |
| 5813 | "test/f32-vmulc-relu.cc", |
| 5814 | "test/vbinaryc-microkernel-tester.h", |
| 5815 | ] + MICROKERNEL_TEST_HDRS, |
| 5816 | deps = MICROKERNEL_TEST_DEPS, |
| 5817 | ) |
| 5818 | |
| 5819 | xnnpack_unit_test( |
| Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 5820 | name = "f16_vmulcaddc_minmax_test", |
| 5821 | srcs = [ |
| 5822 | "test/f16-vmulcaddc-minmax.cc", |
| 5823 | "test/vmulcaddc-microkernel-tester.h", |
| 5824 | "src/xnnpack/AlignedAllocator.h", |
| 5825 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5826 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 5827 | ) |
| 5828 | |
| 5829 | xnnpack_unit_test( |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5830 | name = "f32_vmulcaddc_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5831 | srcs = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5832 | "test/f32-vmulcaddc-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5833 | "test/vmulcaddc-microkernel-tester.h", |
| 5834 | "src/xnnpack/AlignedAllocator.h", |
| 5835 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5836 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5837 | ) |
| 5838 | |
| 5839 | xnnpack_unit_test( |
| Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 5840 | name = "f32_vlrelu_test", |
| 5841 | srcs = [ |
| 5842 | "test/f32-vlrelu.cc", |
| 5843 | "test/vunary-microkernel-tester.h", |
| 5844 | ] + MICROKERNEL_TEST_HDRS, |
| 5845 | deps = MICROKERNEL_TEST_DEPS, |
| 5846 | ) |
| 5847 | |
| 5848 | xnnpack_unit_test( |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5849 | name = "f32_vneg_test", |
| 5850 | srcs = [ |
| 5851 | "test/f32-vneg.cc", |
| 5852 | "test/vunary-microkernel-tester.h", |
| 5853 | ] + MICROKERNEL_TEST_HDRS, |
| 5854 | deps = MICROKERNEL_TEST_DEPS, |
| 5855 | ) |
| 5856 | |
| 5857 | xnnpack_unit_test( |
| Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 5858 | name = "f32_vrndne_test", |
| 5859 | srcs = [ |
| 5860 | "test/f32-vrndne.cc", |
| 5861 | "test/vunary-microkernel-tester.h", |
| 5862 | ] + MICROKERNEL_TEST_HDRS, |
| 5863 | deps = MICROKERNEL_TEST_DEPS, |
| 5864 | ) |
| 5865 | |
| 5866 | xnnpack_unit_test( |
| 5867 | name = "f32_vrndz_test", |
| 5868 | srcs = [ |
| 5869 | "test/f32-vrndz.cc", |
| 5870 | "test/vunary-microkernel-tester.h", |
| 5871 | ] + MICROKERNEL_TEST_HDRS, |
| 5872 | deps = MICROKERNEL_TEST_DEPS, |
| 5873 | ) |
| 5874 | |
| 5875 | xnnpack_unit_test( |
| 5876 | name = "f32_vrndu_test", |
| 5877 | srcs = [ |
| 5878 | "test/f32-vrndu.cc", |
| 5879 | "test/vunary-microkernel-tester.h", |
| 5880 | ] + MICROKERNEL_TEST_HDRS, |
| 5881 | deps = MICROKERNEL_TEST_DEPS, |
| 5882 | ) |
| 5883 | |
| 5884 | xnnpack_unit_test( |
| 5885 | name = "f32_vrndd_test", |
| 5886 | srcs = [ |
| 5887 | "test/f32-vrndd.cc", |
| 5888 | "test/vunary-microkernel-tester.h", |
| 5889 | ] + MICROKERNEL_TEST_HDRS, |
| 5890 | deps = MICROKERNEL_TEST_DEPS, |
| 5891 | ) |
| 5892 | |
| 5893 | xnnpack_unit_test( |
| Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 5894 | name = "f32_vscale_test", |
| 5895 | srcs = [ |
| 5896 | "test/f32-vscale.cc", |
| 5897 | "test/vscale-microkernel-tester.h", |
| 5898 | ] + MICROKERNEL_TEST_HDRS, |
| 5899 | deps = MICROKERNEL_TEST_DEPS, |
| 5900 | ) |
| 5901 | |
| 5902 | xnnpack_unit_test( |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5903 | name = "f32_vscaleexpminusmax_test", |
| 5904 | srcs = [ |
| 5905 | "test/f32-vscaleexpminusmax.cc", |
| 5906 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 5907 | ] + MICROKERNEL_TEST_HDRS, |
| 5908 | deps = MICROKERNEL_TEST_DEPS, |
| 5909 | ) |
| 5910 | |
| 5911 | xnnpack_unit_test( |
| Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 5912 | name = "f32_vscaleextexp_test", |
| 5913 | srcs = [ |
| 5914 | "test/f32-vscaleextexp.cc", |
| 5915 | "test/vscaleextexp-microkernel-tester.h", |
| 5916 | ] + MICROKERNEL_TEST_HDRS, |
| 5917 | deps = MICROKERNEL_TEST_DEPS, |
| 5918 | ) |
| 5919 | |
| 5920 | xnnpack_unit_test( |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5921 | name = "f32_vsqr_test", |
| 5922 | srcs = [ |
| 5923 | "test/f32-vsqr.cc", |
| 5924 | "test/vunary-microkernel-tester.h", |
| 5925 | ] + MICROKERNEL_TEST_HDRS, |
| 5926 | deps = MICROKERNEL_TEST_DEPS, |
| 5927 | ) |
| 5928 | |
| 5929 | xnnpack_unit_test( |
| Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 5930 | name = "f32_vsqrdiff_test", |
| 5931 | srcs = [ |
| 5932 | "test/f32-vsqrdiff.cc", |
| 5933 | "test/vbinary-microkernel-tester.h", |
| 5934 | ] + MICROKERNEL_TEST_HDRS, |
| 5935 | deps = MICROKERNEL_TEST_DEPS, |
| 5936 | ) |
| 5937 | |
| 5938 | xnnpack_unit_test( |
| 5939 | name = "f32_vsqrdiffc_test", |
| 5940 | srcs = [ |
| 5941 | "test/f32-vsqrdiffc.cc", |
| 5942 | "test/vbinaryc-microkernel-tester.h", |
| 5943 | ] + MICROKERNEL_TEST_HDRS, |
| 5944 | deps = MICROKERNEL_TEST_DEPS, |
| 5945 | ) |
| 5946 | |
| 5947 | xnnpack_unit_test( |
| Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 5948 | name = "f32_vsqrt_test", |
| 5949 | srcs = [ |
| 5950 | "test/f32-vsqrt.cc", |
| 5951 | "test/vunary-microkernel-tester.h", |
| 5952 | ] + MICROKERNEL_TEST_HDRS, |
| 5953 | deps = MICROKERNEL_TEST_DEPS, |
| 5954 | ) |
| 5955 | |
| 5956 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5957 | name = "f32_vsub_test", |
| 5958 | srcs = [ |
| 5959 | "test/f32-vsub.cc", |
| 5960 | "test/vbinary-microkernel-tester.h", |
| 5961 | ] + MICROKERNEL_TEST_HDRS, |
| 5962 | deps = MICROKERNEL_TEST_DEPS, |
| 5963 | ) |
| 5964 | |
| 5965 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5966 | name = "f32_vsub_minmax_test", |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5967 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5968 | "test/f32-vsub-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5969 | "test/vbinary-microkernel-tester.h", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5970 | ] + MICROKERNEL_TEST_HDRS, |
| 5971 | deps = MICROKERNEL_TEST_DEPS, |
| 5972 | ) |
| 5973 | |
| 5974 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5975 | name = "f32_vsub_relu_test", |
| 5976 | srcs = [ |
| 5977 | "test/f32-vsub-relu.cc", |
| 5978 | "test/vbinary-microkernel-tester.h", |
| 5979 | ] + MICROKERNEL_TEST_HDRS, |
| 5980 | deps = MICROKERNEL_TEST_DEPS, |
| 5981 | ) |
| 5982 | |
| 5983 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5984 | name = "f32_vsubc_test", |
| 5985 | srcs = [ |
| 5986 | "test/f32-vsubc.cc", |
| 5987 | "test/vbinaryc-microkernel-tester.h", |
| 5988 | ] + MICROKERNEL_TEST_HDRS, |
| 5989 | deps = MICROKERNEL_TEST_DEPS, |
| 5990 | ) |
| 5991 | |
| 5992 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5993 | name = "f32_vsubc_minmax_test", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5994 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5995 | "test/f32-vsubc-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5996 | "test/vbinaryc-microkernel-tester.h", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5997 | ] + MICROKERNEL_TEST_HDRS, |
| 5998 | deps = MICROKERNEL_TEST_DEPS, |
| 5999 | ) |
| 6000 | |
| 6001 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 6002 | name = "f32_vsubc_relu_test", |
| 6003 | srcs = [ |
| 6004 | "test/f32-vsubc-relu.cc", |
| 6005 | "test/vbinaryc-microkernel-tester.h", |
| 6006 | ] + MICROKERNEL_TEST_HDRS, |
| 6007 | deps = MICROKERNEL_TEST_DEPS, |
| 6008 | ) |
| 6009 | |
| 6010 | xnnpack_unit_test( |
| Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 6011 | name = "f32_vrsubc_test", |
| 6012 | srcs = [ |
| 6013 | "test/f32-vrsubc.cc", |
| 6014 | "test/vbinaryc-microkernel-tester.h", |
| 6015 | ] + MICROKERNEL_TEST_HDRS, |
| 6016 | deps = MICROKERNEL_TEST_DEPS, |
| 6017 | ) |
| 6018 | |
| 6019 | xnnpack_unit_test( |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6020 | name = "f32_vrsubc_minmax_test", |
| Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 6021 | srcs = [ |
| Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 6022 | "test/f32-vrsubc-minmax.cc", |
| Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 6023 | "test/vbinaryc-microkernel-tester.h", |
| Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 6024 | ] + MICROKERNEL_TEST_HDRS, |
| 6025 | deps = MICROKERNEL_TEST_DEPS, |
| 6026 | ) |
| 6027 | |
| 6028 | xnnpack_unit_test( |
| Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 6029 | name = "f32_vrsubc_relu_test", |
| 6030 | srcs = [ |
| 6031 | "test/f32-vrsubc-relu.cc", |
| 6032 | "test/vbinaryc-microkernel-tester.h", |
| 6033 | ] + MICROKERNEL_TEST_HDRS, |
| 6034 | deps = MICROKERNEL_TEST_DEPS, |
| 6035 | ) |
| 6036 | |
| 6037 | xnnpack_unit_test( |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6038 | name = "qs8_dwconv_minmax_test", |
| 6039 | srcs = [ |
| 6040 | "test/qs8-dwconv-minmax.cc", |
| 6041 | "test/dwconv-microkernel-tester.h", |
| 6042 | "src/xnnpack/AlignedAllocator.h", |
| 6043 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 6044 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 6045 | ) |
| 6046 | |
| 6047 | xnnpack_unit_test( |
| Marat Dukhan | 4ed53f4 | 2020-08-06 01:12:55 -0700 | [diff] [blame] | 6048 | name = "qs8_gavgpool_minmax_test", |
| 6049 | srcs = [ |
| 6050 | "test/qs8-gavgpool-minmax.cc", |
| 6051 | "test/gavgpool-microkernel-tester.h", |
| 6052 | "src/xnnpack/AlignedAllocator.h", |
| 6053 | ] + MICROKERNEL_TEST_HDRS, |
| 6054 | deps = MICROKERNEL_TEST_DEPS, |
| 6055 | ) |
| 6056 | |
| 6057 | xnnpack_unit_test( |
| Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 6058 | name = "qs8_gemm_minmax_test", |
| 6059 | srcs = [ |
| 6060 | "test/qs8-gemm-minmax.cc", |
| 6061 | "test/gemm-microkernel-tester.h", |
| 6062 | "src/xnnpack/AlignedAllocator.h", |
| 6063 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 6064 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 6065 | ) |
| 6066 | |
| 6067 | xnnpack_unit_test( |
| Marat Dukhan | f948068 | 2020-07-31 14:50:24 -0700 | [diff] [blame] | 6068 | name = "qs8_igemm_minmax_test", |
| 6069 | srcs = [ |
| 6070 | "test/qs8-igemm-minmax.cc", |
| 6071 | "test/gemm-microkernel-tester.h", |
| 6072 | "src/xnnpack/AlignedAllocator.h", |
| 6073 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 6074 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 6075 | ) |
| 6076 | |
| 6077 | xnnpack_unit_test( |
| 6078 | name = "qs8_requantization_test", |
| 6079 | srcs = [ |
| 6080 | "src/xnnpack/requantization-stubs.h", |
| 6081 | "test/qs8-requantization.cc", |
| 6082 | "test/requantization-tester.h", |
| 6083 | ] + MICROKERNEL_TEST_HDRS, |
| 6084 | deps = MICROKERNEL_TEST_DEPS, |
| 6085 | ) |
| 6086 | |
| 6087 | xnnpack_unit_test( |
| Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 6088 | name = "qs8_vadd_minmax_test", |
| 6089 | srcs = [ |
| 6090 | "test/qs8-vadd-minmax.cc", |
| 6091 | "test/vadd-microkernel-tester.h", |
| 6092 | ] + MICROKERNEL_TEST_HDRS, |
| 6093 | deps = MICROKERNEL_TEST_DEPS, |
| 6094 | ) |
| 6095 | |
| 6096 | xnnpack_unit_test( |
| Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 6097 | name = "qs8_vaddc_minmax_test", |
| 6098 | srcs = [ |
| 6099 | "test/qs8-vaddc-minmax.cc", |
| 6100 | "test/vaddc-microkernel-tester.h", |
| 6101 | ] + MICROKERNEL_TEST_HDRS, |
| 6102 | deps = MICROKERNEL_TEST_DEPS, |
| 6103 | ) |
| 6104 | |
| 6105 | xnnpack_unit_test( |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6106 | name = "qu8_avgpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6107 | srcs = [ |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6108 | "test/qu8-avgpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6109 | "test/avgpool-microkernel-tester.h", |
| 6110 | "src/xnnpack/AlignedAllocator.h", |
| 6111 | ] + MICROKERNEL_TEST_HDRS, |
| 6112 | deps = MICROKERNEL_TEST_DEPS, |
| 6113 | ) |
| 6114 | |
| 6115 | xnnpack_unit_test( |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6116 | name = "qu8_dwconv_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6117 | srcs = [ |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6118 | "test/qu8-dwconv-minmax.cc", |
| 6119 | "test/dwconv-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6120 | "src/xnnpack/AlignedAllocator.h", |
| 6121 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 6122 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6123 | ) |
| 6124 | |
| 6125 | xnnpack_unit_test( |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6126 | name = "qu8_igemm_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6127 | srcs = [ |
| Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 6128 | "test/qu8-igemm-minmax.cc", |
| 6129 | "test/gemm-microkernel-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6130 | "src/xnnpack/AlignedAllocator.h", |
| 6131 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 6132 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6133 | ) |
| 6134 | |
| 6135 | xnnpack_unit_test( |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6136 | name = "qu8_gavgpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6137 | srcs = [ |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6138 | "test/qu8-gavgpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6139 | "test/gavgpool-microkernel-tester.h", |
| 6140 | "src/xnnpack/AlignedAllocator.h", |
| 6141 | ] + MICROKERNEL_TEST_HDRS, |
| 6142 | deps = MICROKERNEL_TEST_DEPS, |
| 6143 | ) |
| 6144 | |
| 6145 | xnnpack_unit_test( |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6146 | name = "qu8_gemm_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6147 | srcs = [ |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6148 | "test/qu8-gemm-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6149 | "test/gemm-microkernel-tester.h", |
| 6150 | "src/xnnpack/AlignedAllocator.h", |
| 6151 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 6152 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6153 | ) |
| 6154 | |
| 6155 | xnnpack_unit_test( |
| Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 6156 | name = "qu8_requantization_test", |
| 6157 | srcs = [ |
| 6158 | "src/xnnpack/requantization-stubs.h", |
| 6159 | "test/qu8-requantization.cc", |
| 6160 | "test/requantization-tester.h", |
| 6161 | ] + MICROKERNEL_TEST_HDRS, |
| 6162 | deps = MICROKERNEL_TEST_DEPS, |
| 6163 | ) |
| 6164 | |
| 6165 | xnnpack_unit_test( |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6166 | name = "qu8_vadd_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6167 | srcs = [ |
| Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6168 | "test/qu8-vadd-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6169 | "test/vadd-microkernel-tester.h", |
| 6170 | ] + MICROKERNEL_TEST_HDRS, |
| 6171 | deps = MICROKERNEL_TEST_DEPS, |
| 6172 | ) |
| 6173 | |
| 6174 | xnnpack_unit_test( |
| 6175 | name = "u8_clamp_test", |
| 6176 | srcs = [ |
| 6177 | "test/u8-clamp.cc", |
| 6178 | "test/clamp-microkernel-tester.h", |
| 6179 | ] + MICROKERNEL_TEST_HDRS, |
| 6180 | deps = MICROKERNEL_TEST_DEPS, |
| 6181 | ) |
| 6182 | |
| 6183 | xnnpack_unit_test( |
| 6184 | name = "u8_lut32norm_test", |
| 6185 | srcs = [ |
| 6186 | "test/u8-lut32norm.cc", |
| 6187 | "test/lut-norm-microkernel-tester.h", |
| 6188 | ] + MICROKERNEL_TEST_HDRS, |
| 6189 | deps = MICROKERNEL_TEST_DEPS, |
| 6190 | ) |
| 6191 | |
| 6192 | xnnpack_unit_test( |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 6193 | name = "u8_maxpool_minmax_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6194 | srcs = [ |
| Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 6195 | "test/u8-maxpool-minmax.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6196 | "test/maxpool-microkernel-tester.h", |
| 6197 | ] + MICROKERNEL_TEST_HDRS, |
| 6198 | deps = MICROKERNEL_TEST_DEPS, |
| 6199 | ) |
| 6200 | |
| 6201 | xnnpack_unit_test( |
| 6202 | name = "u8_rmax_test", |
| 6203 | srcs = [ |
| 6204 | "test/u8-rmax.cc", |
| 6205 | "test/rmax-microkernel-tester.h", |
| 6206 | ] + MICROKERNEL_TEST_HDRS, |
| 6207 | deps = MICROKERNEL_TEST_DEPS, |
| 6208 | ) |
| 6209 | |
| 6210 | xnnpack_unit_test( |
| Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 6211 | name = "x32_fill_test", |
| 6212 | srcs = [ |
| 6213 | "test/x32-fill.cc", |
| 6214 | "test/fill-microkernel-tester.h", |
| 6215 | ] + MICROKERNEL_TEST_HDRS, |
| 6216 | deps = MICROKERNEL_TEST_DEPS, |
| 6217 | ) |
| 6218 | |
| 6219 | xnnpack_unit_test( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6220 | name = "x32_packx_test", |
| 6221 | srcs = [ |
| 6222 | "test/x32-packx.cc", |
| 6223 | "test/pack-microkernel-tester.h", |
| 6224 | "src/xnnpack/AlignedAllocator.h", |
| 6225 | ] + MICROKERNEL_TEST_HDRS, |
| 6226 | deps = MICROKERNEL_TEST_DEPS, |
| 6227 | ) |
| 6228 | |
| 6229 | xnnpack_unit_test( |
| 6230 | name = "x32_pad_test", |
| 6231 | srcs = [ |
| 6232 | "test/x32-pad.cc", |
| 6233 | "test/pad-microkernel-tester.h", |
| 6234 | ] + MICROKERNEL_TEST_HDRS, |
| 6235 | deps = MICROKERNEL_TEST_DEPS, |
| 6236 | ) |
| 6237 | |
| 6238 | xnnpack_unit_test( |
| 6239 | name = "x32_unpool_test", |
| 6240 | srcs = [ |
| 6241 | "test/x32-unpool.cc", |
| 6242 | "test/unpool-microkernel-tester.h", |
| 6243 | ] + MICROKERNEL_TEST_HDRS, |
| 6244 | deps = MICROKERNEL_TEST_DEPS, |
| 6245 | ) |
| 6246 | |
| 6247 | xnnpack_unit_test( |
| 6248 | name = "x32_zip_test", |
| 6249 | srcs = [ |
| 6250 | "test/x32-zip.cc", |
| 6251 | "test/zip-microkernel-tester.h", |
| 6252 | ] + MICROKERNEL_TEST_HDRS, |
| 6253 | deps = MICROKERNEL_TEST_DEPS, |
| 6254 | ) |
| 6255 | |
| 6256 | xnnpack_unit_test( |
| 6257 | name = "x8_lut_test", |
| 6258 | srcs = [ |
| 6259 | "test/x8-lut.cc", |
| 6260 | "test/lut-microkernel-tester.h", |
| 6261 | ] + MICROKERNEL_TEST_HDRS, |
| 6262 | deps = MICROKERNEL_TEST_DEPS, |
| 6263 | ) |
| 6264 | |
| 6265 | xnnpack_unit_test( |
| 6266 | name = "x8_zip_test", |
| 6267 | srcs = [ |
| 6268 | "test/x8-zip.cc", |
| 6269 | "test/zip-microkernel-tester.h", |
| 6270 | ] + MICROKERNEL_TEST_HDRS, |
| 6271 | deps = MICROKERNEL_TEST_DEPS, |
| 6272 | ) |
| 6273 | |
| Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6274 | ########################## Size tests for the library ######################### |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6275 | |
| 6276 | xnnpack_binary( |
| Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6277 | name = "operator_size_test", |
| 6278 | srcs = ["test/operator-size.c"], |
| Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 6279 | deps = [":xnnpack_operators_nhwc_f32"], |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6280 | ) |
| 6281 | |
| Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6282 | xnnpack_binary( |
| 6283 | name = "subgraph_size_test", |
| 6284 | srcs = ["test/subgraph-size.c"], |
| 6285 | deps = [":XNNPACK"], |
| 6286 | ) |
| 6287 | |
| 6288 | ########################### Unit tests for operators ########################## |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6289 | |
| 6290 | xnnpack_unit_test( |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6291 | name = "abs_nc_test", |
| 6292 | srcs = [ |
| 6293 | "test/abs-nc.cc", |
| 6294 | "test/abs-operator-tester.h", |
| 6295 | ], |
| 6296 | deps = OPERATOR_TEST_DEPS, |
| 6297 | ) |
| 6298 | |
| 6299 | xnnpack_unit_test( |
| Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6300 | name = "add_nd_test", |
| 6301 | srcs = [ |
| 6302 | "test/add-nd.cc", |
| 6303 | "test/binary-elementwise-operator-tester.h", |
| 6304 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6305 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6306 | ) |
| 6307 | |
| 6308 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6309 | name = "argmax_pooling_nhwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6310 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6311 | "test/argmax-pooling-nhwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6312 | "test/argmax-pooling-operator-tester.h", |
| 6313 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6314 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6315 | ) |
| 6316 | |
| 6317 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6318 | name = "average_pooling_nhwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6319 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6320 | "test/average-pooling-nhwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6321 | "test/average-pooling-operator-tester.h", |
| 6322 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6323 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6324 | ) |
| 6325 | |
| 6326 | xnnpack_unit_test( |
| Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6327 | name = "bankers_rounding_nc_test", |
| 6328 | srcs = [ |
| 6329 | "test/bankers-rounding-nc.cc", |
| 6330 | "test/bankers-rounding-operator-tester.h", |
| 6331 | ], |
| 6332 | deps = OPERATOR_TEST_DEPS, |
| 6333 | ) |
| 6334 | |
| 6335 | xnnpack_unit_test( |
| 6336 | name = "ceiling_nc_test", |
| 6337 | srcs = [ |
| 6338 | "test/ceiling-nc.cc", |
| 6339 | "test/ceiling-operator-tester.h", |
| 6340 | ], |
| 6341 | deps = OPERATOR_TEST_DEPS, |
| 6342 | ) |
| 6343 | |
| 6344 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6345 | name = "channel_shuffle_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6346 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6347 | "test/channel-shuffle-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6348 | "test/channel-shuffle-operator-tester.h", |
| 6349 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6350 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6351 | ) |
| 6352 | |
| 6353 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6354 | name = "clamp_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6355 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6356 | "test/clamp-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6357 | "test/clamp-operator-tester.h", |
| 6358 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6359 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6360 | ) |
| 6361 | |
| 6362 | xnnpack_unit_test( |
| Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 6363 | name = "constant_pad_nd_test", |
| 6364 | srcs = [ |
| 6365 | "test/constant-pad-nd.cc", |
| 6366 | "test/constant-pad-operator-tester.h", |
| 6367 | ], |
| 6368 | deps = OPERATOR_TEST_DEPS, |
| 6369 | ) |
| 6370 | |
| 6371 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6372 | name = "convolution_nhwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6373 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6374 | "test/convolution-nhwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6375 | "test/convolution-operator-tester.h", |
| 6376 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6377 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6378 | ) |
| 6379 | |
| 6380 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6381 | name = "convolution_nchw_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6382 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6383 | "test/convolution-nchw.cc", |
| 6384 | "test/convolution-operator-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6385 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6386 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6387 | ) |
| 6388 | |
| 6389 | xnnpack_unit_test( |
| Marat Dukhan | 4e21b27 | 2020-06-04 18:45:01 -0700 | [diff] [blame] | 6390 | name = "copy_nc_test", |
| 6391 | srcs = [ |
| 6392 | "test/copy-nc.cc", |
| 6393 | "test/copy-operator-tester.h", |
| 6394 | ], |
| 6395 | deps = OPERATOR_TEST_DEPS, |
| 6396 | ) |
| 6397 | |
| 6398 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6399 | name = "deconvolution_nhwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6400 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6401 | "test/deconvolution-nhwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6402 | "test/deconvolution-operator-tester.h", |
| 6403 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6404 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6405 | ) |
| 6406 | |
| 6407 | xnnpack_unit_test( |
| Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 6408 | name = "divide_nd_test", |
| 6409 | srcs = [ |
| 6410 | "test/binary-elementwise-operator-tester.h", |
| 6411 | "test/divide-nd.cc", |
| 6412 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6413 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 6414 | ) |
| 6415 | |
| 6416 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6417 | name = "fully_connected_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6418 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6419 | "test/fully-connected-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6420 | "test/fully-connected-operator-tester.h", |
| 6421 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6422 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6423 | ) |
| 6424 | |
| 6425 | xnnpack_unit_test( |
| Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6426 | name = "floor_nc_test", |
| 6427 | srcs = [ |
| 6428 | "test/floor-nc.cc", |
| 6429 | "test/floor-operator-tester.h", |
| 6430 | ], |
| 6431 | deps = OPERATOR_TEST_DEPS, |
| 6432 | ) |
| 6433 | |
| 6434 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6435 | name = "global_average_pooling_nwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6436 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6437 | "test/global-average-pooling-nwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6438 | "test/global-average-pooling-operator-tester.h", |
| Marat Dukhan | ef61d02 | 2020-06-19 13:54:49 -0700 | [diff] [blame] | 6439 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6440 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6441 | ) |
| 6442 | |
| 6443 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6444 | name = "global_average_pooling_ncw_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6445 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6446 | "test/global-average-pooling-ncw.cc", |
| 6447 | "test/global-average-pooling-operator-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6448 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6449 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6450 | ) |
| 6451 | |
| 6452 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6453 | name = "hardswish_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6454 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6455 | "test/hardswish-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6456 | "test/hardswish-operator-tester.h", |
| 6457 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6458 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6459 | ) |
| 6460 | |
| 6461 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6462 | name = "leaky_relu_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6463 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6464 | "test/leaky-relu-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6465 | "test/leaky-relu-operator-tester.h", |
| 6466 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6467 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6468 | ) |
| 6469 | |
| 6470 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6471 | name = "max_pooling_nhwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6472 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6473 | "test/max-pooling-nhwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6474 | "test/max-pooling-operator-tester.h", |
| 6475 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6476 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6477 | ) |
| 6478 | |
| 6479 | xnnpack_unit_test( |
| Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6480 | name = "maximum_nd_test", |
| 6481 | srcs = [ |
| 6482 | "test/binary-elementwise-operator-tester.h", |
| 6483 | "test/maximum-nd.cc", |
| 6484 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6485 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6486 | ) |
| 6487 | |
| 6488 | xnnpack_unit_test( |
| 6489 | name = "minimum_nd_test", |
| 6490 | srcs = [ |
| 6491 | "test/binary-elementwise-operator-tester.h", |
| 6492 | "test/minimum-nd.cc", |
| 6493 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6494 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6495 | ) |
| 6496 | |
| 6497 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6498 | name = "multiply_nd_test", |
| Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6499 | srcs = [ |
| Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6500 | "test/binary-elementwise-operator-tester.h", |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6501 | "test/multiply-nd.cc", |
| Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6502 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6503 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6504 | ) |
| 6505 | |
| 6506 | xnnpack_unit_test( |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6507 | name = "negate_nc_test", |
| 6508 | srcs = [ |
| 6509 | "test/negate-nc.cc", |
| 6510 | "test/negate-operator-tester.h", |
| 6511 | ], |
| 6512 | deps = OPERATOR_TEST_DEPS, |
| 6513 | ) |
| 6514 | |
| 6515 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6516 | name = "prelu_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6517 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6518 | "test/prelu-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6519 | "test/prelu-operator-tester.h", |
| 6520 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6521 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6522 | ) |
| 6523 | |
| 6524 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6525 | name = "resize_bilinear_nhwc_test", |
| Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6526 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6527 | "test/resize-bilinear-nhwc.cc", |
| Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6528 | "test/resize-bilinear-operator-tester.h", |
| 6529 | ] + OPERATOR_TEST_PARAMS_HDRS, |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6530 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6531 | ) |
| 6532 | |
| 6533 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6534 | name = "sigmoid_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6535 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6536 | "test/sigmoid-nc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6537 | "test/sigmoid-operator-tester.h", |
| 6538 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6539 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6540 | ) |
| 6541 | |
| 6542 | xnnpack_unit_test( |
| Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 6543 | name = "softmax_nc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6544 | srcs = [ |
| Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 6545 | "test/softmax-nc.cc", |
| 6546 | "test/softmax-operator-tester.h", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6547 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6548 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6549 | ) |
| 6550 | |
| 6551 | xnnpack_unit_test( |
| Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6552 | name = "square_nc_test", |
| 6553 | srcs = [ |
| 6554 | "test/square-nc.cc", |
| 6555 | "test/square-operator-tester.h", |
| 6556 | ], |
| 6557 | deps = OPERATOR_TEST_DEPS, |
| 6558 | ) |
| 6559 | |
| 6560 | xnnpack_unit_test( |
| Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 6561 | name = "square_root_nc_test", |
| 6562 | srcs = [ |
| 6563 | "test/square-root-nc.cc", |
| 6564 | "test/square-root-operator-tester.h", |
| 6565 | ], |
| 6566 | deps = OPERATOR_TEST_DEPS, |
| 6567 | ) |
| 6568 | |
| 6569 | xnnpack_unit_test( |
| Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 6570 | name = "squared_difference_nd_test", |
| 6571 | srcs = [ |
| 6572 | "test/binary-elementwise-operator-tester.h", |
| 6573 | "test/squared-difference-nd.cc", |
| 6574 | ], |
| 6575 | deps = OPERATOR_TEST_DEPS, |
| 6576 | ) |
| 6577 | |
| 6578 | xnnpack_unit_test( |
| Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 6579 | name = "subtract_nd_test", |
| 6580 | srcs = [ |
| 6581 | "test/binary-elementwise-operator-tester.h", |
| 6582 | "test/subtract-nd.cc", |
| 6583 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6584 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 6585 | ) |
| 6586 | |
| 6587 | xnnpack_unit_test( |
| Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6588 | name = "truncation_nc_test", |
| 6589 | srcs = [ |
| 6590 | "test/truncation-nc.cc", |
| 6591 | "test/truncation-operator-tester.h", |
| 6592 | ], |
| 6593 | deps = OPERATOR_TEST_DEPS, |
| 6594 | ) |
| 6595 | |
| 6596 | xnnpack_unit_test( |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6597 | name = "unpooling_nhwc_test", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6598 | srcs = [ |
| Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6599 | "test/unpooling-nhwc.cc", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6600 | "test/unpooling-operator-tester.h", |
| 6601 | ], |
| Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6602 | deps = OPERATOR_TEST_DEPS, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6603 | ) |
| 6604 | |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 6605 | ############################### Misc unit tests ############################### |
| 6606 | |
| 6607 | xnnpack_unit_test( |
| 6608 | name = "memory_planner_test", |
| 6609 | srcs = [ |
| 6610 | "test/memory-planner-test.cc", |
| 6611 | ], |
| 6612 | deps = [ |
| 6613 | ":XNNPACK", |
| 6614 | ":memory_planner", |
| 6615 | ], |
| 6616 | ) |
| 6617 | |
| XNNPACK Team | ab8c4c8 | 2020-10-09 08:05:51 -0700 | [diff] [blame] | 6618 | xnnpack_unit_test( |
| 6619 | name = "subgraph_nchw_test", |
| 6620 | srcs = [ |
| 6621 | "src/xnnpack/subgraph.h", |
| 6622 | "test/subgraph-nchw.cc", |
| 6623 | "test/subgraph-tester.h", |
| 6624 | ], |
| 6625 | deps = [ |
| 6626 | ":XNNPACK", |
| 6627 | ], |
| 6628 | ) |
| 6629 | |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6630 | ############################# Build configurations ############################# |
| 6631 | |
| Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6632 | # Enables usage of assembly kernels. |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6633 | config_setting( |
| Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6634 | name = "xnn_enable_assembly_explicit_true", |
| 6635 | define_values = {"xnn_enable_assembly": "true"}, |
| 6636 | ) |
| 6637 | |
| 6638 | # Disables usage of assembly kernels. |
| 6639 | config_setting( |
| 6640 | name = "xnn_enable_assembly_explicit_false", |
| 6641 | define_values = {"xnn_enable_assembly": "false"}, |
| 6642 | ) |
| 6643 | |
| Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 6644 | # Enables usage of sparse inference. |
| 6645 | config_setting( |
| 6646 | name = "xnn_enable_sparse_explicit_true", |
| 6647 | define_values = {"xnn_enable_sparse": "true"}, |
| 6648 | ) |
| 6649 | |
| 6650 | # Disables usage of sparse inference. |
| 6651 | config_setting( |
| 6652 | name = "xnn_enable_sparse_explicit_false", |
| 6653 | define_values = {"xnn_enable_sparse": "false"}, |
| 6654 | ) |
| 6655 | |
| Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 6656 | # Disables usage of HMP-aware optimizations. |
| 6657 | config_setting( |
| 6658 | name = "xnn_enable_hmp_explicit_false", |
| 6659 | define_values = {"xnn_enable_hmp": "false"}, |
| 6660 | ) |
| 6661 | |
| Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 6662 | # Enable usage of optimized memory allocation |
| 6663 | config_setting( |
| 6664 | name = "xnn_enable_memopt_explicit_true", |
| 6665 | define_values = {"xnnpack_enable_memopt": "true"}, |
| 6666 | ) |
| 6667 | |
| 6668 | # Disable usage of optimized memory allocation |
| 6669 | config_setting( |
| 6670 | name = "xnn_enable_memopt_explicit_false", |
| 6671 | define_values = {"xnnpack_enable_memopt": "false"}, |
| 6672 | ) |
| 6673 | |
| Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6674 | # Builds with -c dbg |
| 6675 | config_setting( |
| 6676 | name = "debug_build", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6677 | values = { |
| Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6678 | "compilation_mode": "dbg", |
| 6679 | }, |
| 6680 | ) |
| 6681 | |
| 6682 | # Builds with -c opt |
| 6683 | config_setting( |
| 6684 | name = "optimized_build", |
| 6685 | values = { |
| 6686 | "compilation_mode": "opt", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6687 | }, |
| 6688 | ) |
| 6689 | |
| 6690 | config_setting( |
| Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6691 | name = "linux_k8", |
| 6692 | values = {"cpu": "k8"}, |
| 6693 | ) |
| 6694 | |
| 6695 | config_setting( |
| Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 6696 | name = "linux_arm", |
| 6697 | values = {"cpu": "arm"}, |
| Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 6698 | ) |
| 6699 | |
| 6700 | config_setting( |
| Marat Dukhan | f0bd4de | 2020-06-15 15:53:19 -0700 | [diff] [blame] | 6701 | name = "linux_armeabi", |
| 6702 | values = {"cpu": "armeabi"}, |
| 6703 | ) |
| 6704 | |
| 6705 | config_setting( |
| Terry Heo | 68eef3f | 2020-04-13 22:53:52 -0700 | [diff] [blame] | 6706 | name = "linux_armhf", |
| 6707 | values = {"cpu": "armhf"}, |
| 6708 | ) |
| 6709 | |
| 6710 | config_setting( |
| Marat Dukhan | a720e93 | 2020-06-10 13:01:11 -0700 | [diff] [blame] | 6711 | name = "linux_armv7a", |
| 6712 | values = {"cpu": "armv7a"}, |
| 6713 | ) |
| 6714 | |
| 6715 | config_setting( |
| Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 6716 | name = "linux_aarch64", |
| 6717 | values = {"cpu": "aarch64"}, |
| 6718 | ) |
| 6719 | |
| 6720 | config_setting( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6721 | name = "android", |
| 6722 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 6723 | ) |
| 6724 | |
| 6725 | config_setting( |
| 6726 | name = "android_armv7", |
| 6727 | values = { |
| 6728 | "crosstool_top": "//external:android/crosstool", |
| 6729 | "cpu": "armeabi-v7a", |
| 6730 | }, |
| 6731 | ) |
| 6732 | |
| 6733 | config_setting( |
| 6734 | name = "android_arm64", |
| 6735 | values = { |
| 6736 | "crosstool_top": "//external:android/crosstool", |
| 6737 | "cpu": "arm64-v8a", |
| 6738 | }, |
| 6739 | ) |
| 6740 | |
| 6741 | config_setting( |
| 6742 | name = "android_x86", |
| 6743 | values = { |
| 6744 | "crosstool_top": "//external:android/crosstool", |
| 6745 | "cpu": "x86", |
| 6746 | }, |
| 6747 | ) |
| 6748 | |
| 6749 | config_setting( |
| 6750 | name = "android_x86_64", |
| 6751 | values = { |
| 6752 | "crosstool_top": "//external:android/crosstool", |
| 6753 | "cpu": "x86_64", |
| 6754 | }, |
| 6755 | ) |
| 6756 | |
| 6757 | config_setting( |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 6758 | name = "windows_x86_64", |
| 6759 | values = {"cpu": "x64_windows"}, |
| Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 6760 | ) |
| 6761 | |
| 6762 | config_setting( |
| Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 6763 | name = "windows_x86_64_clang", |
| 6764 | values = { |
| 6765 | "compiler": "clang-cl", |
| 6766 | "cpu": "x64_windows", |
| 6767 | }, |
| 6768 | ) |
| 6769 | |
| 6770 | config_setting( |
| 6771 | name = "windows_x86_64_mingw", |
| 6772 | values = { |
| 6773 | "compiler": "mingw-gcc", |
| 6774 | "cpu": "x64_windows", |
| 6775 | }, |
| 6776 | ) |
| 6777 | |
| 6778 | config_setting( |
| 6779 | name = "windows_x86_64_msys", |
| 6780 | values = { |
| 6781 | "compiler": "msys-gcc", |
| 6782 | "cpu": "x64_windows", |
| 6783 | }, |
| Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 6784 | ) |
| 6785 | |
| 6786 | config_setting( |
| Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 6787 | name = "macos_x86_64", |
| 6788 | values = { |
| 6789 | "apple_platform_type": "macos", |
| 6790 | "cpu": "darwin", |
| 6791 | }, |
| 6792 | ) |
| 6793 | |
| 6794 | config_setting( |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6795 | name = "emscripten", |
| Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6796 | values = {"crosstool_top": "//toolchain:emscripten"}, |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6797 | ) |
| 6798 | |
| 6799 | config_setting( |
| 6800 | name = "emscripten_wasm", |
| 6801 | values = { |
| Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6802 | "crosstool_top": "//toolchain:emscripten", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6803 | "cpu": "wasm", |
| 6804 | }, |
| 6805 | ) |
| 6806 | |
| 6807 | config_setting( |
| 6808 | name = "emscripten_wasmsimd", |
| 6809 | values = { |
| Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6810 | "crosstool_top": "//toolchain:emscripten", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6811 | "cpu": "wasm", |
| Marat Dukhan | 81c6260 | 2020-05-29 13:22:49 -0700 | [diff] [blame] | 6812 | "copt": "-msimd128", |
| Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6813 | }, |
| 6814 | ) |
| 6815 | |
| 6816 | config_setting( |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6817 | name = "ios_armv7", |
| 6818 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6819 | "apple_platform_type": "ios", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6820 | "cpu": "ios_armv7", |
| 6821 | }, |
| 6822 | ) |
| 6823 | |
| 6824 | config_setting( |
| 6825 | name = "ios_arm64", |
| 6826 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6827 | "apple_platform_type": "ios", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6828 | "cpu": "ios_arm64", |
| 6829 | }, |
| 6830 | ) |
| 6831 | |
| 6832 | config_setting( |
| 6833 | name = "ios_arm64e", |
| 6834 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6835 | "apple_platform_type": "ios", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6836 | "cpu": "ios_arm64e", |
| 6837 | }, |
| 6838 | ) |
| 6839 | |
| 6840 | config_setting( |
| 6841 | name = "ios_x86", |
| 6842 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6843 | "apple_platform_type": "ios", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6844 | "cpu": "ios_i386", |
| 6845 | }, |
| 6846 | ) |
| 6847 | |
| 6848 | config_setting( |
| 6849 | name = "ios_x86_64", |
| 6850 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6851 | "apple_platform_type": "ios", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6852 | "cpu": "ios_x86_64", |
| 6853 | }, |
| 6854 | ) |
| 6855 | |
| 6856 | config_setting( |
| 6857 | name = "watchos_armv7k", |
| 6858 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6859 | "apple_platform_type": "watchos", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6860 | "cpu": "watchos_armv7k", |
| 6861 | }, |
| 6862 | ) |
| 6863 | |
| 6864 | config_setting( |
| 6865 | name = "watchos_arm64_32", |
| 6866 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6867 | "apple_platform_type": "watchos", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6868 | "cpu": "watchos_arm64_32", |
| 6869 | }, |
| 6870 | ) |
| 6871 | |
| 6872 | config_setting( |
| 6873 | name = "watchos_x86", |
| 6874 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6875 | "apple_platform_type": "watchos", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6876 | "cpu": "watchos_i386", |
| 6877 | }, |
| 6878 | ) |
| 6879 | |
| 6880 | config_setting( |
| 6881 | name = "watchos_x86_64", |
| 6882 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6883 | "apple_platform_type": "watchos", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6884 | "cpu": "watchos_x86_64", |
| 6885 | }, |
| 6886 | ) |
| 6887 | |
| 6888 | config_setting( |
| 6889 | name = "tvos_arm64", |
| 6890 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6891 | "apple_platform_type": "tvos", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6892 | "cpu": "tvos_arm64", |
| 6893 | }, |
| 6894 | ) |
| 6895 | |
| 6896 | config_setting( |
| 6897 | name = "tvos_x86_64", |
| 6898 | values = { |
| Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6899 | "apple_platform_type": "tvos", |
| Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6900 | "cpu": "tvos_x86_64", |
| 6901 | }, |
| 6902 | ) |