Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1 | # Copyright 2019 Google LLC |
| 2 | # |
| 3 | # This source code is licensed under the BSD-style license found in the |
| 4 | # LICENSE file in the root directory of this source tree. |
| 5 | # |
| 6 | # Description: |
| 7 | # XNNPACK - optimized floating-point neural network operators library |
| 8 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 9 | load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility") |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 10 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 11 | licenses(["notice"]) |
| 12 | |
| 13 | exports_files(["LICENSE"]) |
| 14 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 15 | OPERATOR_BENCHMARK_DEPS = [ |
| 16 | ":XNNPACK", |
| 17 | ":bench_utils", |
| 18 | "@cpuinfo", |
Frank Barchard | 0c84973 | 2020-06-12 13:31:32 -0700 | [diff] [blame] | 19 | "@FP16", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 20 | "@pthreadpool", |
| 21 | ] |
| 22 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 23 | MICROKERNEL_BENCHMARK_DEPS = [ |
| 24 | ":ukernels", |
| 25 | ":bench_utils", |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 26 | ":enable_assembly", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 27 | "@cpuinfo", |
| 28 | "@FP16", |
| 29 | "@pthreadpool", |
| 30 | ] |
| 31 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 32 | ACCURACY_EVAL_DEPS = [ |
| 33 | ":XNNPACK", |
| 34 | ":ukernels", |
| 35 | "@FP16", |
| 36 | "@pthreadpool", |
| 37 | ] |
| 38 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 39 | MICROKERNEL_TEST_DEPS = [ |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 40 | ":ukernels_test_mode", |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 41 | ":enable_assembly", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 42 | "@cpuinfo", |
| 43 | "@FP16", |
| 44 | "@pthreadpool", |
| 45 | ] |
| 46 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 47 | OPERATOR_TEST_DEPS = [ |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 48 | ":XNNPACK_test_mode", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 49 | "@pthreadpool", |
| 50 | "@FP16", |
| 51 | ] |
| 52 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 53 | OPERATOR_SRCS = [ |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 54 | "src/operators/argmax-pooling-nhwc.c", |
| 55 | "src/operators/average-pooling-nhwc.c", |
| 56 | "src/operators/binary-elementwise-nd.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 57 | "src/operators/channel-shuffle-nc.c", |
Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 58 | "src/operators/constant-pad-nd.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 59 | "src/operators/convolution-nchw.c", |
| 60 | "src/operators/convolution-nhwc.c", |
| 61 | "src/operators/deconvolution-nhwc.c", |
| 62 | "src/operators/fully-connected-nc.c", |
| 63 | "src/operators/global-average-pooling-ncw.c", |
| 64 | "src/operators/global-average-pooling-nwc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 65 | "src/operators/leaky-relu-nc.c", |
| 66 | "src/operators/max-pooling-nhwc.c", |
| 67 | "src/operators/prelu-nc.c", |
| 68 | "src/operators/resize-bilinear-nhwc.c", |
| 69 | "src/operators/sigmoid-nc.c", |
| 70 | "src/operators/softmax-nc.c", |
Marat Dukhan | c3065f5 | 2020-06-04 13:33:32 -0700 | [diff] [blame] | 71 | "src/operators/unary-elementwise-nc.c", |
Marat Dukhan | e826543 | 2020-04-28 18:42:59 -0700 | [diff] [blame] | 72 | "src/operators/unpooling-nhwc.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 73 | ] |
| 74 | |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 75 | SUBGRAPH_SRCS = [ |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 76 | "src/subgraph/abs.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 77 | "src/subgraph/add2.c", |
| 78 | "src/subgraph/argmax-pooling-2d.c", |
| 79 | "src/subgraph/average-pooling-2d.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 80 | "src/subgraph/bankers-rounding.c", |
| 81 | "src/subgraph/ceiling.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 82 | "src/subgraph/clamp.c", |
| 83 | "src/subgraph/convolution-2d.c", |
| 84 | "src/subgraph/deconvolution-2d.c", |
| 85 | "src/subgraph/depthwise-convolution-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 86 | "src/subgraph/divide.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 87 | "src/subgraph/fully-connected.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 88 | "src/subgraph/floor.c", |
Marat Dukhan | a059b7d | 2020-06-11 11:41:27 -0700 | [diff] [blame] | 89 | "src/subgraph/global-average-pooling-2d.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 90 | "src/subgraph/hardswish.c", |
Marat Dukhan | 5bbebac | 2020-06-10 19:42:15 -0700 | [diff] [blame] | 91 | "src/subgraph/leaky-relu.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 92 | "src/subgraph/max-pooling-2d.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 93 | "src/subgraph/maximum2.c", |
| 94 | "src/subgraph/minimum2.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 95 | "src/subgraph/multiply2.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 96 | "src/subgraph/negate.c", |
Marat Dukhan | 95e8b7a | 2020-06-03 12:46:26 -0700 | [diff] [blame] | 97 | "src/subgraph/prelu.c", |
| 98 | "src/subgraph/sigmoid.c", |
| 99 | "src/subgraph/softmax.c", |
| 100 | "src/subgraph/static-constant-pad.c", |
Marat Dukhan | d27202d | 2020-07-09 23:43:40 -0700 | [diff] [blame] | 101 | "src/subgraph/static-reshape.c", |
Marat Dukhan | aff24e2 | 2020-07-23 01:43:58 -0700 | [diff] [blame] | 102 | "src/subgraph/static-resize-bilinear-2d.c", |
Marat Dukhan | 5fab409 | 2020-06-10 01:28:28 -0700 | [diff] [blame] | 103 | "src/subgraph/square.c", |
Marat Dukhan | 51a01c6 | 2020-07-09 03:26:57 -0700 | [diff] [blame] | 104 | "src/subgraph/square-root.c", |
Marat Dukhan | 9d3a459 | 2020-06-05 16:52:42 -0700 | [diff] [blame] | 105 | "src/subgraph/squared-difference.c", |
| 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", |
Marat Dukhan | 6b7dfae | 2019-12-04 16:00:52 -0800 | [diff] [blame] | 126 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 127 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 128 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c", |
| 129 | "src/f32-dwconv-chw/3x3p1-scalar.c", |
| 130 | "src/f32-dwconv-chw/3x3s2p1-scalar.c", |
| 131 | "src/f32-dwconv-chw/5x5p2-scalar.c", |
| 132 | "src/f32-dwconv-chw/5x5s2p2-scalar.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 133 | "src/f32-dwconv/gen/up1x4-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 134 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 135 | "src/f32-dwconv/gen/up1x9-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 136 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 137 | "src/f32-dwconv/gen/up1x25-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 138 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 139 | "src/f32-dwconv/gen/up2x4-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 140 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 141 | "src/f32-dwconv/gen/up2x9-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 142 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 143 | "src/f32-dwconv/gen/up2x25-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 144 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 145 | "src/f32-dwconv/gen/up1x4-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 146 | "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 147 | "src/f32-dwconv/gen/up1x9-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 148 | "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 149 | "src/f32-dwconv/gen/up1x25-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 150 | "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 151 | "src/f32-dwconv/gen/up2x4-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 152 | "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 153 | "src/f32-dwconv/gen/up2x9-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 154 | "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 155 | "src/f32-dwconv/gen/up2x25-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 156 | "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 157 | "src/f32-gavgpool-cw/scalar-x1.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 158 | "src/f32-gavgpool/7p7x-minmax-scalar-c1.c", |
| 159 | "src/f32-gavgpool/7x-minmax-scalar-c1.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 160 | "src/f32-gemm/gen/1x4-scalar.c", |
| 161 | "src/f32-gemm/gen/2x4-scalar.c", |
| 162 | "src/f32-gemm/gen/4x2-scalar.c", |
| 163 | "src/f32-gemm/gen/4x4-scalar.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 164 | "src/f32-gemm/gen/1x4-relu-scalar.c", |
| 165 | "src/f32-gemm/gen/2x4-relu-scalar.c", |
| 166 | "src/f32-gemm/gen/4x2-relu-scalar.c", |
| 167 | "src/f32-gemm/gen/4x4-relu-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 168 | "src/f32-gemm/gen/1x4-minmax-scalar.c", |
| 169 | "src/f32-gemm/gen/2x4-minmax-scalar.c", |
| 170 | "src/f32-gemm/gen/4x2-minmax-scalar.c", |
| 171 | "src/f32-gemm/gen/4x4-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 172 | "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c", |
| 173 | "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c", |
| 174 | "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 175 | "src/f32-hswish/gen/hswish-scalar-x1.c", |
| 176 | "src/f32-hswish/gen/hswish-scalar-x2.c", |
| 177 | "src/f32-hswish/gen/hswish-scalar-x4.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 178 | "src/f32-ibilinear/gen/scalar-c1.c", |
| 179 | "src/f32-ibilinear/gen/scalar-c2.c", |
| 180 | "src/f32-ibilinear/gen/scalar-c4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 181 | "src/f32-igemm/gen/1x4-scalar.c", |
| 182 | "src/f32-igemm/gen/2x4-scalar.c", |
| 183 | "src/f32-igemm/gen/4x2-scalar.c", |
| 184 | "src/f32-igemm/gen/4x4-scalar.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 185 | "src/f32-igemm/gen/1x4-relu-scalar.c", |
| 186 | "src/f32-igemm/gen/2x4-relu-scalar.c", |
| 187 | "src/f32-igemm/gen/4x2-relu-scalar.c", |
| 188 | "src/f32-igemm/gen/4x4-relu-scalar.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 189 | "src/f32-igemm/gen/1x4-minmax-scalar.c", |
| 190 | "src/f32-igemm/gen/2x4-minmax-scalar.c", |
| 191 | "src/f32-igemm/gen/4x2-minmax-scalar.c", |
| 192 | "src/f32-igemm/gen/4x4-minmax-scalar.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 193 | "src/f32-vlrelu/gen/vlrelu-scalar-x1.c", |
| 194 | "src/f32-vlrelu/gen/vlrelu-scalar-x2.c", |
| 195 | "src/f32-vlrelu/gen/vlrelu-scalar-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 196 | "src/f32-maxpool/9p8x-minmax-scalar-c1.c", |
| 197 | "src/f32-pavgpool/9p8x-minmax-scalar-c1.c", |
| 198 | "src/f32-pavgpool/9x-minmax-scalar-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 199 | "src/f32-ppmm/gen/2x4-minmax-scalar.c", |
| 200 | "src/f32-ppmm/gen/3x3-minmax-scalar.c", |
| 201 | "src/f32-ppmm/gen/4x2-minmax-scalar.c", |
| 202 | "src/f32-ppmm/gen/4x4-minmax-scalar.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 203 | "src/f32-prelu/gen/scalar-2x1.c", |
| 204 | "src/f32-prelu/gen/scalar-2x4.c", |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 205 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
| 206 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
| 207 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
| 208 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
| 209 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 210 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
| 211 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
| 212 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
| 213 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
| 214 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
| 215 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 216 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 217 | "src/f32-rmax/scalar.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 218 | "src/f32-relu/gen/scalar-x1.c", |
| 219 | "src/f32-relu/gen/scalar-x2.c", |
| 220 | "src/f32-relu/gen/scalar-x4.c", |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 221 | "src/f32-relu/gen/scalar-x8.c", |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 222 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 223 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 224 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
| 225 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 226 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 227 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
| 228 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 229 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 230 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 231 | "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c", |
| 232 | "src/f32-spmm/gen/1x1-minmax-scalar.c", |
| 233 | "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c", |
| 234 | "src/f32-spmm/gen/2x1-minmax-scalar.c", |
| 235 | "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c", |
| 236 | "src/f32-spmm/gen/4x1-minmax-scalar.c", |
| 237 | "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c", |
| 238 | "src/f32-spmm/gen/8x1-minmax-scalar.c", |
| 239 | "src/f32-spmm/gen/8x2-minmax-scalar.c", |
| 240 | "src/f32-spmm/gen/8x4-minmax-scalar.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 241 | "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c", |
| 242 | "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c", |
| 243 | "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 244 | "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 245 | "src/f32-vbinary/gen/vadd-relu-scalar-x1.c", |
| 246 | "src/f32-vbinary/gen/vadd-relu-scalar-x2.c", |
| 247 | "src/f32-vbinary/gen/vadd-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 248 | "src/f32-vbinary/gen/vadd-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 249 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 250 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 251 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 252 | "src/f32-vbinary/gen/vadd-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 253 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c", |
| 254 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c", |
| 255 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 256 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 257 | "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c", |
| 258 | "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c", |
| 259 | "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 260 | "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 261 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 262 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 263 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 264 | "src/f32-vbinary/gen/vaddc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 265 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c", |
| 266 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c", |
| 267 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 268 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 269 | "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c", |
| 270 | "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c", |
| 271 | "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 272 | "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 273 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 274 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 275 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 276 | "src/f32-vbinary/gen/vdiv-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 277 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c", |
| 278 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c", |
| 279 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 280 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 281 | "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c", |
| 282 | "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c", |
| 283 | "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 284 | "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 285 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 286 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 287 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 288 | "src/f32-vbinary/gen/vdivc-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 289 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 290 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 291 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 292 | "src/f32-vbinary/gen/vmax-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 293 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 294 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 295 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 296 | "src/f32-vbinary/gen/vmaxc-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 297 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 298 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 299 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 300 | "src/f32-vbinary/gen/vmin-scalar-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 301 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 302 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 303 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 304 | "src/f32-vbinary/gen/vminc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 305 | "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c", |
| 306 | "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c", |
| 307 | "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 308 | "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 309 | "src/f32-vbinary/gen/vmul-relu-scalar-x1.c", |
| 310 | "src/f32-vbinary/gen/vmul-relu-scalar-x2.c", |
| 311 | "src/f32-vbinary/gen/vmul-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 312 | "src/f32-vbinary/gen/vmul-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 313 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 314 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 315 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 316 | "src/f32-vbinary/gen/vmul-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 317 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c", |
| 318 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c", |
| 319 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 320 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 321 | "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c", |
| 322 | "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c", |
| 323 | "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 324 | "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 325 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 326 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 327 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 328 | "src/f32-vbinary/gen/vmulc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 329 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c", |
| 330 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c", |
| 331 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 332 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 333 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c", |
| 334 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c", |
| 335 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 336 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 337 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 338 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 339 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 340 | "src/f32-vbinary/gen/vrdivc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 341 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c", |
| 342 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c", |
| 343 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 344 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 345 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c", |
| 346 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c", |
| 347 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 348 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 349 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 350 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 351 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 352 | "src/f32-vbinary/gen/vrsubc-scalar-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 353 | "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c", |
| 354 | "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c", |
| 355 | "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 356 | "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 357 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c", |
| 358 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c", |
| 359 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 360 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 361 | "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c", |
| 362 | "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c", |
| 363 | "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 364 | "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 365 | "src/f32-vbinary/gen/vsub-relu-scalar-x1.c", |
| 366 | "src/f32-vbinary/gen/vsub-relu-scalar-x2.c", |
| 367 | "src/f32-vbinary/gen/vsub-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 368 | "src/f32-vbinary/gen/vsub-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 369 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 370 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 371 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 372 | "src/f32-vbinary/gen/vsub-scalar-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 373 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c", |
| 374 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c", |
| 375 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 376 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 377 | "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c", |
| 378 | "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c", |
| 379 | "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 380 | "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c", |
Frank Barchard | 8e229db | 2020-07-06 23:31:35 -0700 | [diff] [blame] | 381 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 382 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 383 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 384 | "src/f32-vbinary/gen/vsubc-scalar-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 385 | "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c", |
| 386 | "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c", |
| 387 | "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 388 | "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c", |
| 389 | "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c", |
| 390 | "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 391 | "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c", |
| 392 | "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c", |
| 393 | "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c", |
| 394 | "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c", |
| 395 | "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c", |
| 396 | "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c", |
| 397 | "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c", |
| 398 | "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c", |
| 399 | "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 400 | "src/f32-vsqrt/gen/scalar-sqrt-x1.c", |
| 401 | "src/f32-vsqrt/gen/scalar-sqrt-x2.c", |
| 402 | "src/f32-vsqrt/gen/scalar-sqrt-x4.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 403 | "src/f32-vunary/gen/vabs-scalar-x1.c", |
| 404 | "src/f32-vunary/gen/vabs-scalar-x2.c", |
| 405 | "src/f32-vunary/gen/vabs-scalar-x4.c", |
| 406 | "src/f32-vunary/gen/vneg-scalar-x1.c", |
| 407 | "src/f32-vunary/gen/vneg-scalar-x2.c", |
| 408 | "src/f32-vunary/gen/vneg-scalar-x4.c", |
| 409 | "src/f32-vunary/gen/vsqr-scalar-x1.c", |
| 410 | "src/f32-vunary/gen/vsqr-scalar-x2.c", |
| 411 | "src/f32-vunary/gen/vsqr-scalar-x4.c", |
Marat Dukhan | 5739f70 | 2019-12-22 19:45:09 -0800 | [diff] [blame] | 412 | "src/math/expminus-scalar-lut2048-p1.c", |
| 413 | "src/math/expminus-scalar-lut64-p2.c", |
| 414 | "src/math/expminus-scalar-p5.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 415 | "src/math/roundne-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 416 | "src/math/roundne-scalar-nearbyint.c", |
| 417 | "src/math/roundne-scalar-rint.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 418 | "src/math/roundd-scalar-addsub.c", |
| 419 | "src/math/roundd-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 420 | "src/math/roundd-scalar-floor.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 421 | "src/math/roundu-scalar-addsub.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 422 | "src/math/roundu-scalar-ceil.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 423 | "src/math/roundu-scalar-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 424 | "src/math/roundz-scalar-addsub.c", |
| 425 | "src/math/roundz-scalar-cvt.c", |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 426 | "src/math/roundz-scalar-trunc.c", |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 427 | "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c", |
| 428 | "src/math/sigmoid-scalar-rr2-lut64-p2-div.c", |
| 429 | "src/math/sigmoid-scalar-rr2-p5-div.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 430 | "src/qs8-gemm/gen/8x8c4-minmax-scalar.c", |
| 431 | "src/qs8-gemm/gen/12x4c4-minmax-scalar.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 432 | "src/qs8-requantization/fp32-scalar-lrintf.c", |
| 433 | "src/qs8-requantization/fp32-scalar-magic.c", |
| 434 | "src/qs8-requantization/precise-scalar-signed64.c", |
| 435 | "src/qs8-requantization/precise-scalar-unsigned32.c", |
| 436 | "src/qs8-requantization/precise-scalar-unsigned64.c", |
| 437 | "src/qs8-requantization/q31-scalar.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 438 | "src/qu8-avgpool/9p8x-minmax-scalar-c1.c", |
| 439 | "src/qu8-avgpool/9x-minmax-scalar-c1.c", |
| 440 | "src/qu8-dwconv/up1x9-minmax-scalar.c", |
| 441 | "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c", |
| 442 | "src/qu8-gavgpool/7x-minmax-scalar-c1.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 443 | "src/qu8-gemm/gen/8x8c4-minmax-scalar.c", |
| 444 | "src/qu8-gemm/gen/12x4c4-minmax-scalar.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 445 | "src/qu8-gemm/2x2-minmax-scalar.c", |
| 446 | "src/qu8-igemm/2x2-minmax-scalar.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 447 | "src/qu8-requantization/fp32-scalar-lrintf.c", |
| 448 | "src/qu8-requantization/fp32-scalar-magic.c", |
| 449 | "src/qu8-requantization/precise-scalar-signed64.c", |
| 450 | "src/qu8-requantization/precise-scalar-unsigned32.c", |
| 451 | "src/qu8-requantization/precise-scalar-unsigned64.c", |
| 452 | "src/qu8-requantization/q31-scalar.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 453 | "src/qu8-vadd/minmax-scalar.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 454 | "src/u8-clamp/scalar-x4.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 455 | "src/u8-lut32norm/scalar.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 456 | "src/u8-maxpool/9p8x-minmax-scalar-c1.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 457 | "src/u8-rmax/scalar.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 458 | "src/x32-fill/scalar-float.c", |
| 459 | "src/x32-fill/scalar-int.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 460 | "src/x32-packx/x2-scalar.c", |
| 461 | "src/x32-packx/x3-scalar.c", |
| 462 | "src/x32-packx/x4-scalar.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 463 | "src/x32-pad/scalar-float.c", |
Frank Barchard | c9c320e | 2020-08-07 22:12:46 -0700 | [diff] [blame] | 464 | "src/x32-pad/scalar-int.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 465 | "src/x32-unpool/scalar.c", |
| 466 | "src/x32-zip/x2-scalar.c", |
| 467 | "src/x32-zip/x3-scalar.c", |
| 468 | "src/x32-zip/x4-scalar.c", |
| 469 | "src/x32-zip/xm-scalar.c", |
| 470 | "src/x8-lut/scalar.c", |
| 471 | "src/x8-zip/x2-scalar.c", |
| 472 | "src/x8-zip/x3-scalar.c", |
| 473 | "src/x8-zip/x4-scalar.c", |
| 474 | "src/x8-zip/xm-scalar.c", |
| 475 | ] |
| 476 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 477 | WASM_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 478 | "src/f32-avgpool/9p8x-minmax-wasm-c1.c", |
| 479 | "src/f32-avgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 480 | "src/f32-clamp/gen/wasm-x1.c", |
| 481 | "src/f32-clamp/gen/wasm-x2.c", |
| 482 | "src/f32-clamp/gen/wasm-x4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 483 | "src/f32-dwconv/gen/up1x4-wasm-acc2.c", |
| 484 | "src/f32-dwconv/gen/up1x4-wasm.c", |
| 485 | "src/f32-dwconv/gen/up1x9-wasm-acc2.c", |
| 486 | "src/f32-dwconv/gen/up1x9-wasm.c", |
| 487 | "src/f32-dwconv/gen/up1x25-wasm-acc2.c", |
| 488 | "src/f32-dwconv/gen/up1x25-wasm.c", |
| 489 | "src/f32-dwconv/gen/up2x4-wasm-acc2.c", |
| 490 | "src/f32-dwconv/gen/up2x4-wasm.c", |
| 491 | "src/f32-dwconv/gen/up2x9-wasm-acc2.c", |
| 492 | "src/f32-dwconv/gen/up2x9-wasm.c", |
| 493 | "src/f32-dwconv/gen/up2x25-wasm-acc2.c", |
| 494 | "src/f32-dwconv/gen/up2x25-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 495 | "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c", |
| 496 | "src/f32-dwconv/gen/up1x4-minmax-wasm.c", |
| 497 | "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c", |
| 498 | "src/f32-dwconv/gen/up1x9-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 499 | "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c", |
| 500 | "src/f32-dwconv/gen/up1x25-minmax-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 501 | "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c", |
| 502 | "src/f32-dwconv/gen/up2x4-minmax-wasm.c", |
| 503 | "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c", |
| 504 | "src/f32-dwconv/gen/up2x9-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 505 | "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c", |
| 506 | "src/f32-dwconv/gen/up2x25-minmax-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 507 | "src/f32-gavgpool/7p7x-minmax-wasm-c1.c", |
| 508 | "src/f32-gavgpool/7x-minmax-wasm-c1.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 509 | "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c", |
| 510 | "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c", |
| 511 | "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 512 | "src/f32-gemm/gen/1x4-wasm.c", |
| 513 | "src/f32-gemm/gen/2x4-wasm.c", |
| 514 | "src/f32-gemm/gen/4x2-wasm.c", |
| 515 | "src/f32-gemm/gen/4x4-wasm.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 516 | "src/f32-gemm/gen/1x4-relu-wasm.c", |
| 517 | "src/f32-gemm/gen/2x4-relu-wasm.c", |
| 518 | "src/f32-gemm/gen/4x2-relu-wasm.c", |
| 519 | "src/f32-gemm/gen/4x4-relu-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 520 | "src/f32-gemm/gen/1x4-minmax-wasm.c", |
| 521 | "src/f32-gemm/gen/2x4-minmax-wasm.c", |
| 522 | "src/f32-gemm/gen/4x2-minmax-wasm.c", |
| 523 | "src/f32-gemm/gen/4x4-minmax-wasm.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 524 | "src/f32-hswish/gen/hswish-wasm-x1.c", |
| 525 | "src/f32-hswish/gen/hswish-wasm-x2.c", |
| 526 | "src/f32-hswish/gen/hswish-wasm-x4.c", |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 527 | "src/f32-igemm/gen/1x4-wasm.c", |
| 528 | "src/f32-igemm/gen/2x4-wasm.c", |
| 529 | "src/f32-igemm/gen/4x2-wasm.c", |
| 530 | "src/f32-igemm/gen/4x4-wasm.c", |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 531 | "src/f32-igemm/gen/1x4-relu-wasm.c", |
| 532 | "src/f32-igemm/gen/2x4-relu-wasm.c", |
| 533 | "src/f32-igemm/gen/4x2-relu-wasm.c", |
| 534 | "src/f32-igemm/gen/4x4-relu-wasm.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 535 | "src/f32-igemm/gen/1x4-minmax-wasm.c", |
| 536 | "src/f32-igemm/gen/2x4-minmax-wasm.c", |
| 537 | "src/f32-igemm/gen/4x2-minmax-wasm.c", |
| 538 | "src/f32-igemm/gen/4x4-minmax-wasm.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 539 | "src/f32-maxpool/9p8x-minmax-wasm-c1.c", |
| 540 | "src/f32-pavgpool/9p8x-minmax-wasm-c1.c", |
| 541 | "src/f32-pavgpool/9x-minmax-wasm-c1.c", |
Marat Dukhan | 7c1f808 | 2020-06-25 13:26:20 -0700 | [diff] [blame] | 542 | "src/f32-prelu/gen/wasm-2x1.c", |
| 543 | "src/f32-prelu/gen/wasm-2x4.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 544 | "src/f32-relu/gen/wasm-x1.c", |
| 545 | "src/f32-relu/gen/wasm-x2.c", |
| 546 | "src/f32-relu/gen/wasm-x4.c", |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 547 | "src/f32-relu/gen/wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 548 | "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c", |
| 549 | "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c", |
| 550 | "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 551 | "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 552 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c", |
| 553 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c", |
| 554 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 555 | "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 556 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c", |
| 557 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c", |
| 558 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 559 | "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 560 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c", |
| 561 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c", |
| 562 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 563 | "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 564 | "src/f32-vbinary/gen/vadd-relu-wasm-x1.c", |
| 565 | "src/f32-vbinary/gen/vadd-relu-wasm-x2.c", |
| 566 | "src/f32-vbinary/gen/vadd-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 567 | "src/f32-vbinary/gen/vadd-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 568 | "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c", |
| 569 | "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c", |
| 570 | "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 571 | "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 572 | "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c", |
| 573 | "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c", |
| 574 | "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 575 | "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 576 | "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c", |
| 577 | "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c", |
| 578 | "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 579 | "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 580 | "src/f32-vbinary/gen/vmax-wasm-x1.c", |
| 581 | "src/f32-vbinary/gen/vmax-wasm-x2.c", |
| 582 | "src/f32-vbinary/gen/vmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 583 | "src/f32-vbinary/gen/vmax-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 584 | "src/f32-vbinary/gen/vmaxc-wasm-x1.c", |
| 585 | "src/f32-vbinary/gen/vmaxc-wasm-x2.c", |
| 586 | "src/f32-vbinary/gen/vmaxc-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 587 | "src/f32-vbinary/gen/vmaxc-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 588 | "src/f32-vbinary/gen/vmin-wasm-x1.c", |
| 589 | "src/f32-vbinary/gen/vmin-wasm-x2.c", |
| 590 | "src/f32-vbinary/gen/vmin-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 591 | "src/f32-vbinary/gen/vmin-wasm-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 592 | "src/f32-vbinary/gen/vminc-wasm-x1.c", |
| 593 | "src/f32-vbinary/gen/vminc-wasm-x2.c", |
| 594 | "src/f32-vbinary/gen/vminc-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 595 | "src/f32-vbinary/gen/vminc-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 596 | "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c", |
| 597 | "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c", |
| 598 | "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 599 | "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 600 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c", |
| 601 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c", |
| 602 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 603 | "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 604 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c", |
| 605 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c", |
| 606 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 607 | "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 608 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c", |
| 609 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c", |
| 610 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 611 | "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 612 | "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c", |
| 613 | "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c", |
| 614 | "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 615 | "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 616 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c", |
| 617 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c", |
| 618 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 619 | "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 620 | "src/f32-vbinary/gen/vmul-relu-wasm-x1.c", |
| 621 | "src/f32-vbinary/gen/vmul-relu-wasm-x2.c", |
| 622 | "src/f32-vbinary/gen/vmul-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 623 | "src/f32-vbinary/gen/vmul-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 624 | "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c", |
| 625 | "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c", |
| 626 | "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 627 | "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 628 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c", |
| 629 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c", |
| 630 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 631 | "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 632 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c", |
| 633 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c", |
| 634 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 635 | "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 636 | "src/f32-vbinary/gen/vsub-relu-wasm-x1.c", |
| 637 | "src/f32-vbinary/gen/vsub-relu-wasm-x2.c", |
| 638 | "src/f32-vbinary/gen/vsub-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 639 | "src/f32-vbinary/gen/vsub-relu-wasm-x8.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 640 | "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c", |
| 641 | "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c", |
| 642 | "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 643 | "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 644 | "src/f32-vlrelu/gen/vlrelu-wasm-x1.c", |
| 645 | "src/f32-vlrelu/gen/vlrelu-wasm-x2.c", |
| 646 | "src/f32-vlrelu/gen/vlrelu-wasm-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 647 | "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c", |
| 648 | "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c", |
| 649 | "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 650 | ] |
| 651 | |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 652 | WASMSIMD_UKERNELS = [ |
Marat Dukhan | 40f0552 | 2020-07-16 22:33:12 -0700 | [diff] [blame] | 653 | "src/f32-argmaxpool/4x-wasmsimd-c4.c", |
| 654 | "src/f32-argmaxpool/9p8x-wasmsimd-c4.c", |
| 655 | "src/f32-argmaxpool/9x-wasmsimd-c4.c", |
Marat Dukhan | 3b7432d | 2020-07-16 17:46:32 -0700 | [diff] [blame] | 656 | "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 657 | "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| 658 | "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c", |
| 659 | "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 660 | "src/f32-clamp/gen/wasmsimd-arm-x4.c", |
| 661 | "src/f32-clamp/gen/wasmsimd-arm-x8.c", |
| 662 | "src/f32-clamp/gen/wasmsimd-x86-x4.c", |
| 663 | "src/f32-clamp/gen/wasmsimd-x86-x8.c", |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 664 | "src/f32-dwconv/gen/up4x25-wasmsimd.c", |
| 665 | "src/f32-dwconv/gen/up4x4-wasmsimd.c", |
| 666 | "src/f32-dwconv/gen/up4x9-wasmsimd.c", |
| 667 | "src/f32-dwconv/gen/up8x25-wasmsimd.c", |
| 668 | "src/f32-dwconv/gen/up8x4-wasmsimd.c", |
| 669 | "src/f32-dwconv/gen/up8x9-wasmsimd.c", |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 670 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c", |
| 671 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c", |
| 672 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c", |
| 673 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c", |
| 674 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c", |
| 675 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c", |
| 676 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c", |
| 677 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c", |
| 678 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c", |
| 679 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c", |
| 680 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c", |
| 681 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c", |
| 682 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c", |
| 683 | "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c", |
| 684 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c", |
| 685 | "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c", |
| 686 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c", |
| 687 | "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c", |
| 688 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c", |
| 689 | "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c", |
| 690 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c", |
| 691 | "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c", |
| 692 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c", |
| 693 | "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c", |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 694 | "src/f32-gemm/gen/1x8-wasmsimd-splat.c", |
| 695 | "src/f32-gemm/gen/4x8-wasmsimd-splat.c", |
| 696 | "src/f32-gemm/gen/5x8-wasmsimd-splat.c", |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 697 | "src/f32-gemm/gen/4x2c4-wasmsimd.c", |
Marat Dukhan | 688f6d8 | 2020-07-14 17:02:11 -0700 | [diff] [blame] | 698 | "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c", |
| 699 | "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c", |
| 700 | "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c", |
| 701 | "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c", |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 702 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 703 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 704 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 705 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 706 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 707 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 708 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 709 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 710 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 711 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 712 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 713 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 714 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 715 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 716 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 717 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 718 | "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 719 | "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 720 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 721 | "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 722 | "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 723 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 724 | "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 725 | "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 726 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 727 | "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 728 | "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 729 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 730 | "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 731 | "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 732 | "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c", |
| 733 | "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c", |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 734 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 735 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c", |
| 736 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c", |
| 737 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 738 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c", |
| 739 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c", |
| 740 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 741 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c", |
| 742 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c", |
| 743 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 744 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c", |
| 745 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c", |
| 746 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c", |
| 747 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c", |
| 748 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c", |
| 749 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 750 | "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c", |
| 751 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c", |
| 752 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 753 | "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c", |
| 754 | "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c", |
| 755 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 756 | "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c", |
| 757 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c", |
| 758 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 759 | "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c", |
| 760 | "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c", |
| 761 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c", |
| 762 | "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c", |
| 763 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c", |
Marat Dukhan | 00d1d6e | 2020-07-09 01:37:27 -0700 | [diff] [blame] | 764 | "src/f32-ibilinear/gen/wasmsimd-c4.c", |
| 765 | "src/f32-ibilinear/gen/wasmsimd-c8.c", |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 766 | "src/f32-igemm/gen/1x8-wasmsimd-splat.c", |
| 767 | "src/f32-igemm/gen/4x8-wasmsimd-splat.c", |
| 768 | "src/f32-igemm/gen/5x8-wasmsimd-splat.c", |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 769 | "src/f32-igemm/gen/4x2c4-wasmsimd.c", |
Marat Dukhan | 688f6d8 | 2020-07-14 17:02:11 -0700 | [diff] [blame] | 770 | "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c", |
| 771 | "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c", |
| 772 | "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c", |
| 773 | "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c", |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 774 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c", |
| 775 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c", |
| 776 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c", |
| 777 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c", |
| 778 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c", |
| 779 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c", |
| 780 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c", |
| 781 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 782 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c", |
| 783 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c", |
| 784 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c", |
| 785 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c", |
| 786 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c", |
| 787 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c", |
| 788 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c", |
| 789 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c", |
| 790 | "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c", |
| 791 | "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c", |
| 792 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c", |
| 793 | "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c", |
| 794 | "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c", |
| 795 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c", |
| 796 | "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
| 797 | "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c", |
| 798 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c", |
| 799 | "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c", |
| 800 | "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c", |
| 801 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c", |
| 802 | "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c", |
| 803 | "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c", |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 804 | "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c", |
| 805 | "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c", |
Marat Dukhan | c601680 | 2020-07-16 18:51:28 -0700 | [diff] [blame] | 806 | "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c", |
| 807 | "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c", |
| 808 | "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c", |
| 809 | "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 1849351 | 2020-07-24 14:13:48 -0700 | [diff] [blame] | 810 | "src/f32-gavgpool-cw/wasmsimd-arm-x4.c", |
| 811 | "src/f32-gavgpool-cw/wasmsimd-x86-x4.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 812 | "src/f32-hswish/gen/hswish-wasmsimd-x4.c", |
| 813 | "src/f32-hswish/gen/hswish-wasmsimd-x8.c", |
| 814 | "src/f32-hswish/gen/hswish-wasmsimd-x16.c", |
Marat Dukhan | f6e2480 | 2020-07-08 22:20:40 -0700 | [diff] [blame] | 815 | "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 816 | "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 1483c53 | 2020-07-16 18:08:19 -0700 | [diff] [blame] | 817 | "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c", |
| 818 | "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c", |
| 819 | "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c", |
| 820 | "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c", |
Marat Dukhan | 0d0d882 | 2020-07-23 23:37:56 -0700 | [diff] [blame] | 821 | "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c", |
| 822 | "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 823 | "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c", |
| 824 | "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c", |
| 825 | "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c", |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 826 | "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c", |
| 827 | "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 828 | "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c", |
| 829 | "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c", |
| 830 | "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c", |
| 831 | "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c", |
| 832 | "src/f32-prelu/gen/wasmsimd-minmax-1x4.c", |
| 833 | "src/f32-prelu/gen/wasmsimd-minmax-1x8.c", |
| 834 | "src/f32-prelu/gen/wasmsimd-minmax-1x16.c", |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 835 | "src/f32-prelu/gen/wasmsimd-minmax-2x4.c", |
| 836 | "src/f32-prelu/gen/wasmsimd-minmax-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 837 | "src/f32-prelu/gen/wasmsimd-minmax-2x16.c", |
| 838 | "src/f32-prelu/gen/wasmsimd-minmax-4x4.c", |
| 839 | "src/f32-prelu/gen/wasmsimd-minmax-4x8.c", |
| 840 | "src/f32-prelu/gen/wasmsimd-minmax-4x16.c", |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 841 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c", |
| 842 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c", |
| 843 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c", |
| 844 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c", |
| 845 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c", |
| 846 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c", |
| 847 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c", |
| 848 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c", |
| 849 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c", |
| 850 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c", |
| 851 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c", |
| 852 | "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 853 | "src/f32-relu/gen/wasmsimd-x4.c", |
| 854 | "src/f32-relu/gen/wasmsimd-x8.c", |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 855 | "src/f32-relu/gen/wasmsimd-x16.c", |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 856 | "src/f32-rmax/wasmsimd-arm.c", |
| 857 | "src/f32-rmax/wasmsimd-x86.c", |
Marat Dukhan | b3635ed | 2020-07-16 12:36:28 -0700 | [diff] [blame] | 858 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c", |
| 859 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c", |
| 860 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c", |
| 861 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c", |
| 862 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c", |
| 863 | "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c", |
Marat Dukhan | d187a5b | 2020-07-20 01:07:17 -0700 | [diff] [blame] | 864 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c", |
| 865 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c", |
| 866 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c", |
| 867 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c", |
| 868 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c", |
| 869 | "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 870 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c", |
| 871 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 872 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 873 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c", |
| 874 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 875 | "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 876 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c", |
| 877 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 878 | "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 879 | "src/f32-vbinary/gen/vadd-wasmsimd-x4.c", |
| 880 | "src/f32-vbinary/gen/vadd-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 881 | "src/f32-vbinary/gen/vadd-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 882 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c", |
| 883 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 884 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 885 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c", |
| 886 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 887 | "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 888 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c", |
| 889 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 890 | "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 891 | "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c", |
| 892 | "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 893 | "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 894 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c", |
| 895 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 896 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 897 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c", |
| 898 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 899 | "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 900 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c", |
| 901 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 902 | "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 903 | "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c", |
| 904 | "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 905 | "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 906 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c", |
| 907 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 908 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 909 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c", |
| 910 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 911 | "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 912 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c", |
| 913 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 914 | "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 915 | "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c", |
| 916 | "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 917 | "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 918 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c", |
| 919 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 920 | "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 921 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c", |
| 922 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 923 | "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 924 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c", |
| 925 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 926 | "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 927 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c", |
| 928 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 929 | "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 930 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c", |
| 931 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 932 | "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 933 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c", |
| 934 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 935 | "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 936 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c", |
| 937 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 938 | "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 939 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c", |
| 940 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 941 | "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 942 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c", |
| 943 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 944 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 945 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c", |
| 946 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 947 | "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 948 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c", |
| 949 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 950 | "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 951 | "src/f32-vbinary/gen/vmul-wasmsimd-x4.c", |
| 952 | "src/f32-vbinary/gen/vmul-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 953 | "src/f32-vbinary/gen/vmul-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 954 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c", |
| 955 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 956 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 957 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c", |
| 958 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 959 | "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 960 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c", |
| 961 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 962 | "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 963 | "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c", |
| 964 | "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 965 | "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 966 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c", |
| 967 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 968 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 969 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c", |
| 970 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 971 | "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 972 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c", |
| 973 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 974 | "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 975 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c", |
| 976 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 977 | "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 978 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c", |
| 979 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 980 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 981 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c", |
| 982 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 983 | "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 984 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c", |
| 985 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 986 | "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 987 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c", |
| 988 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 989 | "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 990 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c", |
| 991 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 992 | "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 993 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c", |
| 994 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 995 | "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 996 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c", |
| 997 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 998 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 999 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c", |
| 1000 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1001 | "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1002 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c", |
| 1003 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1004 | "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1005 | "src/f32-vbinary/gen/vsub-wasmsimd-x4.c", |
| 1006 | "src/f32-vbinary/gen/vsub-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1007 | "src/f32-vbinary/gen/vsub-wasmsimd-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1008 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c", |
| 1009 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1010 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c", |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 1011 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c", |
| 1012 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1013 | "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c", |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 1014 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c", |
| 1015 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1016 | "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c", |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 1017 | "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c", |
| 1018 | "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c", |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 1019 | "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1020 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c", |
| 1021 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c", |
| 1022 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c", |
| 1023 | "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c", |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 1024 | "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c", |
| 1025 | "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c", |
| 1026 | "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c", |
| 1027 | "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c", |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 1028 | "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c", |
| 1029 | "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c", |
| 1030 | "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c", |
| 1031 | "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c", |
| 1032 | "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c", |
| 1033 | "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c", |
| 1034 | "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c", |
| 1035 | "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c", |
| 1036 | "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c", |
| 1037 | "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c", |
| 1038 | "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c", |
| 1039 | "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c", |
| 1040 | "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c", |
| 1041 | "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1042 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c", |
| 1043 | "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c", |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 1044 | "src/f32-vunary/gen/vabs-wasmsimd-x4.c", |
| 1045 | "src/f32-vunary/gen/vabs-wasmsimd-x8.c", |
| 1046 | "src/f32-vunary/gen/vneg-wasmsimd-x4.c", |
| 1047 | "src/f32-vunary/gen/vneg-wasmsimd-x8.c", |
| 1048 | "src/f32-vunary/gen/vsqr-wasmsimd-x4.c", |
| 1049 | "src/f32-vunary/gen/vsqr-wasmsimd-x8.c", |
Marat Dukhan | cc8f34c | 2020-08-05 16:36:38 -0700 | [diff] [blame] | 1050 | "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c", |
| 1051 | "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c", |
| 1052 | "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c", |
Marat Dukhan | ef45180 | 2020-08-06 11:53:47 -0700 | [diff] [blame] | 1053 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c", |
| 1054 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c", |
| 1055 | "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c", |
Marat Dukhan | b5e3d17 | 2020-08-06 13:29:53 -0700 | [diff] [blame] | 1056 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c", |
| 1057 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c", |
| 1058 | "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c", |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1059 | "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c", |
| 1060 | "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c", |
| 1061 | "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c", |
| 1062 | "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c", |
| 1063 | "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c", |
| 1064 | "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c", |
| 1065 | "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c", |
| 1066 | "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c", |
| 1067 | "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c", |
| 1068 | "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c", |
| 1069 | "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c", |
| 1070 | "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c", |
| 1071 | "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c", |
| 1072 | "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c", |
| 1073 | "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1074 | "src/qs8-requantization/fp32-wasmsimd.c", |
Marat Dukhan | 138560c | 2020-08-03 18:57:34 -0700 | [diff] [blame] | 1075 | "src/qs8-requantization/q31-wasmsimd.c", |
Marat Dukhan | 5df27f8 | 2020-09-02 23:59:21 -0700 | [diff] [blame] | 1076 | "src/qs8-vadd/gen/minmax-wasmsimd-x8.c", |
| 1077 | "src/qs8-vadd/gen/minmax-wasmsimd-x16.c", |
| 1078 | "src/qs8-vadd/gen/minmax-wasmsimd-x24.c", |
| 1079 | "src/qs8-vadd/gen/minmax-wasmsimd-x32.c", |
| 1080 | "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c", |
| 1081 | "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c", |
| 1082 | "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c", |
| 1083 | "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1084 | "src/qu8-requantization/fp32-wasmsimd.c", |
Marat Dukhan | 22de5e7 | 2020-08-03 21:53:18 -0700 | [diff] [blame] | 1085 | "src/qu8-requantization/q31-wasmsimd.c", |
Marat Dukhan | 8ee3701 | 2020-07-16 13:17:13 -0700 | [diff] [blame] | 1086 | "src/x32-fill/wasmsimd.c", |
Marat Dukhan | 66d99e9 | 2020-07-16 12:56:21 -0700 | [diff] [blame] | 1087 | "src/x32-packx/x4-wasmsimd.c", |
Marat Dukhan | 9306ae0 | 2020-07-16 15:51:13 -0700 | [diff] [blame] | 1088 | "src/x32-pad/wasmsimd.c", |
Marat Dukhan | 9d4bfa2 | 2020-07-16 19:07:04 -0700 | [diff] [blame] | 1089 | "src/x32-unpool/wasmsimd.c", |
Marat Dukhan | e3b7876 | 2020-07-16 20:02:58 -0700 | [diff] [blame] | 1090 | "src/x32-zip/x2-wasmsimd.c", |
| 1091 | "src/x32-zip/x3-wasmsimd.c", |
| 1092 | "src/x32-zip/x4-wasmsimd.c", |
| 1093 | "src/x32-zip/xm-wasmsimd.c", |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 1094 | "src/math/roundne-wasmsimd-addsub.c", |
| 1095 | "src/math/roundd-wasmsimd-addsub.c", |
Marat Dukhan | 88da62c | 2020-07-15 17:56:09 -0700 | [diff] [blame] | 1096 | "src/math/roundd-wasmsimd-cvt.c", |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 1097 | "src/math/roundu-wasmsimd-addsub.c", |
Marat Dukhan | 88da62c | 2020-07-15 17:56:09 -0700 | [diff] [blame] | 1098 | "src/math/roundu-wasmsimd-cvt.c", |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 1099 | "src/math/roundz-wasmsimd-addsub.c", |
Marat Dukhan | 88da62c | 2020-07-15 17:56:09 -0700 | [diff] [blame] | 1100 | "src/math/roundz-wasmsimd-cvt.c", |
Marat Dukhan | 773da43 | 2020-07-17 18:55:36 -0700 | [diff] [blame] | 1101 | "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c", |
| 1102 | "src/math/sigmoid-wasmsimd-rr2-p5-div.c", |
Marat Dukhan | 290055c | 2020-06-09 12:24:29 -0700 | [diff] [blame] | 1103 | ] |
| 1104 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1105 | PSIMD_FASTMATH_UKERNELS = [ |
Erich Elsen | 0a1970e | 2020-06-10 09:24:59 -0700 | [diff] [blame] | 1106 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c", |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 1107 | "src/f32-dwconv-chw/3x3p1-psimd.c", |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 1108 | "src/f32-dwconv-chw/3x3s2p1-psimd.c", |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 1109 | "src/f32-dwconv-chw/5x5p2-psimd.c", |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 1110 | "src/f32-dwconv-chw/5x5s2p2-psimd.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 1111 | "src/f32-spmm/gen/4x1-minmax-psimd.c", |
| 1112 | "src/f32-spmm/gen/8x1-minmax-psimd.c", |
| 1113 | "src/f32-spmm/gen/16x1-minmax-psimd.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1114 | ] |
| 1115 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1116 | PSIMD_ACCMATH_UKERNELS = [ |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1117 | "src/qs8-requantization/precise-psimd.c", |
| 1118 | "src/qs8-requantization/fp32-psimd.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1119 | "src/qu8-requantization/precise-psimd.c", |
| 1120 | "src/qu8-requantization/fp32-psimd.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1121 | ] |
| 1122 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1123 | # ISA-specific micro-kernels |
| 1124 | NEON_UKERNELS = [ |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 1125 | "src/f32-argmaxpool/4x-neon-c4.c", |
| 1126 | "src/f32-argmaxpool/9p8x-neon-c4.c", |
| 1127 | "src/f32-argmaxpool/9x-neon-c4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1128 | "src/f32-avgpool/9p8x-minmax-neon-c4.c", |
| 1129 | "src/f32-avgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1130 | "src/f32-clamp/gen/neon-x4.c", |
| 1131 | "src/f32-clamp/gen/neon-x8.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1132 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c", |
| 1133 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1134 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c", |
| 1135 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1136 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c", |
| 1137 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1138 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c", |
| 1139 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1140 | "src/f32-dwconv/gen/up4x4-minmax-neon.c", |
| 1141 | "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", |
| 1142 | "src/f32-dwconv/gen/up8x4-minmax-neon.c", |
| 1143 | "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1144 | "src/f32-dwconv/gen/up4x9-minmax-neon.c", |
| 1145 | "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", |
| 1146 | "src/f32-dwconv/gen/up8x9-minmax-neon.c", |
| 1147 | "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1148 | "src/f32-dwconv/gen/up4x25-minmax-neon.c", |
| 1149 | "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", |
| 1150 | "src/f32-dwconv/gen/up8x25-minmax-neon.c", |
| 1151 | "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1152 | "src/f32-gavgpool-cw/neon-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1153 | "src/f32-gavgpool/7p7x-minmax-neon-c4.c", |
| 1154 | "src/f32-gavgpool/7x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1155 | "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 1156 | "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 1157 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 1158 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 1159 | "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c", |
| 1160 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 1161 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 1162 | "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 1163 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 1164 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 1165 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 1166 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 1167 | "src/f32-gemm/gen/1x8s4-minmax-neon.c", |
| 1168 | "src/f32-gemm/gen/4x8s4-minmax-neon.c", |
| 1169 | "src/f32-gemm/gen/6x8s4-minmax-neon.c", |
| 1170 | "src/f32-gemm/gen/8x8s4-minmax-neon.c", |
| 1171 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c", |
| 1172 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c", |
| 1173 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c", |
| 1174 | "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c", |
| 1175 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c", |
| 1176 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c", |
| 1177 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c", |
| 1178 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c", |
| 1179 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c", |
| 1180 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c", |
| 1181 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c", |
| 1182 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c", |
| 1183 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c", |
| 1184 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c", |
| 1185 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1186 | "src/f32-hswish/gen/hswish-neon-x4.c", |
| 1187 | "src/f32-hswish/gen/hswish-neon-x8.c", |
| 1188 | "src/f32-hswish/gen/hswish-neon-x16.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1189 | "src/f32-ibilinear/gen/neon-c4.c", |
| 1190 | "src/f32-ibilinear/gen/neon-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1191 | "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 1192 | "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 1193 | "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c", |
| 1194 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 1195 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 1196 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 1197 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 1198 | "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 1199 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 1200 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 1201 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 1202 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 1203 | "src/f32-igemm/gen/1x8s4-minmax-neon.c", |
| 1204 | "src/f32-igemm/gen/4x8s4-minmax-neon.c", |
| 1205 | "src/f32-igemm/gen/6x8s4-minmax-neon.c", |
| 1206 | "src/f32-igemm/gen/8x8s4-minmax-neon.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1207 | "src/f32-maxpool/9p8x-minmax-neon-c4.c", |
| 1208 | "src/f32-pavgpool/9p8x-minmax-neon-c4.c", |
| 1209 | "src/f32-pavgpool/9x-minmax-neon-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1210 | "src/f32-ppmm/gen/4x8-minmax-neon.c", |
| 1211 | "src/f32-ppmm/gen/8x8-minmax-neon.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1212 | "src/f32-prelu/gen/neon-1x4.c", |
| 1213 | "src/f32-prelu/gen/neon-1x8.c", |
| 1214 | "src/f32-prelu/gen/neon-1x16.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1215 | "src/f32-prelu/gen/neon-2x4.c", |
| 1216 | "src/f32-prelu/gen/neon-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1217 | "src/f32-prelu/gen/neon-2x16.c", |
| 1218 | "src/f32-prelu/gen/neon-4x4.c", |
| 1219 | "src/f32-prelu/gen/neon-4x8.c", |
| 1220 | "src/f32-prelu/gen/neon-4x16.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1221 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 1222 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 1223 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 1224 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 1225 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 1226 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 1227 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 1228 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 1229 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 1230 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
| 1231 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 1232 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 1233 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
| 1234 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
| 1235 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
| 1236 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
| 1237 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 1238 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
| 1239 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
| 1240 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 1241 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
| 1242 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 1243 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 1244 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1245 | "src/f32-relu/gen/neon-x4.c", |
| 1246 | "src/f32-relu/gen/neon-x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1247 | "src/f32-rmax/neon.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1248 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1249 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 1250 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 1251 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 1252 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 1253 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 1254 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
| 1255 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 1256 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 1257 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 1258 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 1259 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 1260 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 1261 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 1262 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 1263 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 1264 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 1265 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 1266 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1267 | "src/f32-vbinary/gen/vadd-minmax-neon-x4.c", |
| 1268 | "src/f32-vbinary/gen/vadd-minmax-neon-x8.c", |
| 1269 | "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c", |
| 1270 | "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1271 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 1272 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 1273 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 1274 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 1275 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 1276 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 1277 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 1278 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1279 | "src/f32-vbinary/gen/vmul-minmax-neon-x4.c", |
| 1280 | "src/f32-vbinary/gen/vmul-minmax-neon-x8.c", |
| 1281 | "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c", |
| 1282 | "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c", |
| 1283 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c", |
| 1284 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1285 | "src/f32-vbinary/gen/vsqrdiff-neon-x4.c", |
| 1286 | "src/f32-vbinary/gen/vsqrdiff-neon-x8.c", |
| 1287 | "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c", |
| 1288 | "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1289 | "src/f32-vbinary/gen/vsub-minmax-neon-x4.c", |
| 1290 | "src/f32-vbinary/gen/vsub-minmax-neon-x8.c", |
| 1291 | "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c", |
| 1292 | "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1293 | "src/f32-vlrelu/gen/vlrelu-neon-x4.c", |
| 1294 | "src/f32-vlrelu/gen/vlrelu-neon-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1295 | "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c", |
| 1296 | "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1297 | "src/f32-vrnd/gen/vrndne-neon-x4.c", |
| 1298 | "src/f32-vrnd/gen/vrndne-neon-x8.c", |
| 1299 | "src/f32-vrnd/gen/vrndz-neon-x4.c", |
| 1300 | "src/f32-vrnd/gen/vrndz-neon-x8.c", |
| 1301 | "src/f32-vrnd/gen/vrndu-neon-x4.c", |
| 1302 | "src/f32-vrnd/gen/vrndu-neon-x8.c", |
| 1303 | "src/f32-vrnd/gen/vrndd-neon-x4.c", |
| 1304 | "src/f32-vrnd/gen/vrndd-neon-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1305 | "src/f32-vunary/gen/vabs-neon-x4.c", |
| 1306 | "src/f32-vunary/gen/vabs-neon-x8.c", |
| 1307 | "src/f32-vunary/gen/vneg-neon-x4.c", |
| 1308 | "src/f32-vunary/gen/vneg-neon-x8.c", |
| 1309 | "src/f32-vunary/gen/vsqr-neon-x4.c", |
| 1310 | "src/f32-vunary/gen/vsqr-neon-x8.c", |
Marat Dukhan | 023bcf9 | 2020-08-10 12:40:50 -0700 | [diff] [blame] | 1311 | "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c", |
| 1312 | "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c", |
| 1313 | "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c", |
| 1314 | "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c", |
Marat Dukhan | 281262d | 2020-08-10 13:23:21 -0700 | [diff] [blame] | 1315 | "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c", |
| 1316 | "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c", |
| 1317 | "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c", |
| 1318 | "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c", |
| 1319 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c", |
| 1320 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c", |
| 1321 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c", |
| 1322 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c", |
Marat Dukhan | 40bbafe | 2020-08-04 02:04:22 -0700 | [diff] [blame] | 1323 | "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c", |
| 1324 | "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c", |
| 1325 | "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c", |
| 1326 | "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c", |
| 1327 | "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c", |
| 1328 | "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c", |
| 1329 | "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c", |
| 1330 | "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 1331 | "src/qs8-requantization/precise-neon.c", |
| 1332 | "src/qs8-requantization/fp32-neon.c", |
| 1333 | "src/qs8-requantization/q31-neon.c", |
Marat Dukhan | ba7b279 | 2020-09-02 14:26:45 -0700 | [diff] [blame] | 1334 | "src/qs8-vadd/gen/minmax-neon-ld64-x8.c", |
| 1335 | "src/qs8-vadd/gen/minmax-neon-ld64-x16.c", |
| 1336 | "src/qs8-vadd/gen/minmax-neon-ld64-x24.c", |
| 1337 | "src/qs8-vadd/gen/minmax-neon-ld64-x32.c", |
| 1338 | "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c", |
| 1339 | "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c", |
| 1340 | "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c", |
| 1341 | "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1342 | "src/qu8-avgpool/9p8x-minmax-neon-c8.c", |
| 1343 | "src/qu8-avgpool/9x-minmax-neon-c8.c", |
| 1344 | "src/qu8-dwconv/up8x9-minmax-neon.c", |
| 1345 | "src/qu8-gavgpool/7p7x-minmax-neon-c8.c", |
| 1346 | "src/qu8-gavgpool/7x-minmax-neon-c8.c", |
| 1347 | "src/qu8-gemm/4x8-minmax-neon.c", |
| 1348 | "src/qu8-gemm/8x8-minmax-neon.c", |
| 1349 | "src/qu8-igemm/4x8-minmax-neon.c", |
| 1350 | "src/qu8-igemm/8x8-minmax-neon.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1351 | "src/qu8-requantization/precise-neon.c", |
| 1352 | "src/qu8-requantization/fp32-neon.c", |
| 1353 | "src/qu8-requantization/q31-neon.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1354 | "src/qu8-vadd/minmax-neon.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1355 | "src/u8-clamp/neon-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1356 | "src/u8-maxpool/9p8x-minmax-neon-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1357 | "src/u8-rmax/neon.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1358 | "src/x32-fill/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1359 | "src/x32-packx/x4-neon-st4.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1360 | "src/x32-pad/neon.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1361 | "src/x32-unpool/neon.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1362 | "src/x32-zip/x2-neon.c", |
| 1363 | "src/x32-zip/x3-neon.c", |
| 1364 | "src/x32-zip/x4-neon.c", |
| 1365 | "src/x32-zip/xm-neon.c", |
| 1366 | "src/x8-zip/x2-neon.c", |
| 1367 | "src/x8-zip/x3-neon.c", |
| 1368 | "src/x8-zip/x4-neon.c", |
| 1369 | "src/x8-zip/xm-neon.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1370 | "src/math/roundne-neon-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1371 | "src/math/roundd-neon-addsub.c", |
| 1372 | "src/math/roundd-neon-cvt.c", |
| 1373 | "src/math/roundu-neon-addsub.c", |
| 1374 | "src/math/roundu-neon-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1375 | "src/math/roundz-neon-addsub.c", |
| 1376 | "src/math/roundz-neon-cvt.c", |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 1377 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1378 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 1379 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 1380 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 1381 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 1382 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 1383 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1384 | "src/math/sqrt-neon-nr1rsqrts.c", |
| 1385 | "src/math/sqrt-neon-nr2rsqrts.c", |
| 1386 | "src/math/sqrt-neon-nr3rsqrts.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1387 | ] |
| 1388 | |
| 1389 | NEONFMA_UKERNELS = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1390 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 1391 | "src/f32-ibilinear/gen/neonfma-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1392 | "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1393 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1394 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1395 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1396 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1397 | "src/f32-igemm/gen/1x8s4-minmax-neonfma.c", |
| 1398 | "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", |
| 1399 | "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", |
| 1400 | "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1401 | "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", |
| 1402 | "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", |
| 1403 | "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", |
| 1404 | "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1405 | "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", |
| 1406 | "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", |
| 1407 | "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", |
| 1408 | "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1409 | "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", |
| 1410 | "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", |
| 1411 | "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", |
| 1412 | "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1413 | "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 1414 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 1415 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 1416 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 1417 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 1418 | "src/f32-gemm/gen/1x8s4-minmax-neonfma.c", |
| 1419 | "src/f32-gemm/gen/4x8s4-minmax-neonfma.c", |
| 1420 | "src/f32-gemm/gen/6x8s4-minmax-neonfma.c", |
| 1421 | "src/f32-gemm/gen/8x8s4-minmax-neonfma.c", |
| 1422 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c", |
| 1423 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c", |
| 1424 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c", |
| 1425 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c", |
| 1426 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c", |
| 1427 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c", |
| 1428 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c", |
| 1429 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c", |
| 1430 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1431 | "src/f32-ppmm/gen/4x8-minmax-neonfma.c", |
| 1432 | "src/f32-ppmm/gen/8x8-minmax-neonfma.c", |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 1433 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 1434 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 1435 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 1436 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 1437 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 1438 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 1439 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 1440 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 1441 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 1442 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
| 1443 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 1444 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 1445 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
| 1446 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
| 1447 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
| 1448 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
| 1449 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 1450 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
| 1451 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
| 1452 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 1453 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
| 1454 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 1455 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 1456 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1457 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 1458 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 1459 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 1460 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 1461 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 1462 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 1463 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 1464 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 1465 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 1466 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 1467 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 1468 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 1469 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 1470 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 1471 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 1472 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 1473 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 1474 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
| 1475 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 1476 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 1477 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 1478 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 1479 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 1480 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
| 1481 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 1482 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 1483 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 1484 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 1485 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 1486 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
| 1487 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 1488 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 1489 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 1490 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 1491 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 1492 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
| 1493 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 1494 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 1495 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 1496 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 1497 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 1498 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
| 1499 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 1500 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 1501 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 1502 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 1503 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 1504 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
| 1505 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 1506 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 1507 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 1508 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 1509 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 1510 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1511 | "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c", |
| 1512 | "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1513 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c", |
| 1514 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c", |
| 1515 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c", |
| 1516 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c", |
| 1517 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c", |
| 1518 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c", |
| 1519 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c", |
| 1520 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c", |
| 1521 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c", |
| 1522 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c", |
| 1523 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c", |
| 1524 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c", |
| 1525 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c", |
| 1526 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c", |
| 1527 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c", |
| 1528 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c", |
| 1529 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c", |
| 1530 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c", |
| 1531 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c", |
| 1532 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c", |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 1533 | "src/math/exp-neonfma-lut64-p2.c", |
| 1534 | "src/math/exp-neonfma-p5.c", |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 1535 | "src/math/expminus-neonfma-lut2048-p1.c", |
| 1536 | "src/math/expminus-neonfma-lut64-p2.c", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1537 | "src/math/expminus-neonfma-p5.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1538 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 1539 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 1540 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
| 1541 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 1542 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 1543 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
| 1544 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 1545 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 1546 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
| 1547 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 1548 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 1549 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
| 1550 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 1551 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 1552 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
| 1553 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 1554 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 1555 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1556 | "src/math/sqrt-neonfma-nr1fma.c", |
| 1557 | "src/math/sqrt-neonfma-nr2fma.c", |
| 1558 | "src/math/sqrt-neonfma-nr3fma.c", |
| 1559 | "src/math/sqrt-neonfma-nr2fma1adj.c", |
| 1560 | "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1561 | ] |
| 1562 | |
| 1563 | AARCH64_NEONFMA_UKERNELS = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1564 | "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c", |
| 1565 | "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c", |
| 1566 | "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c", |
| 1567 | "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c", |
| 1568 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c", |
| 1569 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1570 | "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1571 | "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1572 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1573 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1574 | "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c", |
| 1575 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1576 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
| 1577 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c", |
| 1578 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c", |
| 1579 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c", |
| 1580 | "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c", |
| 1581 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c", |
| 1582 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c", |
| 1583 | "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1584 | "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1585 | "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c", |
| 1586 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1587 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1588 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1589 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1590 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c", |
| 1591 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1592 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c", |
| 1593 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 56b10cd | 2020-05-18 09:35:49 -0700 | [diff] [blame] | 1594 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c", |
| 1595 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c", |
Marat Dukhan | ce7a3f8 | 2020-05-17 21:46:44 -0700 | [diff] [blame] | 1596 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c", |
| 1597 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1598 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c", |
| 1599 | "src/f32-dwconv-chw/3x3p1-neonfma.c", |
| 1600 | "src/f32-dwconv-chw/5x5p2-neonfma.c", |
| 1601 | "src/f32-dwconv-chw/3x3s2p1-neonfma.c", |
| 1602 | "src/f32-dwconv-chw/5x5s2p2-neonfma.c", |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1603 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 1604 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 1605 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 1606 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 1607 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 1608 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
| 1609 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 1610 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 1611 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 1612 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 1613 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 1614 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 1615 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 1616 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 1617 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 1618 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 1619 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 1620 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1621 | "src/f32-spmm/gen/12x1-minmax-neonfma.c", |
| 1622 | "src/f32-spmm/gen/12x2-minmax-neonfma.c", |
| 1623 | "src/f32-spmm/gen/12x4-minmax-neonfma.c", |
| 1624 | "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c", |
| 1625 | "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c", |
| 1626 | "src/f32-spmm/gen/16x1-minmax-neonfma.c", |
| 1627 | "src/f32-spmm/gen/16x2-minmax-neonfma.c", |
| 1628 | "src/f32-spmm/gen/16x4-minmax-neonfma.c", |
| 1629 | "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c", |
| 1630 | "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c", |
| 1631 | "src/f32-spmm/gen/4x1-minmax-neonfma.c", |
| 1632 | "src/f32-spmm/gen/4x2-minmax-neonfma.c", |
| 1633 | "src/f32-spmm/gen/4x4-minmax-neonfma.c", |
| 1634 | "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c", |
| 1635 | "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c", |
| 1636 | "src/f32-spmm/gen/8x1-minmax-neonfma.c", |
| 1637 | "src/f32-spmm/gen/8x2-minmax-neonfma.c", |
| 1638 | "src/f32-spmm/gen/8x4-minmax-neonfma.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1639 | "src/f32-vsqrt/gen/neon-sqrt-x4.c", |
| 1640 | "src/f32-vsqrt/gen/neon-sqrt-x8.c", |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 1641 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
| 1642 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
| 1643 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
| 1644 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
| 1645 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
| 1646 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1647 | ] |
| 1648 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1649 | NEONV8_UKERNELS = [ |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1650 | "src/f32-vrnd/gen/vrndne-neonv8-x4.c", |
| 1651 | "src/f32-vrnd/gen/vrndne-neonv8-x8.c", |
| 1652 | "src/f32-vrnd/gen/vrndz-neonv8-x4.c", |
| 1653 | "src/f32-vrnd/gen/vrndz-neonv8-x8.c", |
| 1654 | "src/f32-vrnd/gen/vrndu-neonv8-x4.c", |
| 1655 | "src/f32-vrnd/gen/vrndu-neonv8-x8.c", |
| 1656 | "src/f32-vrnd/gen/vrndd-neonv8-x4.c", |
| 1657 | "src/f32-vrnd/gen/vrndd-neonv8-x8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1658 | "src/math/roundne-neonv8.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1659 | "src/math/roundd-neonv8.c", |
| 1660 | "src/math/roundu-neonv8.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1661 | "src/math/roundz-neonv8.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 1662 | ] |
| 1663 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1664 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1665 | "src/f16-clamp/gen/neonfp16arith-x16.c", |
| 1666 | "src/f16-clamp/gen/neonfp16arith-x8.c", |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 1667 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c", |
| 1668 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c", |
| 1669 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c", |
| 1670 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c", |
| 1671 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c", |
| 1672 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c", |
| 1673 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c", |
| 1674 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c", |
| 1675 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c", |
| 1676 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c", |
| 1677 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c", |
| 1678 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c", |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 1679 | "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c", |
| 1680 | "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1681 | "src/f16-hswish/gen/hswish-neonfp16arith-x16.c", |
| 1682 | "src/f16-hswish/gen/hswish-neonfp16arith-x8.c", |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 1683 | "src/f16-prelu/gen/neonfp16arith-2x8.c", |
Frank Barchard | a531698 | 2020-07-23 13:19:28 -0700 | [diff] [blame] | 1684 | "src/f16-prelu/gen/neonfp16arith-2x16.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1685 | "src/f16-relu/gen/neonfp16arith-x16.c", |
| 1686 | "src/f16-relu/gen/neonfp16arith-x8.c", |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 1687 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c", |
| 1688 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c", |
| 1689 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c", |
| 1690 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c", |
| 1691 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c", |
| 1692 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c", |
| 1693 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c", |
| 1694 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c", |
| 1695 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c", |
| 1696 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c", |
| 1697 | "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c", |
| 1698 | "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c", |
| 1699 | "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c", |
| 1700 | "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c", |
| 1701 | "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c", |
| 1702 | "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c", |
| 1703 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c", |
| 1704 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c", |
| 1705 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c", |
| 1706 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c", |
| 1707 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c", |
| 1708 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c", |
| 1709 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c", |
| 1710 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c", |
| 1711 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c", |
| 1712 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c", |
| 1713 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c", |
| 1714 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c", |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 1715 | "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c", |
| 1716 | "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c", |
Frank Barchard | 83ea939 | 2020-04-20 20:15:14 -0700 | [diff] [blame] | 1717 | "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1718 | "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1719 | "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1720 | "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1721 | "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c", |
| 1722 | "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c", |
| 1723 | "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c", |
| 1724 | "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c", |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 1725 | "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1726 | "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1727 | "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1728 | "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
Frank Barchard | 3f9f99f | 2020-05-06 01:12:04 -0700 | [diff] [blame] | 1729 | "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1730 | "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1731 | "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1732 | "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
| 1733 | "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c", |
| 1734 | "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c", |
| 1735 | "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c", |
| 1736 | "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c", |
| 1737 | "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1738 | "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1739 | "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1740 | "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1741 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c", |
| 1742 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c", |
| 1743 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c", |
| 1744 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c", |
| 1745 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c", |
| 1746 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c", |
| 1747 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c", |
| 1748 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1749 | ] |
| 1750 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1751 | NEONDOT_UKERNELS = [ |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1752 | "src/qs8-gemm/gen/1x8c4-minmax-neondot.c", |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1753 | "src/qs8-gemm/gen/4x8c4-minmax-neondot.c", |
| 1754 | "src/qs8-gemm/gen/6x8c4-minmax-neondot.c", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1755 | "src/qs8-gemm/gen/8x8c4-minmax-neondot.c", |
| 1756 | "src/qs8-gemm/gen/12x8c4-minmax-neondot.c", |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1757 | "src/qs8-gemm/gen/1x16c4-minmax-neondot.c", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1758 | "src/qs8-gemm/gen/4x16c4-minmax-neondot.c", |
Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1759 | "src/qs8-gemm/gen/6x16c4-minmax-neondot.c", |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1760 | "src/qs8-gemm/gen/8x16c4-minmax-neondot.c", |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1761 | "src/qs8-igemm/gen/1x8c4-minmax-neondot.c", |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1762 | "src/qs8-igemm/gen/4x8c4-minmax-neondot.c", |
| 1763 | "src/qs8-igemm/gen/6x8c4-minmax-neondot.c", |
Benoit Jacob | a05487f | 2020-08-14 13:37:21 -0700 | [diff] [blame] | 1764 | "src/qs8-igemm/gen/8x8c4-minmax-neondot.c", |
| 1765 | "src/qs8-igemm/gen/12x8c4-minmax-neondot.c", |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1766 | "src/qs8-igemm/gen/1x16c4-minmax-neondot.c", |
Benoit Jacob | a05487f | 2020-08-14 13:37:21 -0700 | [diff] [blame] | 1767 | "src/qs8-igemm/gen/4x16c4-minmax-neondot.c", |
Frank Barchard | 2fa1745 | 2020-09-22 09:51:46 -0700 | [diff] [blame] | 1768 | "src/qs8-igemm/gen/6x16c4-minmax-neondot.c", |
Frank Barchard | a48848f | 2020-09-24 00:59:24 -0700 | [diff] [blame] | 1769 | "src/qs8-igemm/gen/8x16c4-minmax-neondot.c", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1770 | ] |
| 1771 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1772 | SSE_UKERNELS = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1773 | "src/f32-avgpool/9p8x-minmax-sse-c4.c", |
| 1774 | "src/f32-avgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1775 | "src/f32-clamp/gen/sse-x4.c", |
| 1776 | "src/f32-clamp/gen/sse-x8.c", |
Erich Elsen | b123340 | 2020-06-08 15:53:15 -0700 | [diff] [blame] | 1777 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c", |
| 1778 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1779 | "src/f32-dwconv-chw/3x3p1-sse.c", |
| 1780 | "src/f32-dwconv-chw/3x3s2p1-sse.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1781 | "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c", |
| 1782 | "src/f32-dwconv/gen/up4x25-minmax-sse.c", |
| 1783 | "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c", |
| 1784 | "src/f32-dwconv/gen/up4x4-minmax-sse.c", |
| 1785 | "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c", |
| 1786 | "src/f32-dwconv/gen/up4x9-minmax-sse.c", |
| 1787 | "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c", |
| 1788 | "src/f32-dwconv/gen/up8x25-minmax-sse.c", |
| 1789 | "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c", |
| 1790 | "src/f32-dwconv/gen/up8x4-minmax-sse.c", |
| 1791 | "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c", |
| 1792 | "src/f32-dwconv/gen/up8x9-minmax-sse.c", |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1793 | "src/f32-gavgpool-cw/sse-x4.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1794 | "src/f32-gavgpool/7p7x-minmax-sse-c4.c", |
| 1795 | "src/f32-gavgpool/7x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1796 | "src/f32-gemm/gen/1x8-minmax-sse-dup.c", |
| 1797 | "src/f32-gemm/gen/1x8-minmax-sse-load1.c", |
| 1798 | "src/f32-gemm/gen/1x8s4-minmax-sse.c", |
| 1799 | "src/f32-gemm/gen/4x2c4-minmax-sse.c", |
| 1800 | "src/f32-gemm/gen/4x8-minmax-sse-dup.c", |
| 1801 | "src/f32-gemm/gen/4x8-minmax-sse-load1.c", |
| 1802 | "src/f32-gemm/gen/4x8s4-minmax-sse.c", |
| 1803 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c", |
| 1804 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c", |
| 1805 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c", |
| 1806 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c", |
| 1807 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c", |
| 1808 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 1809 | "src/f32-hswish/gen/hswish-sse-x4.c", |
| 1810 | "src/f32-hswish/gen/hswish-sse-x8.c", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1811 | "src/f32-ibilinear/gen/sse-c4.c", |
| 1812 | "src/f32-ibilinear/gen/sse-c8.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1813 | "src/f32-igemm/gen/1x8-minmax-sse-dup.c", |
| 1814 | "src/f32-igemm/gen/1x8-minmax-sse-load1.c", |
| 1815 | "src/f32-igemm/gen/1x8s4-minmax-sse.c", |
| 1816 | "src/f32-igemm/gen/4x2c4-minmax-sse.c", |
| 1817 | "src/f32-igemm/gen/4x8-minmax-sse-dup.c", |
| 1818 | "src/f32-igemm/gen/4x8-minmax-sse-load1.c", |
| 1819 | "src/f32-igemm/gen/4x8s4-minmax-sse.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1820 | "src/f32-maxpool/9p8x-minmax-sse-c4.c", |
| 1821 | "src/f32-pavgpool/9p8x-minmax-sse-c4.c", |
| 1822 | "src/f32-pavgpool/9x-minmax-sse-c4.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 1823 | "src/f32-ppmm/gen/4x8-minmax-sse.c", |
Marat Dukhan | 39b5e94 | 2020-06-24 15:03:48 -0700 | [diff] [blame] | 1824 | "src/f32-prelu/gen/sse-2x4.c", |
| 1825 | "src/f32-prelu/gen/sse-2x8.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 1826 | "src/f32-relu/gen/sse-x4.c", |
| 1827 | "src/f32-relu/gen/sse-x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1828 | "src/f32-rmax/sse.c", |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1829 | "src/f32-spmm/gen/4x1-minmax-sse.c", |
| 1830 | "src/f32-spmm/gen/8x1-minmax-sse.c", |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 1831 | "src/f32-spmm/gen/16x1-minmax-sse.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1832 | "src/f32-vbinary/gen/vadd-minmax-sse-x4.c", |
| 1833 | "src/f32-vbinary/gen/vadd-minmax-sse-x8.c", |
| 1834 | "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c", |
| 1835 | "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c", |
| 1836 | "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c", |
| 1837 | "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c", |
| 1838 | "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c", |
| 1839 | "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c", |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 1840 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 1841 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 1842 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 1843 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 1844 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 1845 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 1846 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 1847 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1848 | "src/f32-vbinary/gen/vmul-minmax-sse-x4.c", |
| 1849 | "src/f32-vbinary/gen/vmul-minmax-sse-x8.c", |
| 1850 | "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c", |
| 1851 | "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c", |
| 1852 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c", |
| 1853 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c", |
| 1854 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c", |
| 1855 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 1856 | "src/f32-vbinary/gen/vsqrdiff-sse-x4.c", |
| 1857 | "src/f32-vbinary/gen/vsqrdiff-sse-x8.c", |
| 1858 | "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c", |
| 1859 | "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 1860 | "src/f32-vbinary/gen/vsub-minmax-sse-x4.c", |
| 1861 | "src/f32-vbinary/gen/vsub-minmax-sse-x8.c", |
| 1862 | "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c", |
| 1863 | "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1864 | "src/f32-vlrelu/gen/vlrelu-sse-x4.c", |
| 1865 | "src/f32-vlrelu/gen/vlrelu-sse-x8.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1866 | "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c", |
| 1867 | "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 1868 | "src/f32-vsqrt/gen/sse-sqrt-x4.c", |
| 1869 | "src/f32-vsqrt/gen/sse-sqrt-x8.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1870 | "src/f32-vunary/gen/vabs-sse-x4.c", |
| 1871 | "src/f32-vunary/gen/vabs-sse-x8.c", |
| 1872 | "src/f32-vunary/gen/vneg-sse-x4.c", |
| 1873 | "src/f32-vunary/gen/vneg-sse-x8.c", |
| 1874 | "src/f32-vunary/gen/vsqr-sse-x4.c", |
| 1875 | "src/f32-vunary/gen/vsqr-sse-x8.c", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 1876 | "src/x32-fill/sse.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1877 | "src/x32-packx/x4-sse.c", |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1878 | "src/x32-pad/sse.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 1879 | "src/math/roundne-sse-addsub.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 1880 | "src/math/roundd-sse-addsub.c", |
| 1881 | "src/math/roundu-sse-addsub.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 1882 | "src/math/roundz-sse-addsub.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 1883 | "src/math/sqrt-sse-hh1mac.c", |
| 1884 | "src/math/sqrt-sse-nr1mac.c", |
| 1885 | "src/math/sqrt-sse-nr2mac.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1886 | ] |
| 1887 | |
| 1888 | SSE2_UKERNELS = [ |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1889 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
| 1890 | "src/f32-argmaxpool/4x-sse2-c4.c", |
| 1891 | "src/f32-argmaxpool/9x-sse2-c4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1892 | "src/f32-prelu/gen/sse2-2x4.c", |
| 1893 | "src/f32-prelu/gen/sse2-2x8.c", |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 1894 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
| 1895 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
| 1896 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
| 1897 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
| 1898 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 1899 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
| 1900 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
| 1901 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 1902 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
| 1903 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 1904 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 1905 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 1906 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c", |
| 1907 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c", |
| 1908 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c", |
| 1909 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c", |
| 1910 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c", |
| 1911 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1912 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1913 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1914 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 1915 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 1916 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 1917 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 1918 | "src/f32-vlrelu/gen/vlrelu-sse2-x4.c", |
| 1919 | "src/f32-vlrelu/gen/vlrelu-sse2-x8.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1920 | "src/f32-vrnd/gen/vrndne-sse2-x4.c", |
| 1921 | "src/f32-vrnd/gen/vrndne-sse2-x8.c", |
| 1922 | "src/f32-vrnd/gen/vrndz-sse2-x4.c", |
| 1923 | "src/f32-vrnd/gen/vrndz-sse2-x8.c", |
| 1924 | "src/f32-vrnd/gen/vrndu-sse2-x4.c", |
| 1925 | "src/f32-vrnd/gen/vrndu-sse2-x8.c", |
| 1926 | "src/f32-vrnd/gen/vrndd-sse2-x4.c", |
| 1927 | "src/f32-vrnd/gen/vrndd-sse2-x8.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 1928 | "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c", |
| 1929 | "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c", |
| 1930 | "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c", |
Marat Dukhan | 4ed53f4 | 2020-08-06 01:12:55 -0700 | [diff] [blame] | 1931 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c", |
| 1932 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c", |
| 1933 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c", |
Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 1934 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c", |
| 1935 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c", |
| 1936 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 1937 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c", |
| 1938 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 1939 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c", |
| 1940 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1941 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c", |
| 1942 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 1943 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c", |
| 1944 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 1945 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 1946 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c", |
| 1947 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 1948 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1949 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c", |
| 1950 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c", |
| 1951 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 1952 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c", |
| 1953 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 1954 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c", |
| 1955 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 1956 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c", |
| 1957 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 1958 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 1959 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c", |
| 1960 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 1961 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 1962 | "src/qs8-requantization/precise-sse2.c", |
| 1963 | "src/qs8-requantization/fp32-sse2.c", |
| 1964 | "src/qs8-requantization/q31-sse2.c", |
Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 1965 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c", |
| 1966 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c", |
| 1967 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c", |
| 1968 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c", |
Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 1969 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c", |
| 1970 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c", |
| 1971 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c", |
| 1972 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1973 | "src/qu8-avgpool/9p8x-minmax-sse2-c8.c", |
| 1974 | "src/qu8-avgpool/9x-minmax-sse2-c8.c", |
| 1975 | "src/qu8-igemm/4x4c2-minmax-sse2.c", |
| 1976 | "src/qu8-dwconv/up8x9-minmax-sse2.c", |
| 1977 | "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c", |
| 1978 | "src/qu8-gavgpool/7x-minmax-sse2-c8.c", |
| 1979 | "src/qu8-gemm/2x4c8-minmax-sse2.c", |
| 1980 | "src/qu8-gemm/4x4c2-minmax-sse2.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 1981 | "src/qu8-requantization/precise-sse2.c", |
| 1982 | "src/qu8-requantization/fp32-sse2.c", |
| 1983 | "src/qu8-requantization/q31-sse2.c", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1984 | "src/qu8-vadd/minmax-sse2.c", |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1985 | "src/u8-clamp/sse2-x64.c", |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1986 | "src/u8-maxpool/9p8x-minmax-sse2-c16.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1987 | "src/u8-rmax/sse2.c", |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1988 | "src/x32-unpool/sse2.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 1989 | "src/x32-zip/x2-sse2.c", |
| 1990 | "src/x32-zip/x3-sse2.c", |
| 1991 | "src/x32-zip/x4-sse2.c", |
| 1992 | "src/x32-zip/xm-sse2.c", |
| 1993 | "src/x8-zip/x2-sse2.c", |
| 1994 | "src/x8-zip/x3-sse2.c", |
| 1995 | "src/x8-zip/x4-sse2.c", |
| 1996 | "src/x8-zip/xm-sse2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 1997 | "src/math/exp-sse2-p5.c", |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 1998 | "src/math/exp-sse2-lut64-p2.c", |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 1999 | "src/math/expminus-sse2-p5.c", |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 2000 | "src/math/roundne-sse2-cvt.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 2001 | "src/math/roundd-sse2-cvt.c", |
| 2002 | "src/math/roundu-sse2-cvt.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 2003 | "src/math/roundz-sse2-cvt.c", |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 2004 | "src/math/sigmoid-sse2-rr2-p5-div.c", |
| 2005 | "src/math/sigmoid-sse2-rr2-lut64-p2-div.c", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 2006 | ] |
| 2007 | |
| 2008 | SSSE3_UKERNELS = [ |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2009 | "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c", |
| 2010 | "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c", |
| 2011 | "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c", |
Marat Dukhan | 4ed53f4 | 2020-08-06 01:12:55 -0700 | [diff] [blame] | 2012 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c", |
| 2013 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c", |
| 2014 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c", |
Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2015 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c", |
| 2016 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c", |
| 2017 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2018 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c", |
| 2019 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2020 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c", |
| 2021 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2022 | "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c", |
| 2023 | "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2024 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c", |
| 2025 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2026 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2027 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c", |
| 2028 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2029 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2030 | "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c", |
| 2031 | "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c", |
| 2032 | "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2033 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c", |
| 2034 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2035 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c", |
| 2036 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2037 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c", |
| 2038 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2039 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2040 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c", |
| 2041 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2042 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 2043 | "src/qs8-requantization/precise-ssse3.c", |
| 2044 | "src/qs8-requantization/q31-ssse3.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2045 | "src/qu8-requantization/precise-ssse3.c", |
| 2046 | "src/qu8-requantization/q31-ssse3.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2047 | ] |
| 2048 | |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2049 | SSE41_UKERNELS = [ |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2050 | "src/f32-prelu/gen/sse41-2x4.c", |
| 2051 | "src/f32-prelu/gen/sse41-2x8.c", |
Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 2052 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c", |
| 2053 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c", |
| 2054 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c", |
| 2055 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c", |
| 2056 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c", |
| 2057 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2058 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2059 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2060 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
Marat Dukhan | 40a672f | 2019-11-25 03:08:22 -0800 | [diff] [blame] | 2061 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 2062 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 2063 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2064 | "src/f32-vlrelu/gen/vlrelu-sse41-x4.c", |
| 2065 | "src/f32-vlrelu/gen/vlrelu-sse41-x8.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2066 | "src/f32-vrnd/gen/vrndne-sse41-x4.c", |
| 2067 | "src/f32-vrnd/gen/vrndne-sse41-x8.c", |
| 2068 | "src/f32-vrnd/gen/vrndz-sse41-x4.c", |
| 2069 | "src/f32-vrnd/gen/vrndz-sse41-x8.c", |
| 2070 | "src/f32-vrnd/gen/vrndu-sse41-x4.c", |
| 2071 | "src/f32-vrnd/gen/vrndu-sse41-x8.c", |
| 2072 | "src/f32-vrnd/gen/vrndd-sse41-x4.c", |
| 2073 | "src/f32-vrnd/gen/vrndd-sse41-x8.c", |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 2074 | "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c", |
| 2075 | "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c", |
| 2076 | "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c", |
Marat Dukhan | 4ed53f4 | 2020-08-06 01:12:55 -0700 | [diff] [blame] | 2077 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c", |
| 2078 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c", |
| 2079 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c", |
Marat Dukhan | 159688f | 2020-08-06 10:34:29 -0700 | [diff] [blame] | 2080 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c", |
| 2081 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c", |
| 2082 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2083 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c", |
| 2084 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2085 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c", |
| 2086 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2087 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c", |
| 2088 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2089 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c", |
| 2090 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2091 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2092 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c", |
| 2093 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2094 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2095 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c", |
| 2096 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c", |
| 2097 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c", |
Marat Dukhan | 14d3ce8 | 2020-07-31 16:19:15 -0700 | [diff] [blame] | 2098 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c", |
| 2099 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c", |
Marat Dukhan | dee732b | 2020-07-31 17:32:04 -0700 | [diff] [blame] | 2100 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c", |
| 2101 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2102 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c", |
| 2103 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2104 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c", |
Marat Dukhan | 07bd252 | 2020-07-31 19:12:39 -0700 | [diff] [blame] | 2105 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c", |
| 2106 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2107 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c", |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 2108 | "src/qs8-requantization/fp32-sse4.c", |
| 2109 | "src/qs8-requantization/precise-sse4.c", |
| 2110 | "src/qs8-requantization/q31-sse4.c", |
Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 2111 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c", |
| 2112 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c", |
| 2113 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c", |
| 2114 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c", |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2115 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c", |
| 2116 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c", |
| 2117 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c", |
| 2118 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c", |
Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 2119 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c", |
| 2120 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c", |
| 2121 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c", |
| 2122 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c", |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2123 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c", |
| 2124 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c", |
| 2125 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c", |
| 2126 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c", |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 2127 | "src/qu8-requantization/precise-sse4.c", |
| 2128 | "src/qu8-requantization/q31-sse4.c", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 2129 | "src/math/roundne-sse41.c", |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 2130 | "src/math/roundd-sse41.c", |
| 2131 | "src/math/roundu-sse41.c", |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 2132 | "src/math/roundz-sse41.c", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 2133 | ] |
| 2134 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2135 | AVX_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2136 | "src/f32-clamp/gen/avx-x8.c", |
| 2137 | "src/f32-clamp/gen/avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2138 | "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c", |
| 2139 | "src/f32-dwconv/gen/up16x4-minmax-avx.c", |
| 2140 | "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c", |
| 2141 | "src/f32-dwconv/gen/up8x4-minmax-avx.c", |
| 2142 | "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c", |
| 2143 | "src/f32-dwconv/gen/up16x9-minmax-avx.c", |
| 2144 | "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c", |
| 2145 | "src/f32-dwconv/gen/up8x9-minmax-avx.c", |
| 2146 | "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c", |
| 2147 | "src/f32-dwconv/gen/up16x25-minmax-avx.c", |
| 2148 | "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c", |
| 2149 | "src/f32-dwconv/gen/up8x25-minmax-avx.c", |
| 2150 | "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c", |
| 2151 | "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c", |
| 2152 | "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c", |
| 2153 | "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c", |
| 2154 | "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c", |
| 2155 | "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c", |
| 2156 | "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c", |
| 2157 | "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c", |
| 2158 | "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c", |
| 2159 | "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c", |
| 2160 | "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c", |
| 2161 | "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c", |
| 2162 | "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c", |
| 2163 | "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c", |
| 2164 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c", |
| 2165 | "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c", |
| 2166 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c", |
| 2167 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2168 | "src/f32-hswish/gen/hswish-avx-x8.c", |
| 2169 | "src/f32-hswish/gen/hswish-avx-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2170 | "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c", |
| 2171 | "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c", |
| 2172 | "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c", |
| 2173 | "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c", |
| 2174 | "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c", |
| 2175 | "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c", |
| 2176 | "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c", |
| 2177 | "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c", |
| 2178 | "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2179 | "src/f32-prelu/gen/avx-2x8.c", |
| 2180 | "src/f32-prelu/gen/avx-2x16.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2181 | "src/f32-relu/gen/avx-x16.c", |
| 2182 | "src/f32-relu/gen/avx-x8.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2183 | "src/f32-rmax/avx.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2184 | "src/f32-vbinary/gen/vadd-minmax-avx-x8.c", |
| 2185 | "src/f32-vbinary/gen/vadd-minmax-avx-x16.c", |
| 2186 | "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c", |
| 2187 | "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c", |
| 2188 | "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c", |
| 2189 | "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c", |
| 2190 | "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c", |
| 2191 | "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2192 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 2193 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 2194 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 2195 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 2196 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 2197 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 2198 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 2199 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2200 | "src/f32-vbinary/gen/vmul-minmax-avx-x8.c", |
| 2201 | "src/f32-vbinary/gen/vmul-minmax-avx-x16.c", |
| 2202 | "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c", |
| 2203 | "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c", |
| 2204 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c", |
| 2205 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c", |
| 2206 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c", |
| 2207 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2208 | "src/f32-vbinary/gen/vsqrdiff-avx-x8.c", |
| 2209 | "src/f32-vbinary/gen/vsqrdiff-avx-x16.c", |
| 2210 | "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c", |
| 2211 | "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2212 | "src/f32-vbinary/gen/vsub-minmax-avx-x8.c", |
| 2213 | "src/f32-vbinary/gen/vsub-minmax-avx-x16.c", |
| 2214 | "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c", |
| 2215 | "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2216 | "src/f32-vlrelu/gen/vlrelu-avx-x8.c", |
| 2217 | "src/f32-vlrelu/gen/vlrelu-avx-x16.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2218 | "src/f32-vscale/avx-unroll32.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2219 | "src/f32-vrnd/gen/vrndne-avx-x8.c", |
| 2220 | "src/f32-vrnd/gen/vrndne-avx-x16.c", |
| 2221 | "src/f32-vrnd/gen/vrndz-avx-x8.c", |
| 2222 | "src/f32-vrnd/gen/vrndz-avx-x16.c", |
| 2223 | "src/f32-vrnd/gen/vrndu-avx-x8.c", |
| 2224 | "src/f32-vrnd/gen/vrndu-avx-x16.c", |
| 2225 | "src/f32-vrnd/gen/vrndd-avx-x8.c", |
| 2226 | "src/f32-vrnd/gen/vrndd-avx-x16.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2227 | "src/f32-vsqrt/gen/avx-sqrt-x8.c", |
| 2228 | "src/f32-vsqrt/gen/avx-sqrt-x16.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2229 | "src/f32-vunary/gen/vabs-avx-x8.c", |
| 2230 | "src/f32-vunary/gen/vabs-avx-x16.c", |
| 2231 | "src/f32-vunary/gen/vneg-avx-x8.c", |
| 2232 | "src/f32-vunary/gen/vneg-avx-x16.c", |
| 2233 | "src/f32-vunary/gen/vsqr-avx-x8.c", |
| 2234 | "src/f32-vunary/gen/vsqr-avx-x16.c", |
Marat Dukhan | d08abcb | 2020-09-21 00:26:25 -0700 | [diff] [blame] | 2235 | "src/math/exp-avx-p5.c", |
Marat Dukhan | 2e4727b | 2020-09-24 00:48:39 -0700 | [diff] [blame] | 2236 | "src/math/sigmoid-avx-rr2-p5-div.c", |
| 2237 | "src/math/sigmoid-avx-rr2-lut64-p2-div.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2238 | ] |
| 2239 | |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2240 | XOP_UKERNELS = [ |
| 2241 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c", |
| 2242 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c", |
| 2243 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c", |
| 2244 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2245 | "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c", |
| 2246 | "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2247 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c", |
| 2248 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2249 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2250 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c", |
| 2251 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2252 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2253 | "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c", |
| 2254 | "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c", |
| 2255 | "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2256 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c", |
| 2257 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c", |
| 2258 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c", |
| 2259 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c", |
| 2260 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c", |
| 2261 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2262 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2263 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c", |
| 2264 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c", |
Marat Dukhan | e7edc80 | 2020-08-03 16:43:57 -0700 | [diff] [blame] | 2265 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c", |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 2266 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c", |
| 2267 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c", |
| 2268 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c", |
| 2269 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c", |
| 2270 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c", |
| 2271 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c", |
| 2272 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c", |
| 2273 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 2274 | ] |
| 2275 | |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2276 | FMA3_UKERNELS = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2277 | "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c", |
| 2278 | "src/f32-dwconv/gen/up16x4-minmax-fma3.c", |
| 2279 | "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c", |
| 2280 | "src/f32-dwconv/gen/up8x4-minmax-fma3.c", |
| 2281 | "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c", |
| 2282 | "src/f32-dwconv/gen/up16x9-minmax-fma3.c", |
| 2283 | "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c", |
| 2284 | "src/f32-dwconv/gen/up8x9-minmax-fma3.c", |
| 2285 | "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c", |
| 2286 | "src/f32-dwconv/gen/up16x25-minmax-fma3.c", |
| 2287 | "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c", |
| 2288 | "src/f32-dwconv/gen/up8x25-minmax-fma3.c", |
| 2289 | "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c", |
| 2290 | "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c", |
| 2291 | "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c", |
| 2292 | "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c", |
| 2293 | "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c", |
| 2294 | "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c", |
| 2295 | "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c", |
| 2296 | "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c", |
| 2297 | "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c", |
| 2298 | "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c", |
| 2299 | "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 2300 | "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 2301 | "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 2302 | "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| 2303 | "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c", |
| 2304 | "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c", |
| 2305 | "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c", |
| 2306 | "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c", |
| 2307 | "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c", |
| 2308 | "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c", |
| 2309 | "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c", |
| 2310 | "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c", |
| 2311 | "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c", |
| 2312 | "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c", |
| 2313 | "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c", |
| 2314 | "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c", |
| 2315 | "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c", |
| 2316 | "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2317 | "src/f32-hswish/gen/hswish-fma3-x8.c", |
| 2318 | "src/f32-hswish/gen/hswish-fma3-x16.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2319 | "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c", |
| 2320 | "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c", |
| 2321 | "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c", |
| 2322 | "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c", |
| 2323 | "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c", |
| 2324 | "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c", |
| 2325 | "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c", |
| 2326 | "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c", |
| 2327 | "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c", |
| 2328 | "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c", |
| 2329 | "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 2330 | "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 2331 | "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 2332 | "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2333 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c", |
| 2334 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c", |
| 2335 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c", |
| 2336 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c", |
| 2337 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c", |
| 2338 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c", |
| 2339 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c", |
| 2340 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2341 | "src/math/sqrt-fma3-nr1fma.c", |
| 2342 | "src/math/sqrt-fma3-nr2fma.c", |
| 2343 | "src/math/sqrt-fma3-nr1fma1adj.c", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 2344 | ] |
| 2345 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2346 | AVX2_UKERNELS = [ |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2347 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
| 2348 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 2349 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 2350 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
| 2351 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 2352 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
| 2353 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 2354 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 2355 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
| 2356 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 2357 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 2358 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
| 2359 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
| 2360 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 2361 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
| 2362 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
| 2363 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
| 2364 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
| 2365 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 2366 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
| 2367 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
| 2368 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 2369 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 2370 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
| 2371 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
| 2372 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 2373 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
| 2374 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
| 2375 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
| 2376 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
| 2377 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 2378 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
| 2379 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
| 2380 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 2381 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 2382 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 2383 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 2384 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 2385 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 2386 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 2387 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 2388 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 2389 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 2390 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 2391 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 2392 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 2393 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 2394 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 2395 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 2396 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 2397 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 2398 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 2399 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 2400 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 2401 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 2402 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 2403 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 2404 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 2405 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 2406 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 2407 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 2408 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 2409 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 2410 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 2411 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 2412 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2413 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 2414 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 2415 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 2416 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 2417 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 2418 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 2419 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 2420 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 2421 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 2422 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 2423 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 2424 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 2425 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 2426 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 2427 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 2428 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 2429 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 2430 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 2431 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 2432 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 2433 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 2434 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 2435 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 2436 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Marat Dukhan | 4013552 | 2020-08-07 01:21:00 -0700 | [diff] [blame] | 2437 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c", |
| 2438 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c", |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 2439 | "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c", |
| 2440 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c", |
| 2441 | "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c", |
| 2442 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c", |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 2443 | "src/qs8-gemm/gen/1x8c8-minmax-avx2.c", |
| 2444 | "src/qs8-gemm/gen/2x8c8-minmax-avx2.c", |
| 2445 | "src/qs8-gemm/gen/3x8c8-minmax-avx2.c", |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 2446 | "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c", |
| 2447 | "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c", |
| 2448 | "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c", |
Marat Dukhan | 1280952 | 2020-08-02 22:23:51 -0700 | [diff] [blame] | 2449 | "src/qs8-igemm/gen/1x8c8-minmax-avx2.c", |
| 2450 | "src/qs8-igemm/gen/2x8c8-minmax-avx2.c", |
| 2451 | "src/qs8-igemm/gen/3x8c8-minmax-avx2.c", |
Marat Dukhan | e6dc0b6 | 2020-09-08 23:57:14 -0700 | [diff] [blame] | 2452 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2453 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2454 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2455 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c", |
| 2456 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2457 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2458 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2459 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2460 | "src/math/exp-avx2-p5.c", |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 2461 | "src/math/exp-avx2-lut8-p3-perm.c", |
| 2462 | "src/math/exp-avx2-lut8-p4-perm.c", |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 2463 | "src/math/expminus-avx2-p5.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2464 | "src/math/extexp-avx2-p5.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 2465 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2466 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 2467 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2468 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 2469 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2470 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2471 | ] |
| 2472 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2473 | AVX512F_UKERNELS = [ |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2474 | "src/f32-clamp/gen/avx512f-x16.c", |
| 2475 | "src/f32-clamp/gen/avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2476 | "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c", |
| 2477 | "src/f32-dwconv/gen/up32x4-minmax-avx512f.c", |
| 2478 | "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c", |
| 2479 | "src/f32-dwconv/gen/up16x4-minmax-avx512f.c", |
| 2480 | "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c", |
| 2481 | "src/f32-dwconv/gen/up32x9-minmax-avx512f.c", |
| 2482 | "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c", |
| 2483 | "src/f32-dwconv/gen/up16x9-minmax-avx512f.c", |
| 2484 | "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c", |
| 2485 | "src/f32-dwconv/gen/up32x25-minmax-avx512f.c", |
| 2486 | "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c", |
| 2487 | "src/f32-dwconv/gen/up16x25-minmax-avx512f.c", |
| 2488 | "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2489 | "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2490 | "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2491 | "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2492 | "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2493 | "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c", |
| 2494 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c", |
| 2495 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c", |
| 2496 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c", |
| 2497 | "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c", |
| 2498 | "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c", |
| 2499 | "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c", |
Marat Dukhan | efaac27 | 2020-07-16 16:58:07 -0700 | [diff] [blame] | 2500 | "src/f32-hswish/gen/hswish-avx512f-x16.c", |
| 2501 | "src/f32-hswish/gen/hswish-avx512f-x32.c", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2502 | "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2503 | "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2504 | "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2505 | "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2506 | "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2507 | "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c", |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 2508 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 2509 | "src/f32-prelu/gen/avx512f-2x32.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2510 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2511 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2512 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 2513 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2514 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 2515 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2516 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2517 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 2518 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2519 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2520 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2521 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
| 2522 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
| 2523 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2524 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
| 2525 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
| 2526 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
| 2527 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
| 2528 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2529 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
| 2530 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
| 2531 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2532 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2533 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
| 2534 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2535 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2536 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
| 2537 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2538 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
| 2539 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2540 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2541 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
| 2542 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2543 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2544 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2545 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 2546 | "src/f32-relu/gen/avx512f-x16.c", |
| 2547 | "src/f32-relu/gen/avx512f-x32.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2548 | "src/f32-rmax/avx512f.c", |
Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 2549 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c", |
| 2550 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c", |
| 2551 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c", |
| 2552 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c", |
| 2553 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c", |
| 2554 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c", |
| 2555 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c", |
| 2556 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c", |
| 2557 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c", |
| 2558 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c", |
| 2559 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c", |
| 2560 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c", |
| 2561 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c", |
| 2562 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c", |
| 2563 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c", |
| 2564 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c", |
| 2565 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c", |
| 2566 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c", |
| 2567 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c", |
| 2568 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c", |
| 2569 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c", |
| 2570 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c", |
| 2571 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c", |
| 2572 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c", |
| 2573 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c", |
| 2574 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c", |
| 2575 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c", |
| 2576 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c", |
| 2577 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c", |
| 2578 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c", |
| 2579 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c", |
| 2580 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c", |
| 2581 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c", |
| 2582 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c", |
| 2583 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c", |
| 2584 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c", |
| 2585 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c", |
| 2586 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c", |
| 2587 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c", |
| 2588 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c", |
| 2589 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c", |
| 2590 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c", |
| 2591 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c", |
| 2592 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c", |
| 2593 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c", |
| 2594 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c", |
| 2595 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c", |
| 2596 | "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] | 2597 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c", |
| 2598 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c", |
| 2599 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c", |
| 2600 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c", |
| 2601 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c", |
| 2602 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c", |
| 2603 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c", |
| 2604 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c", |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 2605 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 2606 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 2607 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 2608 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 2609 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 2610 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 2611 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 2612 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2613 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c", |
| 2614 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c", |
| 2615 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c", |
| 2616 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c", |
| 2617 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c", |
| 2618 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c", |
| 2619 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c", |
| 2620 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 2621 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c", |
| 2622 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c", |
| 2623 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c", |
| 2624 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c", |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 2625 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c", |
| 2626 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c", |
| 2627 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c", |
| 2628 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c", |
Marat Dukhan | 19dd91d | 2020-07-16 11:12:44 -0700 | [diff] [blame] | 2629 | "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c", |
| 2630 | "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2631 | "src/f32-vscale/avx512f-unroll64.c", |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 2632 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 2633 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 2634 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 2635 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 2636 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 2637 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 2638 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 2639 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2640 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2641 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2642 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 2643 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2644 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 2645 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 2646 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 2647 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 2648 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 2649 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 2650 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 2651 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 2652 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 2653 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 2654 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 2655 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 2656 | "src/f32-vrnd/gen/vrndne-avx512f-x16.c", |
| 2657 | "src/f32-vrnd/gen/vrndne-avx512f-x32.c", |
| 2658 | "src/f32-vrnd/gen/vrndz-avx512f-x16.c", |
| 2659 | "src/f32-vrnd/gen/vrndz-avx512f-x32.c", |
| 2660 | "src/f32-vrnd/gen/vrndu-avx512f-x16.c", |
| 2661 | "src/f32-vrnd/gen/vrndu-avx512f-x32.c", |
| 2662 | "src/f32-vrnd/gen/vrndd-avx512f-x16.c", |
| 2663 | "src/f32-vrnd/gen/vrndd-avx512f-x32.c", |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 2664 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c", |
| 2665 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c", |
| 2666 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c", |
| 2667 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c", |
| 2668 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c", |
| 2669 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c", |
| 2670 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c", |
| 2671 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c", |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2672 | "src/f32-vunary/gen/vabs-avx512f-x16.c", |
| 2673 | "src/f32-vunary/gen/vabs-avx512f-x32.c", |
| 2674 | "src/f32-vunary/gen/vneg-avx512f-x16.c", |
| 2675 | "src/f32-vunary/gen/vneg-avx512f-x32.c", |
| 2676 | "src/f32-vunary/gen/vsqr-avx512f-x16.c", |
| 2677 | "src/f32-vunary/gen/vsqr-avx512f-x32.c", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2678 | "src/math/exp-avx512f-p5-scalef.c", |
| 2679 | "src/math/exp-avx512f-p5.c", |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 2680 | "src/math/exp-avx512f-lut16-p3-perm.c", |
Marat Dukhan | 80a8ac5 | 2020-09-21 17:00:27 -0700 | [diff] [blame] | 2681 | "src/math/exp-avx512f-lut16-p3-perm-scalef.c", |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 2682 | "src/math/exp-avx512f-lut32-p2-perm2.c", |
Marat Dukhan | 80a8ac5 | 2020-09-21 17:00:27 -0700 | [diff] [blame] | 2683 | "src/math/exp-avx512f-lut32-p2-perm2-scalef.c", |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 2684 | "src/math/extexp-avx512f-p5.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2685 | "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2686 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2687 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2688 | "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2689 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c", |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 2690 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c", |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 2691 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c", |
| 2692 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c", |
| 2693 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c", |
| 2694 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c", |
| 2695 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c", |
| 2696 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c", |
| 2697 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c", |
| 2698 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c", |
| 2699 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c", |
| 2700 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c", |
| 2701 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c", |
| 2702 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c", |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 2703 | "src/math/sqrt-avx512f-nr1fma.c", |
| 2704 | "src/math/sqrt-avx512f-nr2fma.c", |
| 2705 | "src/math/sqrt-avx512f-nr1fma1adj.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2706 | ] |
| 2707 | |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 2708 | AVX512SKX_UKERNELS = [ |
Marat Dukhan | 2ffc5e6 | 2020-09-06 22:33:38 -0700 | [diff] [blame] | 2709 | "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c", |
| 2710 | "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 2711 | "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c", |
| 2712 | "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c", |
| 2713 | "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c", |
| 2714 | "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c", |
| 2715 | "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c", |
| 2716 | "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c", |
| 2717 | "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c", |
| 2718 | "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c", |
| 2719 | ] |
| 2720 | |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 2721 | WASM32_ASM_UKERNELS = [ |
| 2722 | "src/f32-relu/wasm_shr_x1.S", |
| 2723 | "src/f32-relu/wasm_shr_x2.S", |
| 2724 | "src/f32-relu/wasm_shr_x4.S", |
| 2725 | ] |
| 2726 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2727 | AARCH32_ASM_UKERNELS = [ |
Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 2728 | "src/f32-gemm/4x4-aarch32-vfp-ld64.S", |
Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 2729 | "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2730 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2731 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 2732 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2733 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 2734 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 2735 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2736 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2737 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 2738 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
| 2739 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 2740 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| 2741 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2742 | ] |
| 2743 | |
| 2744 | AARCH64_ASM_UKERNELS = [ |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 2745 | "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 2746 | "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2747 | "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 2748 | "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 2749 | "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
| 2750 | "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2751 | "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2752 | "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2753 | "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 2754 | "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 2755 | "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| 2756 | "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
| 2757 | "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 2758 | "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2759 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S", |
| 2760 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S", |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 2761 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2762 | "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2763 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2764 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2765 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2766 | "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2767 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2768 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2769 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2770 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2771 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S", |
| 2772 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S", |
| 2773 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2774 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2775 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2776 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2777 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2778 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2779 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2780 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S", |
| 2781 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S", |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 2782 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2783 | "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2784 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2785 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2786 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2787 | "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2788 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2789 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2790 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2791 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2792 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2793 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2794 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2795 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2796 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2797 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2798 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S", |
| 2799 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2800 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 2801 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2802 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2803 | "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2804 | "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2805 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2806 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2807 | "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2808 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2809 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2810 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2811 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2812 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2813 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2814 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2815 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2816 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2817 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2818 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Frank Barchard | f1fd89e | 2020-10-06 23:27:11 -0700 | [diff] [blame^] | 2819 | "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S", |
Frank Barchard | a5237a5 | 2020-10-06 16:18:46 -0700 | [diff] [blame] | 2820 | "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2821 | ] |
| 2822 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2823 | INTERNAL_MICROKERNEL_HDRS = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2824 | "src/xnnpack/argmaxpool.h", |
| 2825 | "src/xnnpack/avgpool.h", |
| 2826 | "src/xnnpack/clamp.h", |
| 2827 | "src/xnnpack/common.h", |
| 2828 | "src/xnnpack/conv.h", |
| 2829 | "src/xnnpack/dwconv.h", |
| 2830 | "src/xnnpack/gavgpool.h", |
| 2831 | "src/xnnpack/gemm.h", |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 2832 | "src/xnnpack/fill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2833 | "src/xnnpack/hswish.h", |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2834 | "src/xnnpack/ibilinear.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2835 | "src/xnnpack/igemm.h", |
Marat Dukhan | cfb3134 | 2019-12-05 10:42:57 -0800 | [diff] [blame] | 2836 | "src/xnnpack/intrinsics-polyfill.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2837 | "src/xnnpack/lut.h", |
| 2838 | "src/xnnpack/math.h", |
| 2839 | "src/xnnpack/maxpool.h", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2840 | "src/xnnpack/memory.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2841 | "src/xnnpack/packx.h", |
| 2842 | "src/xnnpack/pad.h", |
| 2843 | "src/xnnpack/params.h", |
| 2844 | "src/xnnpack/pavgpool.h", |
| 2845 | "src/xnnpack/ppmm.h", |
| 2846 | "src/xnnpack/prelu.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2847 | "src/xnnpack/raddexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2848 | "src/xnnpack/raddextexp.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2849 | "src/xnnpack/raddstoreexpminusmax.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2850 | "src/xnnpack/rmax.h", |
| 2851 | "src/xnnpack/scalar-utils.h", |
| 2852 | "src/xnnpack/spmm.h", |
| 2853 | "src/xnnpack/unpool.h", |
| 2854 | "src/xnnpack/vadd.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2855 | "src/xnnpack/vbinary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2856 | "src/xnnpack/vmulcaddc.h", |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 2857 | "src/xnnpack/vscale.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 2858 | "src/xnnpack/vscaleexpminusmax.h", |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 2859 | "src/xnnpack/vscaleextexp.h", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2860 | "src/xnnpack/vunary.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2861 | "src/xnnpack/zip.h", |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2862 | ] |
| 2863 | |
| 2864 | INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2865 | "include/xnnpack.h", |
| 2866 | "src/xnnpack/allocator.h", |
| 2867 | "src/xnnpack/compute.h", |
| 2868 | "src/xnnpack/im2col.h", |
| 2869 | "src/xnnpack/indirection.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2870 | "src/xnnpack/math-stubs.h", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 2871 | "src/xnnpack/memory-planner.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2872 | "src/xnnpack/operator.h", |
| 2873 | "src/xnnpack/pack.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2874 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2875 | "src/xnnpack/requantization-stubs.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2876 | "src/xnnpack/requantization.h", |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 2877 | "src/xnnpack/subgraph.h", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2878 | ] |
| 2879 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2880 | ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 2881 | "src/xnnpack/math-stubs.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2882 | ] |
| 2883 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2884 | MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2885 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2886 | "include/xnnpack.h", |
| 2887 | ] |
| 2888 | |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 2889 | MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2890 | "src/xnnpack/isa-checks.h", |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 2891 | "src/xnnpack/params-init.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2892 | "src/xnnpack/requantization.h", |
| 2893 | "include/xnnpack.h", |
| 2894 | ] |
| 2895 | |
| 2896 | OPERATOR_TEST_PARAMS_HDRS = [ |
| 2897 | "src/xnnpack/params.h", |
| 2898 | "src/xnnpack/common.h", |
| 2899 | ] |
| 2900 | |
| 2901 | WEIGHTS_PACK_HDRS = [ |
| 2902 | "src/xnnpack/pack.h", |
| 2903 | "src/xnnpack/operator.h", |
| 2904 | "src/xnnpack/compute.h", |
| 2905 | ] |
| 2906 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2907 | LOGGING_COPTS = select({ |
| 2908 | # No logging in optimized mode |
| 2909 | ":optimized_build": ["-DXNN_LOG_LEVEL=0"], |
| 2910 | # Full logging in debug mode |
| 2911 | ":debug_build": ["-DXNN_LOG_LEVEL=5"], |
| 2912 | # Error-only logging in default (fastbuild) mode |
| 2913 | "//conditions:default": ["-DXNN_LOG_LEVEL=2"], |
| 2914 | }) |
| 2915 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 2916 | LOGGING_SRCS = select({ |
| 2917 | # No logging in optimized mode |
| 2918 | ":optimized_build": [], |
| 2919 | "//conditions:default": [ |
| 2920 | "src/operator-strings.c", |
| 2921 | "src/subgraph-strings.c", |
| 2922 | ], |
| 2923 | }) |
| 2924 | |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 2925 | LOGGING_HDRS = [ |
| 2926 | "src/xnnpack/log.h", |
| 2927 | ] |
| 2928 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2929 | xnnpack_cc_library( |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2930 | name = "tables", |
| 2931 | srcs = TABLE_SRCS, |
| 2932 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2933 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2934 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2935 | ) |
| 2936 | |
| 2937 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2938 | name = "scalar_ukernels", |
| 2939 | srcs = SCALAR_UKERNELS, |
| 2940 | hdrs = INTERNAL_HDRS, |
| 2941 | aarch32_copts = ["-marm"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2942 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2943 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2944 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2945 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2946 | "@FP16", |
| 2947 | "@FXdiv", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2948 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 2949 | ], |
| 2950 | ) |
| 2951 | |
| 2952 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2953 | name = "scalar_ukernels_test_mode", |
| 2954 | srcs = SCALAR_UKERNELS, |
| 2955 | hdrs = INTERNAL_HDRS, |
| 2956 | aarch32_copts = ["-marm"], |
| 2957 | copts = [ |
| 2958 | "-UNDEBUG", |
| 2959 | "-DXNN_TEST_MODE=1", |
| 2960 | ], |
| 2961 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2962 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2963 | deps = [ |
| 2964 | ":tables", |
| 2965 | "@FP16", |
| 2966 | "@FXdiv", |
| 2967 | "@pthreadpool", |
| 2968 | ], |
| 2969 | ) |
| 2970 | |
| 2971 | xnnpack_cc_library( |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2972 | name = "wasm_ukernels", |
| 2973 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 2974 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2975 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2976 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 1c6e389 | 2020-06-25 23:56:10 -0700 | [diff] [blame] | 2977 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2978 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2979 | ":tables", |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 2980 | "@FP16", |
| 2981 | "@FXdiv", |
| 2982 | "@pthreadpool", |
| 2983 | ], |
| 2984 | ) |
| 2985 | |
| 2986 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2987 | name = "wasm_ukernels_test_mode", |
| 2988 | hdrs = INTERNAL_HDRS, |
| 2989 | copts = [ |
| 2990 | "-UNDEBUG", |
| 2991 | "-DXNN_TEST_MODE=1", |
| 2992 | ], |
| 2993 | gcc_copts = xnnpack_gcc_std_copts(), |
| 2994 | msvc_copts = xnnpack_msvc_std_copts(), |
| 2995 | wasm_srcs = WASM_UKERNELS, |
Marat Dukhan | 0e97d6f | 2020-06-26 19:35:09 -0700 | [diff] [blame] | 2996 | wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 2997 | deps = [ |
| 2998 | ":tables", |
| 2999 | "@FP16", |
| 3000 | "@FXdiv", |
| 3001 | "@pthreadpool", |
| 3002 | ], |
| 3003 | ) |
| 3004 | |
| 3005 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3006 | name = "psimd_fastmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3007 | hdrs = INTERNAL_HDRS, |
| 3008 | aarch32_copts = [ |
| 3009 | "-marm", |
| 3010 | "-mfpu=neon", |
| 3011 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3012 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3013 | gcc_x86_copts = ["-msse2"], |
| 3014 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3015 | optimized_copts = [ |
| 3016 | "-O3", |
| 3017 | "-ffast-math", |
| 3018 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3019 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3020 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3021 | ":tables", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3022 | "@FP16", |
| 3023 | "@psimd", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3024 | "@pthreadpool", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3025 | ], |
| 3026 | ) |
| 3027 | |
| 3028 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3029 | name = "psimd_fastmath_ukernels_test_mode", |
| 3030 | hdrs = INTERNAL_HDRS, |
| 3031 | aarch32_copts = [ |
| 3032 | "-marm", |
| 3033 | "-mfpu=neon", |
| 3034 | ], |
| 3035 | copts = [ |
| 3036 | "-UNDEBUG", |
| 3037 | "-DXNN_TEST_MODE=1", |
| 3038 | ], |
| 3039 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3040 | gcc_x86_copts = ["-msse2"], |
| 3041 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3042 | optimized_copts = [ |
| 3043 | "-O3", |
| 3044 | "-ffast-math", |
| 3045 | ], |
| 3046 | psimd_srcs = PSIMD_FASTMATH_UKERNELS, |
| 3047 | deps = [ |
| 3048 | ":tables", |
| 3049 | "@FP16", |
| 3050 | "@psimd", |
| 3051 | "@pthreadpool", |
| 3052 | ], |
| 3053 | ) |
| 3054 | |
| 3055 | xnnpack_cc_library( |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3056 | name = "psimd_accmath_ukernels", |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3057 | hdrs = INTERNAL_HDRS, |
| 3058 | aarch32_copts = [ |
| 3059 | "-marm", |
| 3060 | "-mfpu=neon", |
| 3061 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3062 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3063 | gcc_x86_copts = ["-msse2"], |
| 3064 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3065 | optimized_copts = [ |
| 3066 | "-O3", |
| 3067 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3068 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3069 | deps = [ |
| 3070 | ":tables", |
| 3071 | "@FP16", |
| 3072 | "@psimd", |
| 3073 | "@pthreadpool", |
| 3074 | ], |
| 3075 | ) |
| 3076 | |
| 3077 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3078 | name = "psimd_accmath_ukernels_test_mode", |
| 3079 | hdrs = INTERNAL_HDRS, |
| 3080 | aarch32_copts = [ |
| 3081 | "-marm", |
| 3082 | "-mfpu=neon", |
| 3083 | ], |
| 3084 | copts = [ |
| 3085 | "-UNDEBUG", |
| 3086 | "-DXNN_TEST_MODE=1", |
| 3087 | ], |
| 3088 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3089 | gcc_x86_copts = ["-msse2"], |
| 3090 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3091 | optimized_copts = [ |
| 3092 | "-O3", |
| 3093 | ], |
| 3094 | psimd_srcs = PSIMD_ACCMATH_UKERNELS, |
| 3095 | deps = [ |
| 3096 | ":tables", |
| 3097 | "@FP16", |
| 3098 | "@psimd", |
| 3099 | "@pthreadpool", |
| 3100 | ], |
| 3101 | ) |
| 3102 | |
| 3103 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3104 | name = "neon_ukernels", |
| 3105 | hdrs = INTERNAL_HDRS, |
| 3106 | aarch32_copts = [ |
| 3107 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3108 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3109 | "-mfpu=neon", |
| 3110 | ], |
| 3111 | aarch32_srcs = NEON_UKERNELS, |
| 3112 | aarch64_srcs = NEON_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3113 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3114 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3115 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3116 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3117 | "@FP16", |
| 3118 | "@pthreadpool", |
| 3119 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3120 | ) |
| 3121 | |
| 3122 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3123 | name = "neon_ukernels_test_mode", |
| 3124 | hdrs = INTERNAL_HDRS, |
| 3125 | aarch32_copts = [ |
| 3126 | "-marm", |
| 3127 | "-march=armv7-a", |
| 3128 | "-mfpu=neon", |
| 3129 | ], |
| 3130 | aarch32_srcs = NEON_UKERNELS, |
| 3131 | aarch64_srcs = NEON_UKERNELS, |
| 3132 | copts = [ |
| 3133 | "-UNDEBUG", |
| 3134 | "-DXNN_TEST_MODE=1", |
| 3135 | ], |
| 3136 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3137 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3138 | deps = [ |
| 3139 | ":tables", |
| 3140 | "@FP16", |
| 3141 | "@pthreadpool", |
| 3142 | ], |
| 3143 | ) |
| 3144 | |
| 3145 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3146 | name = "neonfma_ukernels", |
| 3147 | hdrs = INTERNAL_HDRS, |
| 3148 | aarch32_copts = [ |
| 3149 | "-marm", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3150 | "-march=armv7-a", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3151 | "-mfpu=neon-vfpv4", |
| 3152 | ], |
| 3153 | aarch32_srcs = NEONFMA_UKERNELS, |
| 3154 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3155 | apple_aarch32_copts = [ |
| 3156 | "-mcpu=swift", |
| 3157 | "-mtune=generic", |
| 3158 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3159 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3160 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3161 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3162 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3163 | "@FP16", |
| 3164 | "@pthreadpool", |
| 3165 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3166 | ) |
| 3167 | |
| 3168 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3169 | name = "neonfma_ukernels_test_mode", |
| 3170 | hdrs = INTERNAL_HDRS, |
| 3171 | aarch32_copts = [ |
| 3172 | "-marm", |
| 3173 | "-march=armv7-a", |
| 3174 | "-mfpu=neon-vfpv4", |
| 3175 | ], |
| 3176 | aarch32_srcs = NEONFMA_UKERNELS, |
| 3177 | aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3178 | apple_aarch32_copts = [ |
| 3179 | "-mcpu=swift", |
| 3180 | "-mtune=generic", |
| 3181 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3182 | copts = [ |
| 3183 | "-UNDEBUG", |
| 3184 | "-DXNN_TEST_MODE=1", |
| 3185 | ], |
| 3186 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3187 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3188 | deps = [ |
| 3189 | ":tables", |
| 3190 | "@FP16", |
| 3191 | "@pthreadpool", |
| 3192 | ], |
| 3193 | ) |
| 3194 | |
| 3195 | xnnpack_cc_library( |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3196 | name = "neonv8_ukernels", |
| 3197 | hdrs = INTERNAL_HDRS, |
| 3198 | aarch32_copts = [ |
| 3199 | "-marm", |
| 3200 | "-march=armv8-a", |
| 3201 | "-mfpu=neon-fp-armv8", |
| 3202 | ], |
| 3203 | aarch32_srcs = NEONV8_UKERNELS, |
| 3204 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3205 | apple_aarch32_copts = [ |
| 3206 | "-mcpu=cyclone", |
| 3207 | "-mtune=generic", |
| 3208 | ], |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3209 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3210 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3211 | deps = [ |
| 3212 | ":tables", |
| 3213 | "@FP16", |
| 3214 | "@pthreadpool", |
| 3215 | ], |
| 3216 | ) |
| 3217 | |
| 3218 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3219 | name = "neonv8_ukernels_test_mode", |
| 3220 | hdrs = INTERNAL_HDRS, |
| 3221 | aarch32_copts = [ |
| 3222 | "-marm", |
| 3223 | "-march=armv8-a", |
| 3224 | "-mfpu=neon-fp-armv8", |
| 3225 | ], |
| 3226 | aarch32_srcs = NEONV8_UKERNELS, |
| 3227 | aarch64_srcs = NEONV8_UKERNELS, |
Marat Dukhan | bc69ed6 | 2020-06-09 21:34:56 -0700 | [diff] [blame] | 3228 | apple_aarch32_copts = [ |
| 3229 | "-mcpu=cyclone", |
| 3230 | "-mtune=generic", |
| 3231 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3232 | copts = [ |
| 3233 | "-UNDEBUG", |
| 3234 | "-DXNN_TEST_MODE=1", |
| 3235 | ], |
| 3236 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3237 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3238 | deps = [ |
| 3239 | ":tables", |
| 3240 | "@FP16", |
| 3241 | "@pthreadpool", |
| 3242 | ], |
| 3243 | ) |
| 3244 | |
| 3245 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3246 | name = "neonfp16arith_ukernels", |
| 3247 | hdrs = INTERNAL_HDRS, |
| 3248 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 3249 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3250 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3251 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3252 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3253 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3254 | "@FP16", |
| 3255 | "@pthreadpool", |
| 3256 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3257 | ) |
| 3258 | |
| 3259 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3260 | name = "neonfp16arith_ukernels_test_mode", |
| 3261 | hdrs = INTERNAL_HDRS, |
| 3262 | aarch64_copts = ["-march=armv8.2-a+fp16"], |
| 3263 | aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS, |
| 3264 | copts = [ |
| 3265 | "-UNDEBUG", |
| 3266 | "-DXNN_TEST_MODE=1", |
| 3267 | ], |
| 3268 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3269 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3270 | deps = [ |
| 3271 | ":tables", |
| 3272 | "@FP16", |
| 3273 | "@pthreadpool", |
| 3274 | ], |
| 3275 | ) |
| 3276 | |
| 3277 | xnnpack_cc_library( |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3278 | name = "neondot_ukernels", |
| 3279 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 799ac75 | 2020-08-13 16:08:03 -0700 | [diff] [blame] | 3280 | aarch32_copts = [ |
| 3281 | "-marm", |
| 3282 | "-march=armv8.2-a+dotprod", |
| 3283 | "-mfpu=neon-fp-armv8", |
| 3284 | ], |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3285 | aarch32_srcs = NEONDOT_UKERNELS, |
| 3286 | aarch64_copts = ["-march=armv8.2-a+dotprod"], |
| 3287 | aarch64_srcs = NEONDOT_UKERNELS, |
| 3288 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3289 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3290 | deps = [ |
| 3291 | ":tables", |
| 3292 | "@FP16", |
| 3293 | "@pthreadpool", |
| 3294 | ], |
| 3295 | ) |
| 3296 | |
| 3297 | xnnpack_cc_library( |
| 3298 | name = "neondot_ukernels_test_mode", |
| 3299 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 799ac75 | 2020-08-13 16:08:03 -0700 | [diff] [blame] | 3300 | aarch32_copts = [ |
| 3301 | "-marm", |
| 3302 | "-march=armv8.2-a+dotprod", |
| 3303 | "-mfpu=neon-fp-armv8", |
| 3304 | ], |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3305 | aarch32_srcs = NEONDOT_UKERNELS, |
| 3306 | aarch64_copts = ["-march=armv8.2-a+dotprod"], |
| 3307 | aarch64_srcs = NEONDOT_UKERNELS, |
| 3308 | copts = [ |
| 3309 | "-UNDEBUG", |
| 3310 | "-DXNN_TEST_MODE=1", |
| 3311 | ], |
| 3312 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3313 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3314 | deps = [ |
| 3315 | ":tables", |
| 3316 | "@FP16", |
| 3317 | "@pthreadpool", |
| 3318 | ], |
| 3319 | ) |
| 3320 | |
| 3321 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3322 | name = "sse2_ukernels", |
| 3323 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3324 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3325 | gcc_x86_copts = ["-msse2"], |
| 3326 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3327 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3328 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3329 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3330 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3331 | "@FP16", |
| 3332 | "@pthreadpool", |
| 3333 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3334 | ) |
| 3335 | |
| 3336 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3337 | name = "sse2_ukernels_test_mode", |
| 3338 | hdrs = INTERNAL_HDRS, |
| 3339 | copts = [ |
| 3340 | "-UNDEBUG", |
| 3341 | "-DXNN_TEST_MODE=1", |
| 3342 | ], |
| 3343 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3344 | gcc_x86_copts = ["-msse2"], |
| 3345 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3346 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3347 | x86_srcs = SSE_UKERNELS + SSE2_UKERNELS, |
| 3348 | deps = [ |
| 3349 | ":tables", |
| 3350 | "@FP16", |
| 3351 | "@pthreadpool", |
| 3352 | ], |
| 3353 | ) |
| 3354 | |
| 3355 | xnnpack_cc_library( |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3356 | name = "ssse3_ukernels", |
| 3357 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3358 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3359 | gcc_x86_copts = ["-mssse3"], |
| 3360 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3361 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3362 | x86_srcs = SSSE3_UKERNELS, |
| 3363 | deps = [ |
| 3364 | ":tables", |
| 3365 | "@FP16", |
| 3366 | "@pthreadpool", |
| 3367 | ], |
| 3368 | ) |
| 3369 | |
| 3370 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3371 | name = "ssse3_ukernels_test_mode", |
| 3372 | hdrs = INTERNAL_HDRS, |
| 3373 | copts = [ |
| 3374 | "-UNDEBUG", |
| 3375 | "-DXNN_TEST_MODE=1", |
| 3376 | ], |
| 3377 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3378 | gcc_x86_copts = ["-mssse3"], |
| 3379 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3380 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3381 | x86_srcs = SSSE3_UKERNELS, |
| 3382 | deps = [ |
| 3383 | ":tables", |
| 3384 | "@FP16", |
| 3385 | "@pthreadpool", |
| 3386 | ], |
| 3387 | ) |
| 3388 | |
| 3389 | xnnpack_cc_library( |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3390 | name = "sse41_ukernels", |
| 3391 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3392 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3393 | gcc_x86_copts = ["-msse4.1"], |
| 3394 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3395 | msvc_x86_32_copts = ["/arch:SSE2"], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3396 | x86_srcs = SSE41_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3397 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3398 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3399 | "@FP16", |
| 3400 | "@pthreadpool", |
| 3401 | ], |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3402 | ) |
| 3403 | |
| 3404 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3405 | name = "sse41_ukernels_test_mode", |
| 3406 | hdrs = INTERNAL_HDRS, |
| 3407 | copts = [ |
| 3408 | "-UNDEBUG", |
| 3409 | "-DXNN_TEST_MODE=1", |
| 3410 | ], |
| 3411 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3412 | gcc_x86_copts = ["-msse4.1"], |
| 3413 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3414 | msvc_x86_32_copts = ["/arch:SSE2"], |
| 3415 | x86_srcs = SSE41_UKERNELS, |
| 3416 | deps = [ |
| 3417 | ":tables", |
| 3418 | "@FP16", |
| 3419 | "@pthreadpool", |
| 3420 | ], |
| 3421 | ) |
| 3422 | |
| 3423 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3424 | name = "avx_ukernels", |
| 3425 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3426 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3427 | gcc_x86_copts = ["-mavx"], |
| 3428 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3429 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3430 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3431 | x86_srcs = AVX_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3432 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3433 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3434 | "@FP16", |
| 3435 | "@pthreadpool", |
| 3436 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3437 | ) |
| 3438 | |
| 3439 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3440 | name = "avx_ukernels_test_mode", |
| 3441 | hdrs = INTERNAL_HDRS, |
| 3442 | copts = [ |
| 3443 | "-UNDEBUG", |
| 3444 | "-DXNN_TEST_MODE=1", |
| 3445 | ], |
| 3446 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3447 | gcc_x86_copts = ["-mavx"], |
| 3448 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3449 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3450 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3451 | x86_srcs = AVX_UKERNELS, |
| 3452 | deps = [ |
| 3453 | ":tables", |
| 3454 | "@FP16", |
| 3455 | "@pthreadpool", |
| 3456 | ], |
| 3457 | ) |
| 3458 | |
| 3459 | xnnpack_cc_library( |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3460 | name = "xop_ukernels", |
| 3461 | hdrs = INTERNAL_HDRS, |
| 3462 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3463 | gcc_x86_copts = ["-mxop"], |
| 3464 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3465 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3466 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3467 | x86_srcs = XOP_UKERNELS, |
| 3468 | deps = [ |
| 3469 | ":tables", |
| 3470 | "@FP16", |
| 3471 | "@pthreadpool", |
| 3472 | ], |
| 3473 | ) |
| 3474 | |
| 3475 | xnnpack_cc_library( |
| 3476 | name = "xop_ukernels_test_mode", |
| 3477 | hdrs = INTERNAL_HDRS, |
| 3478 | copts = [ |
| 3479 | "-UNDEBUG", |
| 3480 | "-DXNN_TEST_MODE=1", |
| 3481 | ], |
| 3482 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3483 | gcc_x86_copts = ["-mxop"], |
| 3484 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3485 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3486 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3487 | x86_srcs = XOP_UKERNELS, |
| 3488 | deps = [ |
| 3489 | ":tables", |
| 3490 | "@FP16", |
| 3491 | "@pthreadpool", |
| 3492 | ], |
| 3493 | ) |
| 3494 | |
| 3495 | xnnpack_cc_library( |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3496 | name = "fma3_ukernels", |
| 3497 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3498 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3499 | gcc_x86_copts = ["-mfma"], |
| 3500 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3501 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3502 | msvc_x86_64_copts = ["/arch:AVX"], |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3503 | x86_srcs = FMA3_UKERNELS, |
| 3504 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3505 | ":tables", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3506 | "@FP16", |
| 3507 | "@pthreadpool", |
| 3508 | ], |
| 3509 | ) |
| 3510 | |
| 3511 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3512 | name = "fma3_ukernels_test_mode", |
| 3513 | hdrs = INTERNAL_HDRS, |
| 3514 | copts = [ |
| 3515 | "-UNDEBUG", |
| 3516 | "-DXNN_TEST_MODE=1", |
| 3517 | ], |
| 3518 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3519 | gcc_x86_copts = ["-mfma"], |
| 3520 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3521 | msvc_x86_32_copts = ["/arch:AVX"], |
| 3522 | msvc_x86_64_copts = ["/arch:AVX"], |
| 3523 | x86_srcs = FMA3_UKERNELS, |
| 3524 | deps = [ |
| 3525 | ":tables", |
| 3526 | "@FP16", |
| 3527 | "@pthreadpool", |
| 3528 | ], |
| 3529 | ) |
| 3530 | |
| 3531 | xnnpack_cc_library( |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3532 | name = "avx2_ukernels", |
| 3533 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3534 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3535 | gcc_x86_copts = [ |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3536 | "-mfma", |
| 3537 | "-mavx2", |
| 3538 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3539 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3540 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 3541 | msvc_x86_64_copts = ["/arch:AVX2"], |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3542 | x86_srcs = AVX2_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 | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3548 | ) |
| 3549 | |
| 3550 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3551 | name = "avx2_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 = [ |
| 3559 | "-mfma", |
| 3560 | "-mavx2", |
| 3561 | ], |
| 3562 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3563 | msvc_x86_32_copts = ["/arch:AVX2"], |
| 3564 | msvc_x86_64_copts = ["/arch:AVX2"], |
| 3565 | x86_srcs = AVX2_UKERNELS, |
| 3566 | deps = [ |
| 3567 | ":tables", |
| 3568 | "@FP16", |
| 3569 | "@pthreadpool", |
| 3570 | ], |
| 3571 | ) |
| 3572 | |
| 3573 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3574 | name = "avx512f_ukernels", |
| 3575 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3576 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3577 | gcc_x86_copts = ["-mavx512f"], |
| 3578 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3579 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3580 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3581 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3582 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3583 | x86_srcs = AVX512F_UKERNELS, |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3584 | deps = [ |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 3585 | ":tables", |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 3586 | "@FP16", |
| 3587 | "@pthreadpool", |
| 3588 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3589 | ) |
| 3590 | |
| 3591 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3592 | name = "avx512f_ukernels_test_mode", |
| 3593 | hdrs = INTERNAL_HDRS, |
| 3594 | copts = [ |
| 3595 | "-UNDEBUG", |
| 3596 | "-DXNN_TEST_MODE=1", |
| 3597 | ], |
| 3598 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3599 | gcc_x86_copts = ["-mavx512f"], |
| 3600 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3601 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3602 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3603 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3604 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3605 | x86_srcs = AVX512F_UKERNELS, |
| 3606 | deps = [ |
| 3607 | ":tables", |
| 3608 | "@FP16", |
| 3609 | "@pthreadpool", |
| 3610 | ], |
| 3611 | ) |
| 3612 | |
| 3613 | xnnpack_cc_library( |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3614 | name = "avx512skx_ukernels", |
| 3615 | hdrs = INTERNAL_HDRS, |
| 3616 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3617 | gcc_x86_copts = [ |
| 3618 | "-mavx512f", |
| 3619 | "-mavx512cd", |
| 3620 | "-mavx512bw", |
| 3621 | "-mavx512dq", |
| 3622 | "-mavx512vl", |
| 3623 | ], |
| 3624 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3625 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3626 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3627 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3628 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3629 | x86_srcs = AVX512SKX_UKERNELS, |
| 3630 | deps = [ |
| 3631 | ":tables", |
| 3632 | "@FP16", |
| 3633 | "@pthreadpool", |
| 3634 | ], |
| 3635 | ) |
| 3636 | |
| 3637 | xnnpack_cc_library( |
| 3638 | name = "avx512skx_ukernels_test_mode", |
| 3639 | hdrs = INTERNAL_HDRS, |
| 3640 | copts = [ |
| 3641 | "-UNDEBUG", |
| 3642 | "-DXNN_TEST_MODE=1", |
| 3643 | ], |
| 3644 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3645 | gcc_x86_copts = [ |
| 3646 | "-mavx512f", |
| 3647 | "-mavx512cd", |
| 3648 | "-mavx512bw", |
| 3649 | "-mavx512dq", |
| 3650 | "-mavx512vl", |
| 3651 | ], |
| 3652 | mingw_copts = ["-fno-asynchronous-unwind-tables"], |
| 3653 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3654 | msvc_x86_32_copts = ["/arch:AVX512"], |
| 3655 | msvc_x86_64_copts = ["/arch:AVX512"], |
| 3656 | msys_copts = ["-fno-asynchronous-unwind-tables"], |
| 3657 | x86_srcs = AVX512SKX_UKERNELS, |
| 3658 | deps = [ |
| 3659 | ":tables", |
| 3660 | "@FP16", |
| 3661 | "@pthreadpool", |
| 3662 | ], |
| 3663 | ) |
| 3664 | |
| 3665 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3666 | name = "asm_ukernels", |
| 3667 | hdrs = ["src/xnnpack/assembly.h"], |
| 3668 | aarch32_srcs = AARCH32_ASM_UKERNELS, |
Frank Barchard | 31bb45b | 2020-10-06 00:26:33 -0700 | [diff] [blame] | 3669 | aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3670 | aarch64_srcs = AARCH64_ASM_UKERNELS, |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3671 | wasm_srcs = WASM32_ASM_UKERNELS, |
| 3672 | wasmsimd_srcs = WASM32_ASM_UKERNELS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3673 | ) |
| 3674 | |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3675 | xnnpack_cc_library( |
| 3676 | name = "logging_utils", |
| 3677 | srcs = LOGGING_SRCS, |
| 3678 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3679 | copts = LOGGING_COPTS + [ |
| 3680 | "-Isrc", |
| 3681 | "-Iinclude", |
| 3682 | ] + select({ |
| 3683 | ":debug_build": [], |
| 3684 | "//conditions:default": xnnpack_min_size_copts(), |
| 3685 | }), |
| 3686 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3687 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3688 | visibility = xnnpack_visibility(), |
| 3689 | deps = [ |
| 3690 | "@FP16", |
| 3691 | "@clog", |
| 3692 | "@pthreadpool", |
| 3693 | ], |
| 3694 | ) |
| 3695 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3696 | xnnpack_aggregate_library( |
| 3697 | name = "ukernels", |
| 3698 | aarch32_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3699 | ":neon_ukernels", |
| 3700 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3701 | ":neonv8_ukernels", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3702 | ":neondot_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3703 | ":asm_ukernels", |
| 3704 | ], |
| 3705 | aarch64_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3706 | ":neon_ukernels", |
| 3707 | ":neonfma_ukernels", |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 3708 | ":neonv8_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3709 | ":neonfp16arith_ukernels", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3710 | ":neondot_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3711 | ":asm_ukernels", |
| 3712 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3713 | generic_deps = [ |
| 3714 | ":scalar_ukernels", |
| 3715 | ], |
Marat Dukhan | 500b889 | 2020-04-15 17:09:50 -0700 | [diff] [blame] | 3716 | psimd_deps = [ |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 3717 | ":psimd_fastmath_ukernels", |
| 3718 | ":psimd_accmath_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3719 | ], |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3720 | wasm_deps = [ |
| 3721 | ":wasm_ukernels", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3722 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3723 | ], |
| 3724 | wasmsimd_deps = [ |
| 3725 | ":wasm_ukernels", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3726 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3727 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3728 | x86_deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3729 | ":sse2_ukernels", |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 3730 | ":ssse3_ukernels", |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 3731 | ":sse41_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3732 | ":avx_ukernels", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3733 | ":xop_ukernels", |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 3734 | ":fma3_ukernels", |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 3735 | ":avx2_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3736 | ":avx512f_ukernels", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3737 | ":avx512skx_ukernels", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3738 | ], |
| 3739 | ) |
| 3740 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3741 | xnnpack_aggregate_library( |
| 3742 | name = "ukernels_test_mode", |
| 3743 | aarch32_deps = [ |
| 3744 | ":neon_ukernels_test_mode", |
| 3745 | ":neonfma_ukernels_test_mode", |
| 3746 | ":neonv8_ukernels_test_mode", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3747 | ":neondot_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3748 | ":asm_ukernels", |
| 3749 | ], |
| 3750 | aarch64_deps = [ |
| 3751 | ":neon_ukernels_test_mode", |
| 3752 | ":neonfma_ukernels_test_mode", |
| 3753 | ":neonv8_ukernels_test_mode", |
| 3754 | ":neonfp16arith_ukernels_test_mode", |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 3755 | ":neondot_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3756 | ":asm_ukernels", |
| 3757 | ], |
| 3758 | generic_deps = [ |
| 3759 | ":scalar_ukernels_test_mode", |
| 3760 | ], |
| 3761 | psimd_deps = [ |
| 3762 | ":psimd_fastmath_ukernels_test_mode", |
| 3763 | ":psimd_accmath_ukernels_test_mode", |
| 3764 | ], |
| 3765 | wasm_deps = [ |
| 3766 | ":wasm_ukernels_test_mode", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3767 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3768 | ], |
| 3769 | wasmsimd_deps = [ |
| 3770 | ":wasm_ukernels_test_mode", |
Frank Barchard | bcedc08 | 2020-08-17 18:00:51 -0700 | [diff] [blame] | 3771 | ":asm_ukernels", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3772 | ], |
| 3773 | x86_deps = [ |
| 3774 | ":sse2_ukernels_test_mode", |
| 3775 | ":ssse3_ukernels_test_mode", |
| 3776 | ":sse41_ukernels_test_mode", |
| 3777 | ":avx_ukernels_test_mode", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3778 | ":xop_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3779 | ":fma3_ukernels_test_mode", |
| 3780 | ":avx2_ukernels_test_mode", |
| 3781 | ":avx512f_ukernels_test_mode", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3782 | ":avx512skx_ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3783 | ], |
| 3784 | ) |
| 3785 | |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3786 | # This target must be compatible with Android NDK r17 and compile with Clang 6. |
| 3787 | # It is similar to :ukernels target, but excludes microkernels which require |
| 3788 | # NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics. |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3789 | xnnpack_aggregate_library( |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 3790 | name = "ukernels_max_compatibility", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3791 | aarch32_deps = [ |
| 3792 | ":neon_ukernels", |
| 3793 | ":neonfma_ukernels", |
| 3794 | ":neonv8_ukernels", |
| 3795 | ":asm_ukernels", |
| 3796 | ], |
| 3797 | aarch64_deps = [ |
| 3798 | ":neon_ukernels", |
| 3799 | ":neonfma_ukernels", |
| 3800 | ":neonv8_ukernels", |
| 3801 | ":asm_ukernels", |
| 3802 | ], |
| 3803 | generic_deps = [ |
| 3804 | ":scalar_ukernels", |
| 3805 | ], |
| 3806 | psimd_deps = [ |
| 3807 | ":psimd_fastmath_ukernels", |
| 3808 | ":psimd_accmath_ukernels", |
| 3809 | ], |
| 3810 | wasm_deps = [ |
| 3811 | ":wasm_ukernels", |
| 3812 | ], |
| 3813 | wasmsimd_deps = [ |
| 3814 | ":wasm_ukernels", |
| 3815 | ], |
| 3816 | x86_deps = [ |
| 3817 | ":sse2_ukernels", |
| 3818 | ":ssse3_ukernels", |
| 3819 | ":sse41_ukernels", |
| 3820 | ":avx_ukernels", |
Marat Dukhan | 1566fee | 2020-08-02 21:55:41 -0700 | [diff] [blame] | 3821 | ":xop_ukernels", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3822 | ":fma3_ukernels", |
| 3823 | ":avx2_ukernels", |
| 3824 | ":avx512f_ukernels", |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 3825 | ":avx512skx_ukernels", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 3826 | ], |
| 3827 | ) |
| 3828 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3829 | xnnpack_cc_library( |
| 3830 | name = "im2col", |
| 3831 | srcs = ["src/im2col.c"], |
| 3832 | hdrs = [ |
| 3833 | "src/xnnpack/common.h", |
| 3834 | "src/xnnpack/im2col.h", |
| 3835 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3836 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3837 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3838 | ) |
| 3839 | |
| 3840 | xnnpack_cc_library( |
| 3841 | name = "indirection", |
| 3842 | srcs = ["src/indirection.c"], |
| 3843 | hdrs = INTERNAL_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3844 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3845 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3846 | deps = [ |
| 3847 | "@FP16", |
| 3848 | "@FXdiv", |
| 3849 | "@pthreadpool", |
| 3850 | ], |
| 3851 | ) |
| 3852 | |
| 3853 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3854 | name = "indirection_test_mode", |
| 3855 | srcs = ["src/indirection.c"], |
| 3856 | hdrs = INTERNAL_HDRS, |
| 3857 | copts = [ |
| 3858 | "-UNDEBUG", |
| 3859 | "-DXNN_TEST_MODE=1", |
| 3860 | ], |
| 3861 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3862 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3863 | deps = [ |
| 3864 | "@FP16", |
| 3865 | "@FXdiv", |
| 3866 | "@pthreadpool", |
| 3867 | ], |
| 3868 | ) |
| 3869 | |
| 3870 | xnnpack_cc_library( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 3871 | name = "packing", |
| 3872 | srcs = ["src/packing.c"], |
| 3873 | hdrs = INTERNAL_HDRS, |
| 3874 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3875 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3876 | deps = [ |
| 3877 | "@FP16", |
| 3878 | "@FXdiv", |
| 3879 | "@pthreadpool", |
| 3880 | ], |
| 3881 | ) |
| 3882 | |
| 3883 | xnnpack_cc_library( |
| 3884 | name = "packing_test_mode", |
| 3885 | srcs = ["src/packing.c"], |
| 3886 | hdrs = INTERNAL_HDRS, |
| 3887 | copts = [ |
| 3888 | "-UNDEBUG", |
| 3889 | "-DXNN_TEST_MODE=1", |
| 3890 | ], |
| 3891 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3892 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3893 | deps = [ |
| 3894 | "@FP16", |
| 3895 | "@FXdiv", |
| 3896 | "@pthreadpool", |
| 3897 | ], |
| 3898 | ) |
| 3899 | |
| 3900 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3901 | name = "operator_run", |
| 3902 | srcs = ["src/operator-run.c"], |
Marat Dukhan | c8e00eb | 2019-10-04 14:55:26 -0700 | [diff] [blame] | 3903 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3904 | copts = LOGGING_COPTS + select({ |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 3905 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3906 | "//conditions:default": [], |
| 3907 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 3908 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3909 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3910 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3911 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3912 | "@FP16", |
| 3913 | "@FXdiv", |
| 3914 | "@clog", |
| 3915 | "@pthreadpool", |
| 3916 | ], |
| 3917 | ) |
| 3918 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3919 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3920 | name = "operator_run_test_mode", |
| 3921 | srcs = ["src/operator-run.c"], |
| 3922 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 3923 | copts = LOGGING_COPTS + [ |
| 3924 | "-UNDEBUG", |
| 3925 | "-DXNN_TEST_MODE=1", |
| 3926 | ] + select({ |
| 3927 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 3928 | "//conditions:default": [], |
| 3929 | }), |
| 3930 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3931 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3932 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3933 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3934 | "@FP16", |
| 3935 | "@FXdiv", |
| 3936 | "@clog", |
| 3937 | "@pthreadpool", |
| 3938 | ], |
| 3939 | ) |
| 3940 | |
| 3941 | xnnpack_cc_library( |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3942 | name = "memory_planner", |
| 3943 | srcs = ["src/memory-planner.c"], |
| 3944 | hdrs = INTERNAL_HDRS, |
| 3945 | defines = select({ |
| 3946 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 3947 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 3948 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 3949 | }), |
| 3950 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3951 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3952 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3953 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 3954 | "@pthreadpool", |
| 3955 | ], |
| 3956 | ) |
| 3957 | |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3958 | xnnpack_cc_library( |
| 3959 | name = "memory_planner_test_mode", |
| 3960 | srcs = ["src/memory-planner.c"], |
| 3961 | hdrs = INTERNAL_HDRS, |
| 3962 | copts = [ |
| 3963 | "-UNDEBUG", |
| 3964 | "-DXNN_TEST_MODE=1", |
| 3965 | ], |
| 3966 | defines = select({ |
| 3967 | ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"], |
| 3968 | ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"], |
| 3969 | "//conditions:default": ["XNN_ENABLE_MEMOPT=1"], |
| 3970 | }), |
| 3971 | gcc_copts = xnnpack_gcc_std_copts(), |
| 3972 | msvc_copts = xnnpack_msvc_std_copts(), |
| 3973 | deps = [ |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 3974 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 3975 | "@pthreadpool", |
| 3976 | ], |
| 3977 | ) |
| 3978 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3979 | cc_library( |
| 3980 | name = "enable_assembly", |
| 3981 | defines = select({ |
| 3982 | ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"], |
| 3983 | ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"], |
Frank Barchard | 810171d | 2019-10-10 10:34:51 -0700 | [diff] [blame] | 3984 | "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3985 | }), |
| 3986 | ) |
| 3987 | |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3988 | cc_library( |
| 3989 | name = "enable_sparse", |
| 3990 | defines = select({ |
| 3991 | ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"], |
| 3992 | ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"], |
Marat Dukhan | 8b283aa | 2020-06-18 16:48:57 -0700 | [diff] [blame] | 3993 | "//conditions:default": ["XNN_ENABLE_SPARSE=0"], |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 3994 | }), |
| 3995 | ) |
| 3996 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 3997 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 3998 | name = "operators", |
| 3999 | srcs = OPERATOR_SRCS + [ |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 4000 | "src/memory.c", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4001 | "src/operator-delete.c", |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4002 | ], |
| 4003 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4004 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4005 | "-Isrc", |
| 4006 | "-Iinclude", |
| 4007 | ] + select({ |
| 4008 | ":debug_build": [], |
| 4009 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4010 | }) + select({ |
| 4011 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4012 | "//conditions:default": [], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4013 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4014 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4015 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4016 | deps = [ |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4017 | ":indirection", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4018 | ":logging_utils", |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4019 | ":packing", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4020 | "@FP16", |
| 4021 | "@FXdiv", |
| 4022 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4023 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4024 | ], |
| 4025 | ) |
| 4026 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4027 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4028 | name = "operators_test_mode", |
| 4029 | srcs = OPERATOR_SRCS + [ |
| 4030 | "src/memory.c", |
| 4031 | "src/operator-delete.c", |
| 4032 | ], |
| 4033 | hdrs = INTERNAL_HDRS + LOGGING_HDRS, |
| 4034 | copts = LOGGING_COPTS + [ |
| 4035 | "-Isrc", |
| 4036 | "-Iinclude", |
| 4037 | "-UNDEBUG", |
| 4038 | "-DXNN_TEST_MODE=1", |
| 4039 | ] + select({ |
| 4040 | ":debug_build": [], |
| 4041 | "//conditions:default": xnnpack_min_size_copts(), |
| 4042 | }) + select({ |
| 4043 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4044 | "//conditions:default": [], |
| 4045 | }), |
| 4046 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4047 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4048 | deps = [ |
| 4049 | ":indirection_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4050 | ":logging_utils", |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4051 | ":packing_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4052 | "@FP16", |
| 4053 | "@FXdiv", |
| 4054 | "@clog", |
| 4055 | "@pthreadpool", |
| 4056 | ], |
| 4057 | ) |
| 4058 | |
| 4059 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4060 | name = "XNNPACK", |
| 4061 | srcs = [ |
| 4062 | "src/init.c", |
Marat Dukhan | ccfdbd1 | 2020-02-03 14:27:45 -0800 | [diff] [blame] | 4063 | "src/runtime.c", |
| 4064 | "src/subgraph.c", |
| 4065 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 4066 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4067 | hdrs = ["include/xnnpack.h"], |
| 4068 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4069 | "-Isrc", |
| 4070 | "-Iinclude", |
| 4071 | ] + select({ |
| 4072 | ":debug_build": [], |
| 4073 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4074 | }) + select({ |
| 4075 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4076 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4077 | }), |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4078 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4079 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4080 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4081 | visibility = xnnpack_visibility(), |
| 4082 | deps = [ |
| 4083 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4084 | ":enable_sparse", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4085 | ":logging_utils", |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 4086 | ":memory_planner", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4087 | ":operator_run", |
| 4088 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4089 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4090 | "@clog", |
Marat Dukhan | ab2946c | 2020-05-21 20:04:13 -0700 | [diff] [blame] | 4091 | "@FP16", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4092 | "@pthreadpool", |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 4093 | ] + select({ |
| 4094 | ":emscripten": [], |
| 4095 | "//conditions:default": ["@cpuinfo"], |
| 4096 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4097 | ) |
| 4098 | |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4099 | xnnpack_cc_library( |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4100 | name = "XNNPACK_test_mode", |
| 4101 | srcs = [ |
| 4102 | "src/init.c", |
| 4103 | "src/runtime.c", |
| 4104 | "src/subgraph.c", |
| 4105 | "src/tensor.c", |
Marat Dukhan | f03da0d | 2020-06-10 16:00:20 -0700 | [diff] [blame] | 4106 | ] + SUBGRAPH_SRCS, |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4107 | hdrs = ["include/xnnpack.h"], |
| 4108 | copts = LOGGING_COPTS + [ |
| 4109 | "-Isrc", |
| 4110 | "-Iinclude", |
| 4111 | "-UNDEBUG", |
| 4112 | "-DXNN_TEST_MODE=1", |
| 4113 | ] + select({ |
| 4114 | ":debug_build": [], |
| 4115 | "//conditions:default": xnnpack_min_size_copts(), |
| 4116 | }) + select({ |
| 4117 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4118 | "//conditions:default": [], |
| 4119 | }), |
| 4120 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4121 | includes = ["include"], |
| 4122 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4123 | visibility = xnnpack_visibility(), |
| 4124 | deps = [ |
| 4125 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4126 | ":enable_sparse", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4127 | ":logging_utils", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4128 | ":memory_planner_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4129 | ":operator_run_test_mode", |
| 4130 | ":operators_test_mode", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4131 | ":ukernels_test_mode", |
Marat Dukhan | 33fcf78 | 2020-05-24 14:27:15 -0700 | [diff] [blame] | 4132 | "@clog", |
| 4133 | "@FP16", |
| 4134 | "@pthreadpool", |
| 4135 | ] + select({ |
| 4136 | ":emscripten": [], |
| 4137 | "//conditions:default": ["@cpuinfo"], |
| 4138 | }), |
| 4139 | ) |
| 4140 | |
| 4141 | xnnpack_cc_library( |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4142 | name = "xnnpack_f32", |
| 4143 | srcs = [ |
| 4144 | "src/init.c", |
| 4145 | "src/runtime.c", |
| 4146 | "src/subgraph.c", |
| 4147 | "src/tensor.c", |
| 4148 | ] + SUBGRAPH_SRCS, |
| 4149 | hdrs = ["include/xnnpack.h"], |
| 4150 | copts = LOGGING_COPTS + [ |
| 4151 | "-Isrc", |
| 4152 | "-Iinclude", |
| 4153 | ] + select({ |
| 4154 | ":debug_build": [], |
| 4155 | "//conditions:default": xnnpack_min_size_copts(), |
| 4156 | }) + select({ |
| 4157 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4158 | "//conditions:default": [], |
| 4159 | }), |
| 4160 | defines = [ |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 4161 | "XNN_NO_QS8_OPERATORS", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4162 | "XNN_NO_QU8_OPERATORS", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4163 | "XNN_NO_U8_OPERATORS", |
| 4164 | "XNN_NO_X8_OPERATORS", |
| 4165 | "XNN_NO_F16_OPERATORS", |
| 4166 | "XNN_NO_X16_OPERATORS", |
| 4167 | ], |
| 4168 | gcc_copts = xnnpack_gcc_std_copts(), |
| 4169 | includes = ["include"], |
| 4170 | msvc_copts = xnnpack_msvc_std_copts(), |
| 4171 | visibility = xnnpack_visibility(), |
| 4172 | deps = [ |
| 4173 | ":enable_assembly", |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 4174 | ":enable_sparse", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4175 | ":logging_utils", |
| 4176 | ":memory_planner", |
| 4177 | ":operator_run", |
| 4178 | ":operators", |
Marat Dukhan | 86df384 | 2020-06-17 14:03:31 -0700 | [diff] [blame] | 4179 | ":ukernels_max_compatibility", |
Marat Dukhan | ae046f5 | 2020-06-15 13:16:14 -0700 | [diff] [blame] | 4180 | "@clog", |
| 4181 | "@FP16", |
| 4182 | "@pthreadpool", |
| 4183 | ] + select({ |
| 4184 | ":emscripten": [], |
| 4185 | "//conditions:default": ["@cpuinfo"], |
| 4186 | }), |
| 4187 | ) |
| 4188 | |
| 4189 | xnnpack_cc_library( |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4190 | name = "xnnpack_operators_nhwc_f32", |
| 4191 | srcs = [ |
| 4192 | "src/init.c", |
| 4193 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4194 | hdrs = ["include/xnnpack.h"], |
| 4195 | copts = LOGGING_COPTS + [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4196 | "-Isrc", |
| 4197 | "-Iinclude", |
| 4198 | ] + select({ |
| 4199 | ":debug_build": [], |
| 4200 | "//conditions:default": xnnpack_min_size_copts(), |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 4201 | }) + select({ |
| 4202 | ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"], |
| 4203 | "//conditions:default": [], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4204 | }), |
| 4205 | defines = [ |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 4206 | "XNN_NO_QS8_OPERATORS", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4207 | "XNN_NO_QU8_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4208 | "XNN_NO_U8_OPERATORS", |
| 4209 | "XNN_NO_X8_OPERATORS", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 4210 | "XNN_NO_NCHW_OPERATORS", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4211 | ], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4212 | gcc_copts = xnnpack_gcc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4213 | includes = ["include"], |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4214 | msvc_copts = xnnpack_msvc_std_copts(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4215 | visibility = xnnpack_visibility(), |
| 4216 | deps = [ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4217 | ":enable_assembly", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4218 | ":logging_utils", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4219 | ":operator_run", |
| 4220 | ":operators", |
Marat Dukhan | 3b59de2 | 2020-06-03 20:15:19 -0700 | [diff] [blame] | 4221 | ":ukernels", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4222 | "@clog", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4223 | "@pthreadpool", |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 4224 | ] + select({ |
| 4225 | ":emscripten": [], |
| 4226 | "//conditions:default": ["@cpuinfo"], |
| 4227 | }), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4228 | ) |
| 4229 | |
Marat Dukhan | cf056b2 | 2019-10-07 10:26:29 -0700 | [diff] [blame] | 4230 | xnnpack_cc_library( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4231 | name = "bench_utils", |
| 4232 | srcs = ["bench/utils.cc"], |
| 4233 | hdrs = ["bench/utils.h"], |
Marat Dukhan | bad48fe | 2019-11-04 10:35:22 -0800 | [diff] [blame] | 4234 | deps = [ |
| 4235 | "@com_google_benchmark//:benchmark", |
| 4236 | "@cpuinfo", |
| 4237 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4238 | ) |
| 4239 | |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 4240 | ######################### Benchmarks for micro-kernels ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4241 | |
| 4242 | xnnpack_benchmark( |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 4243 | name = "qs8_gemm_bench", |
| 4244 | srcs = [ |
| 4245 | "bench/gemm.h", |
| 4246 | "bench/qs8-gemm.cc", |
| 4247 | "src/xnnpack/AlignedAllocator.h", |
| 4248 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
| 4249 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"], |
| 4250 | ) |
| 4251 | |
| 4252 | xnnpack_benchmark( |
Marat Dukhan | 56bdd4a | 2020-08-03 19:47:04 -0700 | [diff] [blame] | 4253 | name = "qs8_requantization_bench", |
| 4254 | srcs = [ |
| 4255 | "bench/qs8-requantization.cc", |
| 4256 | "src/xnnpack/requantization-stubs.h", |
| 4257 | "src/xnnpack/AlignedAllocator.h", |
| 4258 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4259 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4260 | ) |
| 4261 | |
| 4262 | xnnpack_benchmark( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4263 | name = "qu8_gemm_bench", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4264 | srcs = [ |
| 4265 | "bench/gemm.h", |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 4266 | "bench/qu8-gemm.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4267 | "src/xnnpack/AlignedAllocator.h", |
| 4268 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4269 | copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(), |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4270 | 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] | 4271 | ) |
| 4272 | |
| 4273 | xnnpack_benchmark( |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 4274 | name = "qu8_requantization_bench", |
| 4275 | srcs = [ |
| 4276 | "bench/qu8-requantization.cc", |
| 4277 | "src/xnnpack/requantization-stubs.h", |
| 4278 | "src/xnnpack/AlignedAllocator.h", |
| 4279 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4280 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4281 | ) |
| 4282 | |
| 4283 | xnnpack_benchmark( |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 4284 | name = "f16_igemm_bench", |
| 4285 | srcs = [ |
| 4286 | "bench/f16-igemm.cc", |
| 4287 | "bench/conv.h", |
| 4288 | "bench/google/conv.h", |
| 4289 | "src/xnnpack/AlignedAllocator.h", |
| 4290 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4291 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4292 | ":indirection", |
| 4293 | ":packing", |
| 4294 | ], |
Frank Barchard | 40d20fe | 2020-05-05 00:37:45 -0700 | [diff] [blame] | 4295 | ) |
| 4296 | |
| 4297 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4298 | name = "f16_gemm_bench", |
| 4299 | srcs = [ |
| 4300 | "bench/f16-gemm.cc", |
| 4301 | "bench/gemm.h", |
| 4302 | "src/xnnpack/AlignedAllocator.h", |
| 4303 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4304 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4305 | ":packing", |
| 4306 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4307 | ) |
| 4308 | |
| 4309 | xnnpack_benchmark( |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4310 | name = "f16_spmm_bench", |
| 4311 | srcs = [ |
| 4312 | "bench/f16-spmm.cc", |
| 4313 | "bench/gemm.h", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4314 | "src/xnnpack/AlignedAllocator.h", |
| 4315 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4316 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4317 | ) |
| 4318 | |
| 4319 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4320 | name = "f32_igemm_bench", |
| 4321 | srcs = [ |
| 4322 | "bench/f32-igemm.cc", |
| 4323 | "bench/conv.h", |
| 4324 | "src/xnnpack/AlignedAllocator.h", |
| 4325 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4326 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4327 | ":indirection", |
| 4328 | ":packing", |
| 4329 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4330 | ) |
| 4331 | |
| 4332 | xnnpack_benchmark( |
Frank Barchard | c201006 | 2020-07-19 10:04:12 -0700 | [diff] [blame] | 4333 | name = "f16_relu_bench", |
| 4334 | srcs = [ |
| 4335 | "bench/f16-relu.cc", |
| 4336 | "src/xnnpack/AlignedAllocator.h", |
| 4337 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4338 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4339 | ) |
| 4340 | |
| 4341 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4342 | name = "f32_conv_hwc_bench", |
| 4343 | srcs = [ |
| 4344 | "bench/f32-conv-hwc.cc", |
| 4345 | "bench/dconv.h", |
| 4346 | "src/xnnpack/AlignedAllocator.h", |
| 4347 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4348 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4349 | ":packing", |
| 4350 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4351 | ) |
| 4352 | |
| 4353 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4354 | name = "f32_conv_hwc2chw_bench", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4355 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4356 | "bench/f32-conv-hwc2chw.cc", |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4357 | "bench/dconv.h", |
| 4358 | "src/xnnpack/AlignedAllocator.h", |
| 4359 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4360 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4361 | ":packing", |
| 4362 | ], |
Erich Elsen | 563df5f | 2019-10-23 08:02:21 -0700 | [diff] [blame] | 4363 | ) |
| 4364 | |
| 4365 | xnnpack_benchmark( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4366 | name = "f16_dwconv_bench", |
| 4367 | srcs = [ |
| 4368 | "bench/f16-dwconv.cc", |
| 4369 | "bench/dwconv.h", |
| 4370 | "bench/google/dwconv.h", |
| 4371 | "src/xnnpack/AlignedAllocator.h", |
| 4372 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4373 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4374 | ":indirection", |
| 4375 | ":packing", |
| 4376 | ], |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 4377 | ) |
| 4378 | |
| 4379 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4380 | name = "f32_dwconv_bench", |
| 4381 | srcs = [ |
| 4382 | "bench/f32-dwconv.cc", |
| 4383 | "bench/dwconv.h", |
| 4384 | "src/xnnpack/AlignedAllocator.h", |
| 4385 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4386 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4387 | ":indirection", |
| 4388 | ":packing", |
| 4389 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4390 | ) |
| 4391 | |
| 4392 | xnnpack_benchmark( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4393 | name = "f32_dwconv_chw_bench", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4394 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 4395 | "bench/f32-dwconv-chw.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4396 | "bench/dwconv.h", |
| 4397 | "src/xnnpack/AlignedAllocator.h", |
| 4398 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4399 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4400 | ":indirection", |
| 4401 | ":packing", |
| 4402 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4403 | ) |
| 4404 | |
| 4405 | xnnpack_benchmark( |
| 4406 | name = "f32_gemm_bench", |
| 4407 | srcs = [ |
| 4408 | "bench/f32-gemm.cc", |
| 4409 | "bench/gemm.h", |
| 4410 | "src/xnnpack/AlignedAllocator.h", |
| 4411 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4412 | copts = xnnpack_optional_ruy_copts(), |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4413 | deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4414 | ) |
| 4415 | |
| 4416 | xnnpack_benchmark( |
Marat Dukhan | a11ca34 | 2020-06-25 23:45:07 -0700 | [diff] [blame] | 4417 | name = "f32_hswish_bench", |
| 4418 | srcs = [ |
| 4419 | "bench/f32-hswish.cc", |
| 4420 | "src/xnnpack/AlignedAllocator.h", |
| 4421 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4422 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4423 | ) |
| 4424 | |
| 4425 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 4426 | name = "f32_raddexpminusmax_bench", |
| 4427 | srcs = [ |
| 4428 | "bench/f32-raddexpminusmax.cc", |
| 4429 | "src/xnnpack/AlignedAllocator.h", |
| 4430 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4431 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4432 | ) |
| 4433 | |
| 4434 | xnnpack_benchmark( |
| 4435 | name = "f32_raddextexp_bench", |
| 4436 | srcs = [ |
| 4437 | "bench/f32-raddextexp.cc", |
| 4438 | "src/xnnpack/AlignedAllocator.h", |
| 4439 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4440 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4441 | ) |
| 4442 | |
| 4443 | xnnpack_benchmark( |
| 4444 | name = "f32_raddstoreexpminusmax_bench", |
| 4445 | srcs = [ |
| 4446 | "bench/f32-raddstoreexpminusmax.cc", |
| 4447 | "src/xnnpack/AlignedAllocator.h", |
| 4448 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4449 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4450 | ) |
| 4451 | |
| 4452 | xnnpack_benchmark( |
Frank Barchard | 3c4a952 | 2020-07-18 16:11:22 -0700 | [diff] [blame] | 4453 | name = "f32_relu_bench", |
| 4454 | srcs = [ |
| 4455 | "bench/f32-relu.cc", |
| 4456 | "src/xnnpack/AlignedAllocator.h", |
| 4457 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4458 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4459 | ) |
| 4460 | |
| 4461 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4462 | name = "f32_rmax_bench", |
| 4463 | srcs = [ |
| 4464 | "bench/f32-rmax.cc", |
| 4465 | "src/xnnpack/AlignedAllocator.h", |
| 4466 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4467 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4468 | ) |
| 4469 | |
| 4470 | xnnpack_benchmark( |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 4471 | name = "f32_sigmoid_bench", |
| 4472 | srcs = [ |
| 4473 | "bench/f32-sigmoid.cc", |
| 4474 | "src/xnnpack/AlignedAllocator.h", |
| 4475 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 4476 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4477 | ) |
| 4478 | |
| 4479 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4480 | name = "f32_spmm_bench", |
| 4481 | srcs = [ |
| 4482 | "bench/f32-spmm.cc", |
| 4483 | "bench/gemm.h", |
| 4484 | "src/xnnpack/AlignedAllocator.h", |
| 4485 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4486 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4487 | ) |
| 4488 | |
| 4489 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4490 | name = "f32_softmax_bench", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4491 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4492 | "bench/f32-softmax.cc", |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4493 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 4494 | copts = xnnpack_optional_dnnl_copts(), |
Marat Dukhan | 8d3c693 | 2020-03-06 20:27:27 -0800 | [diff] [blame] | 4495 | deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(), |
Marat Dukhan | 4a4a7fa | 2019-10-21 13:46:14 -0700 | [diff] [blame] | 4496 | ) |
| 4497 | |
| 4498 | xnnpack_benchmark( |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 4499 | name = "f32_vscaleexpminusmax_bench", |
| 4500 | srcs = [ |
| 4501 | "bench/f32-vscaleexpminusmax.cc", |
| 4502 | "src/xnnpack/AlignedAllocator.h", |
| 4503 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4504 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4505 | ) |
| 4506 | |
| 4507 | xnnpack_benchmark( |
| 4508 | name = "f32_vscaleextexp_bench", |
| 4509 | srcs = [ |
| 4510 | "bench/f32-vscaleextexp.cc", |
| 4511 | "src/xnnpack/AlignedAllocator.h", |
| 4512 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4513 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4514 | ) |
| 4515 | |
| 4516 | xnnpack_benchmark( |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 4517 | name = "f32_vsqrt_bench", |
| 4518 | srcs = [ |
| 4519 | "bench/f32-vsqrt.cc", |
| 4520 | "src/xnnpack/AlignedAllocator.h", |
| 4521 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4522 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4523 | ) |
| 4524 | |
| 4525 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4526 | name = "f32_im2col_gemm_bench", |
| 4527 | srcs = [ |
| 4528 | "bench/f32-im2col-gemm.cc", |
| 4529 | "bench/conv.h", |
| 4530 | "src/xnnpack/AlignedAllocator.h", |
| 4531 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4532 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4533 | ":im2col", |
| 4534 | ":packing", |
| 4535 | ], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4536 | ) |
| 4537 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 4538 | xnnpack_benchmark( |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 4539 | name = "rounding_bench", |
| 4540 | srcs = [ |
| 4541 | "bench/rounding.cc", |
| 4542 | "src/xnnpack/math-stubs.h", |
| 4543 | "src/xnnpack/AlignedAllocator.h", |
| 4544 | ] + MICROKERNEL_BENCHMARK_HDRS, |
| 4545 | deps = MICROKERNEL_BENCHMARK_DEPS, |
| 4546 | ) |
| 4547 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4548 | ########################### Benchmarks for operators ########################### |
| 4549 | |
| 4550 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4551 | name = "average_pooling_bench", |
| 4552 | srcs = ["bench/average-pooling.cc"], |
Marat Dukhan | 7a16d8b | 2020-03-11 04:22:44 -0700 | [diff] [blame] | 4553 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4554 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4555 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4556 | ) |
| 4557 | |
| 4558 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4559 | name = "bankers_rounding_bench", |
| 4560 | srcs = ["bench/bankers-rounding.cc"], |
| 4561 | copts = xnnpack_optional_tflite_copts(), |
| 4562 | tags = ["nowin32"], |
| 4563 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4564 | ) |
| 4565 | |
| 4566 | xnnpack_benchmark( |
| 4567 | name = "ceiling_bench", |
| 4568 | srcs = ["bench/ceiling.cc"], |
| 4569 | copts = xnnpack_optional_tflite_copts(), |
| 4570 | tags = ["nowin32"], |
| 4571 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4572 | ) |
| 4573 | |
| 4574 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4575 | name = "channel_shuffle_bench", |
| 4576 | srcs = ["bench/channel-shuffle.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4577 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4578 | ) |
| 4579 | |
| 4580 | xnnpack_benchmark( |
| 4581 | name = "convolution_bench", |
| 4582 | srcs = ["bench/convolution.cc"], |
| 4583 | copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4584 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4585 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4586 | ) |
| 4587 | |
| 4588 | xnnpack_benchmark( |
| 4589 | name = "deconvolution_bench", |
| 4590 | srcs = ["bench/deconvolution.cc"], |
| 4591 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4592 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4593 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4594 | ) |
| 4595 | |
| 4596 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4597 | name = "floor_bench", |
| 4598 | srcs = ["bench/floor.cc"], |
| 4599 | copts = xnnpack_optional_tflite_copts(), |
| 4600 | tags = ["nowin32"], |
| 4601 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4602 | ) |
| 4603 | |
| 4604 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4605 | name = "global_average_pooling_bench", |
| 4606 | srcs = ["bench/global-average-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4607 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4608 | ) |
| 4609 | |
| 4610 | xnnpack_benchmark( |
Marat Dukhan | ad35260 | 2020-06-25 21:50:54 -0700 | [diff] [blame] | 4611 | name = "hardswish_bench", |
| 4612 | srcs = ["bench/hardswish.cc"], |
| 4613 | copts = xnnpack_optional_tflite_copts(), |
| 4614 | tags = ["nowin32"], |
| 4615 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4616 | ) |
| 4617 | |
| 4618 | xnnpack_benchmark( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4619 | name = "max_pooling_bench", |
| 4620 | srcs = ["bench/max-pooling.cc"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4621 | deps = OPERATOR_BENCHMARK_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4622 | ) |
| 4623 | |
| 4624 | xnnpack_benchmark( |
| 4625 | name = "sigmoid_bench", |
| 4626 | srcs = ["bench/sigmoid.cc"], |
Marat Dukhan | c3b9e86 | 2019-11-17 13:18:54 -0800 | [diff] [blame] | 4627 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 4628 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4629 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4630 | ) |
| 4631 | |
| 4632 | xnnpack_benchmark( |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 4633 | name = "prelu_bench", |
| 4634 | srcs = ["bench/prelu.cc"], |
| 4635 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | 8ea0b07 | 2020-04-23 16:12:18 -0700 | [diff] [blame] | 4636 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4637 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 95b2243 | 2019-10-30 16:30:14 -0700 | [diff] [blame] | 4638 | ) |
| 4639 | |
| 4640 | xnnpack_benchmark( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 4641 | name = "softmax_bench", |
| 4642 | srcs = ["bench/softmax.cc"], |
Marat Dukhan | 9c0db96 | 2020-01-28 12:30:14 -0800 | [diff] [blame] | 4643 | copts = xnnpack_optional_tflite_copts(), |
Marat Dukhan | ca2ba70 | 2020-04-24 01:31:47 -0700 | [diff] [blame] | 4644 | tags = ["nowin32"], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 4645 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4646 | ) |
| 4647 | |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4648 | xnnpack_benchmark( |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 4649 | name = "square_root_bench", |
| 4650 | srcs = ["bench/square-root.cc"], |
| 4651 | copts = xnnpack_optional_tflite_copts(), |
| 4652 | tags = ["nowin32"], |
| 4653 | deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(), |
| 4654 | ) |
| 4655 | |
| 4656 | xnnpack_benchmark( |
Marat Dukhan | 8772714 | 2020-06-24 15:24:10 -0700 | [diff] [blame] | 4657 | name = "truncation_bench", |
| 4658 | srcs = ["bench/truncation.cc"], |
| 4659 | deps = OPERATOR_BENCHMARK_DEPS, |
| 4660 | ) |
| 4661 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4662 | ############################# End-to-end benchmarks ############################ |
| 4663 | |
| 4664 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4665 | name = "fp32_mobilenet_v1", |
| 4666 | srcs = ["models/fp32-mobilenet-v1.cc"], |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4667 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4668 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4669 | linkstatic = True, |
| 4670 | deps = [ |
| 4671 | ":XNNPACK", |
| 4672 | "@pthreadpool", |
| 4673 | ], |
| 4674 | ) |
| 4675 | |
| 4676 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4677 | name = "fp16_mobilenet_v1", |
| 4678 | srcs = ["models/fp16-mobilenet-v1.cc"], |
| 4679 | hdrs = ["models/models.h"], |
| 4680 | copts = xnnpack_std_cxxopts(), |
| 4681 | linkstatic = True, |
| 4682 | deps = [ |
| 4683 | ":XNNPACK", |
| 4684 | "@FP16", |
| 4685 | "@pthreadpool", |
| 4686 | ], |
| 4687 | ) |
| 4688 | |
| 4689 | cc_library( |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 4690 | name = "qs8_mobilenet_v1", |
| 4691 | srcs = ["models/qs8-mobilenet-v1.cc"], |
| 4692 | hdrs = ["models/models.h"], |
| 4693 | copts = xnnpack_std_cxxopts(), |
| 4694 | linkstatic = True, |
| 4695 | deps = [ |
| 4696 | ":XNNPACK", |
| 4697 | "@pthreadpool", |
| 4698 | ], |
| 4699 | ) |
| 4700 | |
| 4701 | cc_library( |
Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 4702 | name = "qs8_mobilenet_v2", |
| 4703 | srcs = ["models/qs8-mobilenet-v2.cc"], |
| 4704 | hdrs = ["models/models.h"], |
| 4705 | copts = xnnpack_std_cxxopts(), |
| 4706 | linkstatic = True, |
| 4707 | deps = [ |
| 4708 | ":XNNPACK", |
| 4709 | "@pthreadpool", |
| 4710 | ], |
| 4711 | ) |
| 4712 | |
| 4713 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4714 | name = "fp32_mobilenet_v2", |
| 4715 | srcs = ["models/fp32-mobilenet-v2.cc"], |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4716 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4717 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4718 | linkstatic = True, |
| 4719 | deps = [ |
| 4720 | ":XNNPACK", |
| 4721 | "@pthreadpool", |
| 4722 | ], |
| 4723 | ) |
| 4724 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4725 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4726 | name = "fp16_mobilenet_v2", |
| 4727 | srcs = ["models/fp16-mobilenet-v2.cc"], |
| 4728 | hdrs = ["models/models.h"], |
| 4729 | copts = xnnpack_std_cxxopts(), |
| 4730 | linkstatic = True, |
| 4731 | deps = [ |
| 4732 | ":XNNPACK", |
| 4733 | "@FP16", |
| 4734 | "@pthreadpool", |
| 4735 | ], |
| 4736 | ) |
| 4737 | |
| 4738 | cc_library( |
| 4739 | name = "fp32_mobilenet_v3_large", |
| 4740 | srcs = ["models/fp32-mobilenet-v3-large.cc"], |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4741 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4742 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4743 | linkstatic = True, |
| 4744 | deps = [ |
| 4745 | ":XNNPACK", |
| 4746 | "@pthreadpool", |
| 4747 | ], |
| 4748 | ) |
| 4749 | |
| 4750 | cc_library( |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4751 | name = "fp16_mobilenet_v3_large", |
| 4752 | srcs = ["models/fp16-mobilenet-v3-large.cc"], |
| 4753 | hdrs = ["models/models.h"], |
| 4754 | copts = xnnpack_std_cxxopts(), |
| 4755 | linkstatic = True, |
| 4756 | deps = [ |
| 4757 | ":XNNPACK", |
| 4758 | "@FP16", |
| 4759 | "@pthreadpool", |
| 4760 | ], |
| 4761 | ) |
| 4762 | |
| 4763 | cc_library( |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4764 | name = "fp32_mobilenet_v3_small", |
| 4765 | srcs = ["models/fp32-mobilenet-v3-small.cc"], |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4766 | hdrs = ["models/models.h"], |
Marat Dukhan | a84e40b | 2019-12-11 15:38:03 -0800 | [diff] [blame] | 4767 | copts = xnnpack_std_cxxopts(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4768 | linkstatic = True, |
| 4769 | deps = [ |
| 4770 | ":XNNPACK", |
| 4771 | "@pthreadpool", |
| 4772 | ], |
| 4773 | ) |
| 4774 | |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4775 | cc_library( |
| 4776 | name = "fp16_mobilenet_v3_small", |
| 4777 | srcs = ["models/fp16-mobilenet-v3-small.cc"], |
| 4778 | hdrs = ["models/models.h"], |
| 4779 | copts = xnnpack_std_cxxopts(), |
| 4780 | linkstatic = True, |
| 4781 | deps = [ |
| 4782 | ":XNNPACK", |
| 4783 | "@FP16", |
| 4784 | "@pthreadpool", |
| 4785 | ], |
| 4786 | ) |
| 4787 | |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4788 | xnnpack_benchmark( |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 4789 | name = "f32_dwconv_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4790 | srcs = [ |
| 4791 | "bench/f32-dwconv-e2e.cc", |
| 4792 | "bench/end2end.h", |
| 4793 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 4794 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4795 | ":XNNPACK", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4796 | ":fp32_mobilenet_v1", |
| 4797 | ":fp32_mobilenet_v2", |
| 4798 | ":fp32_mobilenet_v3_large", |
| 4799 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | ef4416e | 2019-10-31 13:44:40 -0700 | [diff] [blame] | 4800 | ], |
| 4801 | ) |
| 4802 | |
| 4803 | xnnpack_benchmark( |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4804 | name = "f32_gemm_e2e_bench", |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 4805 | srcs = [ |
| 4806 | "bench/f32-gemm-e2e.cc", |
| 4807 | "bench/end2end.h", |
| 4808 | ] + MICROKERNEL_BENCHMARK_HDRS, |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4809 | deps = MICROKERNEL_BENCHMARK_DEPS + [ |
| 4810 | ":XNNPACK", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4811 | ":fp32_mobilenet_v1", |
| 4812 | ":fp32_mobilenet_v2", |
| 4813 | ":fp32_mobilenet_v3_large", |
| 4814 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | 5f18d26 | 2019-10-31 10:24:14 -0700 | [diff] [blame] | 4815 | ], |
| 4816 | ) |
| 4817 | |
| 4818 | xnnpack_benchmark( |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4819 | name = "end2end_bench", |
| 4820 | srcs = ["bench/end2end.cc"], |
| 4821 | deps = [ |
| 4822 | ":XNNPACK", |
Frank Barchard | c712fa4 | 2019-10-31 14:00:21 -0700 | [diff] [blame] | 4823 | ":bench_utils", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4824 | ":fp16_mobilenet_v1", |
| 4825 | ":fp16_mobilenet_v2", |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 4826 | ":fp16_mobilenet_v3_large", |
| 4827 | ":fp16_mobilenet_v3_small", |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 4828 | ":fp32_mobilenet_v1", |
| 4829 | ":fp32_mobilenet_v2", |
| 4830 | ":fp32_mobilenet_v3_large", |
| 4831 | ":fp32_mobilenet_v3_small", |
Marat Dukhan | 0743cdf | 2020-08-04 18:52:07 -0700 | [diff] [blame] | 4832 | ":qs8_mobilenet_v1", |
Marat Dukhan | 70a9618 | 2020-09-03 17:13:58 -0700 | [diff] [blame] | 4833 | ":qs8_mobilenet_v2", |
Marat Dukhan | c068bb6 | 2019-10-04 13:24:39 -0700 | [diff] [blame] | 4834 | "@pthreadpool", |
| 4835 | ], |
| 4836 | ) |
| 4837 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 4838 | #################### Accuracy evaluation for math functions #################### |
| 4839 | |
| 4840 | xnnpack_benchmark( |
| 4841 | name = "f32_exp_eval", |
| 4842 | srcs = [ |
| 4843 | "eval/f32-exp.cc", |
| 4844 | "src/xnnpack/AlignedAllocator.h", |
| 4845 | ] + ACCURACY_EVAL_HDRS, |
| 4846 | deps = ACCURACY_EVAL_DEPS, |
| 4847 | ) |
| 4848 | |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 4849 | xnnpack_benchmark( |
| 4850 | name = "f32_expminus_eval", |
| 4851 | srcs = [ |
| 4852 | "eval/f32-expminus.cc", |
| 4853 | "src/xnnpack/AlignedAllocator.h", |
| 4854 | ] + ACCURACY_EVAL_HDRS, |
| 4855 | deps = ACCURACY_EVAL_DEPS, |
| 4856 | ) |
| 4857 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 4858 | xnnpack_benchmark( |
| 4859 | name = "f32_extexp_eval", |
| 4860 | srcs = [ |
| 4861 | "eval/f32-extexp.cc", |
| 4862 | "src/xnnpack/AlignedAllocator.h", |
| 4863 | ] + ACCURACY_EVAL_HDRS, |
| 4864 | deps = ACCURACY_EVAL_DEPS, |
| 4865 | ) |
| 4866 | |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 4867 | xnnpack_unit_test( |
| 4868 | name = "f32_roundne_eval", |
| 4869 | srcs = [ |
| 4870 | "eval/f32-roundne.cc", |
| 4871 | "src/xnnpack/AlignedAllocator.h", |
| 4872 | "src/xnnpack/math-stubs.h", |
| 4873 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | 22eed3d | 2020-05-11 20:13:37 -0700 | [diff] [blame] | 4874 | automatic = False, |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 4875 | deps = MICROKERNEL_TEST_DEPS, |
| 4876 | ) |
| 4877 | |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 4878 | xnnpack_unit_test( |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 4879 | name = "f32_roundd_eval", |
| 4880 | srcs = [ |
| 4881 | "eval/f32-roundd.cc", |
| 4882 | "src/xnnpack/AlignedAllocator.h", |
| 4883 | "src/xnnpack/math-stubs.h", |
| 4884 | ] + MICROKERNEL_TEST_HDRS, |
| 4885 | automatic = False, |
| 4886 | deps = MICROKERNEL_TEST_DEPS, |
| 4887 | ) |
| 4888 | |
| 4889 | xnnpack_unit_test( |
| 4890 | name = "f32_roundu_eval", |
| 4891 | srcs = [ |
| 4892 | "eval/f32-roundu.cc", |
| 4893 | "src/xnnpack/AlignedAllocator.h", |
| 4894 | "src/xnnpack/math-stubs.h", |
| 4895 | ] + MICROKERNEL_TEST_HDRS, |
| 4896 | automatic = False, |
| 4897 | deps = MICROKERNEL_TEST_DEPS, |
| 4898 | ) |
| 4899 | |
| 4900 | xnnpack_unit_test( |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 4901 | name = "f32_roundz_eval", |
| 4902 | srcs = [ |
| 4903 | "eval/f32-roundz.cc", |
| 4904 | "src/xnnpack/AlignedAllocator.h", |
| 4905 | "src/xnnpack/math-stubs.h", |
| 4906 | ] + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 4907 | automatic = False, |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 4908 | deps = MICROKERNEL_TEST_DEPS, |
| 4909 | ) |
| 4910 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 4911 | xnnpack_benchmark( |
| 4912 | name = "f32_sigmoid_eval", |
| 4913 | srcs = [ |
| 4914 | "eval/f32-sigmoid.cc", |
| 4915 | "src/xnnpack/AlignedAllocator.h", |
| 4916 | ] + ACCURACY_EVAL_HDRS, |
| 4917 | deps = ACCURACY_EVAL_DEPS, |
| 4918 | ) |
| 4919 | |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 4920 | xnnpack_benchmark( |
| 4921 | name = "f32_sqrt_eval", |
| 4922 | srcs = [ |
| 4923 | "eval/f32-sqrt.cc", |
| 4924 | "src/xnnpack/AlignedAllocator.h", |
| 4925 | ] + ACCURACY_EVAL_HDRS, |
| 4926 | deps = ACCURACY_EVAL_DEPS + [ |
| 4927 | ":bench_utils", |
| 4928 | ], |
| 4929 | ) |
| 4930 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4931 | ######################### Unit tests for micro-kernels ######################### |
| 4932 | |
| 4933 | xnnpack_unit_test( |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4934 | name = "f16_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4935 | srcs = [ |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 4936 | "test/f16-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4937 | "test/gemm-microkernel-tester.h", |
| 4938 | "src/xnnpack/AlignedAllocator.h", |
| 4939 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 4940 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4941 | ) |
| 4942 | |
| 4943 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4944 | name = "f16_spmm_minmax_test", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4945 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 4946 | "test/f16-spmm-minmax.cc", |
Marat Dukhan | bdb56f5 | 2020-02-05 21:42:49 -0800 | [diff] [blame] | 4947 | "test/spmm-microkernel-tester.h", |
| 4948 | "src/xnnpack/AlignedAllocator.h", |
| 4949 | ] + MICROKERNEL_TEST_HDRS, |
| 4950 | deps = MICROKERNEL_TEST_DEPS, |
| 4951 | ) |
| 4952 | |
| 4953 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4954 | name = "f32_argmaxpool_test", |
| 4955 | srcs = [ |
| 4956 | "test/f32-argmaxpool.cc", |
| 4957 | "test/argmaxpool-microkernel-tester.h", |
| 4958 | "src/xnnpack/AlignedAllocator.h", |
| 4959 | ] + MICROKERNEL_TEST_HDRS, |
| 4960 | deps = MICROKERNEL_TEST_DEPS, |
| 4961 | ) |
| 4962 | |
| 4963 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4964 | name = "f32_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4965 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 4966 | "test/f32-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4967 | "test/avgpool-microkernel-tester.h", |
| 4968 | "src/xnnpack/AlignedAllocator.h", |
| 4969 | ] + MICROKERNEL_TEST_HDRS, |
| 4970 | deps = MICROKERNEL_TEST_DEPS, |
| 4971 | ) |
| 4972 | |
| 4973 | xnnpack_unit_test( |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 4974 | name = "f32_ibilinear_test", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 4975 | srcs = [ |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 4976 | "test/f32-ibilinear.cc", |
| 4977 | "test/ibilinear-microkernel-tester.h", |
Marat Dukhan | 35dacfb | 2019-11-07 19:18:16 -0800 | [diff] [blame] | 4978 | "src/xnnpack/AlignedAllocator.h", |
| 4979 | ] + MICROKERNEL_TEST_HDRS, |
| 4980 | deps = MICROKERNEL_TEST_DEPS, |
| 4981 | ) |
| 4982 | |
| 4983 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 4984 | name = "f16_clamp_test", |
| 4985 | srcs = [ |
| 4986 | "test/f16-clamp.cc", |
| 4987 | "test/clamp-microkernel-tester.h", |
| 4988 | ] + MICROKERNEL_TEST_HDRS, |
| 4989 | deps = MICROKERNEL_TEST_DEPS, |
| 4990 | ) |
| 4991 | |
| 4992 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 4993 | name = "f32_clamp_test", |
| 4994 | srcs = [ |
| 4995 | "test/f32-clamp.cc", |
| 4996 | "test/clamp-microkernel-tester.h", |
| 4997 | ] + MICROKERNEL_TEST_HDRS, |
| 4998 | deps = MICROKERNEL_TEST_DEPS, |
| 4999 | ) |
| 5000 | |
| 5001 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5002 | name = "f32_igemm_test", |
| 5003 | srcs = [ |
| 5004 | "test/f32-igemm.cc", |
| 5005 | "test/gemm-microkernel-tester.h", |
| 5006 | "src/xnnpack/AlignedAllocator.h", |
| 5007 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5008 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5009 | ) |
| 5010 | |
| 5011 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5012 | name = "f32_igemm_relu_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5013 | srcs = [ |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5014 | "test/f32-igemm-relu.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5015 | "test/gemm-microkernel-tester.h", |
| 5016 | "src/xnnpack/AlignedAllocator.h", |
| 5017 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5018 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5019 | ) |
| 5020 | |
| 5021 | xnnpack_unit_test( |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 5022 | name = "f32_igemm_minmax_test", |
| 5023 | srcs = [ |
| 5024 | "test/f32-igemm-minmax.cc", |
| 5025 | "test/gemm-microkernel-tester.h", |
| 5026 | "src/xnnpack/AlignedAllocator.h", |
| 5027 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5028 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | e207b7b | 2020-05-28 16:27:42 -0700 | [diff] [blame] | 5029 | ) |
| 5030 | |
| 5031 | xnnpack_unit_test( |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 5032 | name = "f16_igemm_minmax_test", |
| 5033 | srcs = [ |
| 5034 | "test/f16-igemm-minmax.cc", |
| 5035 | "test/gemm-microkernel-tester.h", |
| 5036 | "src/xnnpack/AlignedAllocator.h", |
| 5037 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5038 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -0700 | [diff] [blame] | 5039 | ) |
| 5040 | |
| 5041 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5042 | name = "f32_conv_hwc_test", |
| 5043 | srcs = [ |
| 5044 | "test/f32-conv-hwc.cc", |
| 5045 | "test/conv-hwc-microkernel-tester.h", |
| 5046 | "src/xnnpack/AlignedAllocator.h", |
| 5047 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5048 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5049 | ) |
| 5050 | |
| 5051 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5052 | name = "f32_conv_hwc2chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5053 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5054 | "test/f32-conv-hwc2chw.cc", |
| 5055 | "test/conv-hwc2chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5056 | "src/xnnpack/AlignedAllocator.h", |
| 5057 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5058 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5059 | ) |
| 5060 | |
| 5061 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5062 | name = "f32_dwconv_test", |
| 5063 | srcs = [ |
| 5064 | "test/f32-dwconv.cc", |
| 5065 | "test/dwconv-microkernel-tester.h", |
| 5066 | "src/xnnpack/AlignedAllocator.h", |
| 5067 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5068 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5069 | ) |
| 5070 | |
| 5071 | xnnpack_unit_test( |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 5072 | name = "f16_dwconv_minmax_test", |
| 5073 | srcs = [ |
| 5074 | "test/f16-dwconv-minmax.cc", |
| 5075 | "test/dwconv-microkernel-tester.h", |
| 5076 | "src/xnnpack/AlignedAllocator.h", |
| 5077 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5078 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 5079 | ) |
| 5080 | |
| 5081 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5082 | name = "f32_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5083 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5084 | "test/f32-dwconv-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5085 | "test/dwconv-microkernel-tester.h", |
| 5086 | "src/xnnpack/AlignedAllocator.h", |
| 5087 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5088 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5089 | ) |
| 5090 | |
| 5091 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5092 | name = "f32_dwconv_chw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5093 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5094 | "test/f32-dwconv-chw.cc", |
| 5095 | "test/dwconv-chw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5096 | "src/xnnpack/AlignedAllocator.h", |
| 5097 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5098 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5099 | ) |
| 5100 | |
| 5101 | xnnpack_unit_test( |
Frank Barchard | 0bb49a7 | 2020-06-04 11:35:11 -0700 | [diff] [blame] | 5102 | name = "f16_gavgpool_minmax_test", |
| 5103 | srcs = [ |
| 5104 | "test/f16-gavgpool-minmax.cc", |
| 5105 | "test/gavgpool-microkernel-tester.h", |
| 5106 | "src/xnnpack/AlignedAllocator.h", |
| 5107 | ] + MICROKERNEL_TEST_HDRS, |
| 5108 | deps = MICROKERNEL_TEST_DEPS, |
| 5109 | ) |
| 5110 | |
| 5111 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5112 | name = "f32_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5113 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5114 | "test/f32-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5115 | "test/gavgpool-microkernel-tester.h", |
| 5116 | "src/xnnpack/AlignedAllocator.h", |
| 5117 | ] + MICROKERNEL_TEST_HDRS, |
| 5118 | deps = MICROKERNEL_TEST_DEPS, |
| 5119 | ) |
| 5120 | |
| 5121 | xnnpack_unit_test( |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5122 | name = "f32_gavgpool_cw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5123 | srcs = [ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 5124 | "test/f32-gavgpool-cw.cc", |
| 5125 | "test/gavgpool-cw-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5126 | "src/xnnpack/AlignedAllocator.h", |
| 5127 | ] + MICROKERNEL_TEST_HDRS, |
| 5128 | deps = MICROKERNEL_TEST_DEPS, |
| 5129 | ) |
| 5130 | |
| 5131 | xnnpack_unit_test( |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5132 | name = "f32_gemm_test", |
| 5133 | srcs = [ |
| 5134 | "test/f32-gemm.cc", |
| 5135 | "test/gemm-microkernel-tester.h", |
| 5136 | "src/xnnpack/AlignedAllocator.h", |
| 5137 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5138 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 5139 | ) |
| 5140 | |
| 5141 | xnnpack_unit_test( |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5142 | name = "f32_gemm_relu_test", |
| 5143 | srcs = [ |
| 5144 | "test/f32-gemm-relu.cc", |
| 5145 | "test/gemm-microkernel-tester.h", |
| 5146 | "src/xnnpack/AlignedAllocator.h", |
| 5147 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5148 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 5149 | ) |
| 5150 | |
| 5151 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5152 | name = "f32_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5153 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5154 | "test/f32-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5155 | "test/gemm-microkernel-tester.h", |
| 5156 | "src/xnnpack/AlignedAllocator.h", |
| 5157 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5158 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5159 | ) |
| 5160 | |
| 5161 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5162 | name = "f32_gemminc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5163 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5164 | "test/f32-gemminc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5165 | "test/gemm-microkernel-tester.h", |
| 5166 | "src/xnnpack/AlignedAllocator.h", |
| 5167 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5168 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5169 | ) |
| 5170 | |
| 5171 | xnnpack_unit_test( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5172 | name = "f16_hswish_test", |
| 5173 | srcs = [ |
| 5174 | "test/f16-hswish.cc", |
| 5175 | "test/hswish-microkernel-tester.h", |
| 5176 | ] + MICROKERNEL_TEST_HDRS, |
| 5177 | deps = MICROKERNEL_TEST_DEPS, |
| 5178 | ) |
| 5179 | |
| 5180 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5181 | name = "f32_hswish_test", |
| 5182 | srcs = [ |
| 5183 | "test/f32-hswish.cc", |
| 5184 | "test/hswish-microkernel-tester.h", |
| 5185 | ] + MICROKERNEL_TEST_HDRS, |
| 5186 | deps = MICROKERNEL_TEST_DEPS, |
| 5187 | ) |
| 5188 | |
| 5189 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5190 | name = "f32_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5191 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5192 | "test/f32-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5193 | "test/maxpool-microkernel-tester.h", |
| 5194 | ] + MICROKERNEL_TEST_HDRS, |
| 5195 | deps = MICROKERNEL_TEST_DEPS, |
| 5196 | ) |
| 5197 | |
| 5198 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5199 | name = "f32_pavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5200 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5201 | "test/f32-pavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5202 | "test/avgpool-microkernel-tester.h", |
| 5203 | "src/xnnpack/AlignedAllocator.h", |
| 5204 | ] + MICROKERNEL_TEST_HDRS, |
| 5205 | deps = MICROKERNEL_TEST_DEPS, |
| 5206 | ) |
| 5207 | |
| 5208 | xnnpack_unit_test( |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5209 | name = "f32_ppmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5210 | srcs = [ |
Marat Dukhan | 1c58711 | 2020-04-08 20:04:28 -0700 | [diff] [blame] | 5211 | "test/f32-ppmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5212 | "test/gemm-microkernel-tester.h", |
| 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( |
Frank Barchard | b196659 | 2020-05-12 13:47:06 -0700 | [diff] [blame] | 5219 | name = "f16_prelu_test", |
| 5220 | srcs = [ |
| 5221 | "test/f16-prelu.cc", |
| 5222 | "test/prelu-microkernel-tester.h", |
| 5223 | "src/xnnpack/AlignedAllocator.h", |
| 5224 | ] + MICROKERNEL_TEST_HDRS, |
| 5225 | deps = MICROKERNEL_TEST_DEPS, |
| 5226 | ) |
| 5227 | |
| 5228 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5229 | name = "f32_prelu_test", |
| 5230 | srcs = [ |
| 5231 | "test/f32-prelu.cc", |
| 5232 | "test/prelu-microkernel-tester.h", |
| 5233 | "src/xnnpack/AlignedAllocator.h", |
| 5234 | ] + MICROKERNEL_TEST_HDRS, |
| 5235 | deps = MICROKERNEL_TEST_DEPS, |
| 5236 | ) |
| 5237 | |
| 5238 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5239 | name = "f32_raddexpminusmax_test", |
| 5240 | srcs = [ |
| 5241 | "test/f32-raddexpminusmax.cc", |
| 5242 | "test/raddexpminusmax-microkernel-tester.h", |
| 5243 | ] + MICROKERNEL_TEST_HDRS, |
| 5244 | deps = MICROKERNEL_TEST_DEPS, |
| 5245 | ) |
| 5246 | |
| 5247 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 5248 | name = "f32_raddextexp_test", |
| 5249 | srcs = [ |
| 5250 | "test/f32-raddextexp.cc", |
| 5251 | "test/raddextexp-microkernel-tester.h", |
| 5252 | ] + MICROKERNEL_TEST_HDRS, |
| 5253 | deps = MICROKERNEL_TEST_DEPS, |
| 5254 | ) |
| 5255 | |
| 5256 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5257 | name = "f32_raddstoreexpminusmax_test", |
| 5258 | srcs = [ |
| 5259 | "test/f32-raddstoreexpminusmax.cc", |
| 5260 | "test/raddstoreexpminusmax-microkernel-tester.h", |
| 5261 | ] + MICROKERNEL_TEST_HDRS, |
| 5262 | deps = MICROKERNEL_TEST_DEPS, |
| 5263 | ) |
| 5264 | |
| 5265 | xnnpack_unit_test( |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 5266 | name = "f32_relu_test", |
| 5267 | srcs = [ |
| 5268 | "test/f32-relu.cc", |
| 5269 | "test/vunary-microkernel-tester.h", |
| 5270 | ] + MICROKERNEL_TEST_HDRS, |
| 5271 | deps = MICROKERNEL_TEST_DEPS, |
| 5272 | ) |
| 5273 | |
| 5274 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5275 | name = "f32_rmax_test", |
| 5276 | srcs = [ |
| 5277 | "test/f32-rmax.cc", |
| 5278 | "test/rmax-microkernel-tester.h", |
| 5279 | ] + MICROKERNEL_TEST_HDRS, |
| 5280 | deps = MICROKERNEL_TEST_DEPS, |
| 5281 | ) |
| 5282 | |
| 5283 | xnnpack_unit_test( |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5284 | name = "f32_sigmoid_test", |
| 5285 | srcs = [ |
| 5286 | "test/f32-sigmoid.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5287 | "test/vunary-microkernel-tester.h", |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 5288 | ] + MICROKERNEL_TEST_HDRS, |
| 5289 | deps = MICROKERNEL_TEST_DEPS, |
| 5290 | ) |
| 5291 | |
| 5292 | xnnpack_unit_test( |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5293 | name = "f32_spmm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5294 | srcs = [ |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 5295 | "test/f32-spmm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5296 | "test/spmm-microkernel-tester.h", |
| 5297 | "src/xnnpack/AlignedAllocator.h", |
| 5298 | ] + MICROKERNEL_TEST_HDRS, |
| 5299 | deps = MICROKERNEL_TEST_DEPS, |
| 5300 | ) |
| 5301 | |
| 5302 | xnnpack_unit_test( |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 5303 | name = "f16_vadd_minmax_test", |
| 5304 | srcs = [ |
| 5305 | "test/f16-vadd-minmax.cc", |
| 5306 | "test/vbinary-microkernel-tester.h", |
| 5307 | ] + MICROKERNEL_TEST_HDRS, |
| 5308 | deps = MICROKERNEL_TEST_DEPS, |
| 5309 | ) |
| 5310 | |
| 5311 | xnnpack_unit_test( |
| 5312 | name = "f16_vaddc_minmax_test", |
| 5313 | srcs = [ |
| 5314 | "test/f16-vaddc-minmax.cc", |
| 5315 | "test/vbinaryc-microkernel-tester.h", |
| 5316 | ] + MICROKERNEL_TEST_HDRS, |
| 5317 | deps = MICROKERNEL_TEST_DEPS, |
| 5318 | ) |
| 5319 | |
| 5320 | xnnpack_unit_test( |
| 5321 | name = "f16_vdiv_minmax_test", |
| 5322 | srcs = [ |
| 5323 | "test/f16-vdiv-minmax.cc", |
| 5324 | "test/vbinary-microkernel-tester.h", |
| 5325 | ] + MICROKERNEL_TEST_HDRS, |
| 5326 | deps = MICROKERNEL_TEST_DEPS, |
| 5327 | ) |
| 5328 | |
| 5329 | xnnpack_unit_test( |
| 5330 | name = "f16_vdivc_minmax_test", |
| 5331 | srcs = [ |
| 5332 | "test/f16-vdivc-minmax.cc", |
| 5333 | "test/vbinaryc-microkernel-tester.h", |
| 5334 | ] + MICROKERNEL_TEST_HDRS, |
| 5335 | deps = MICROKERNEL_TEST_DEPS, |
| 5336 | ) |
| 5337 | |
| 5338 | xnnpack_unit_test( |
| 5339 | name = "f16_vrdivc_minmax_test", |
| 5340 | srcs = [ |
| 5341 | "test/f16-vrdivc-minmax.cc", |
| 5342 | "test/vbinaryc-microkernel-tester.h", |
| 5343 | ] + MICROKERNEL_TEST_HDRS, |
| 5344 | deps = MICROKERNEL_TEST_DEPS, |
| 5345 | ) |
| 5346 | |
| 5347 | xnnpack_unit_test( |
| 5348 | name = "f16_vmax_test", |
| 5349 | srcs = [ |
| 5350 | "test/f16-vmax.cc", |
| 5351 | "test/vbinary-microkernel-tester.h", |
| 5352 | ] + MICROKERNEL_TEST_HDRS, |
| 5353 | deps = MICROKERNEL_TEST_DEPS, |
| 5354 | ) |
| 5355 | |
| 5356 | xnnpack_unit_test( |
| 5357 | name = "f16_vmaxc_test", |
| 5358 | srcs = [ |
| 5359 | "test/f16-vmaxc.cc", |
| 5360 | "test/vbinaryc-microkernel-tester.h", |
| 5361 | ] + MICROKERNEL_TEST_HDRS, |
| 5362 | deps = MICROKERNEL_TEST_DEPS, |
| 5363 | ) |
| 5364 | |
| 5365 | xnnpack_unit_test( |
| 5366 | name = "f16_vmin_test", |
| 5367 | srcs = [ |
| 5368 | "test/f16-vmin.cc", |
| 5369 | "test/vbinary-microkernel-tester.h", |
| 5370 | ] + MICROKERNEL_TEST_HDRS, |
| 5371 | deps = MICROKERNEL_TEST_DEPS, |
| 5372 | ) |
| 5373 | |
| 5374 | xnnpack_unit_test( |
| 5375 | name = "f16_vminc_test", |
| 5376 | srcs = [ |
| 5377 | "test/f16-vminc.cc", |
| 5378 | "test/vbinaryc-microkernel-tester.h", |
| 5379 | ] + MICROKERNEL_TEST_HDRS, |
| 5380 | deps = MICROKERNEL_TEST_DEPS, |
| 5381 | ) |
| 5382 | |
| 5383 | xnnpack_unit_test( |
| 5384 | name = "f16_vmul_minmax_test", |
| 5385 | srcs = [ |
| 5386 | "test/f16-vmul-minmax.cc", |
| 5387 | "test/vbinary-microkernel-tester.h", |
| 5388 | ] + MICROKERNEL_TEST_HDRS, |
| 5389 | deps = MICROKERNEL_TEST_DEPS, |
| 5390 | ) |
| 5391 | |
| 5392 | xnnpack_unit_test( |
| 5393 | name = "f16_vmulc_minmax_test", |
| 5394 | srcs = [ |
| 5395 | "test/f16-vmulc-minmax.cc", |
| 5396 | "test/vbinaryc-microkernel-tester.h", |
| 5397 | ] + MICROKERNEL_TEST_HDRS, |
| 5398 | deps = MICROKERNEL_TEST_DEPS, |
| 5399 | ) |
| 5400 | |
| 5401 | xnnpack_unit_test( |
| 5402 | name = "f16_vsub_minmax_test", |
| 5403 | srcs = [ |
| 5404 | "test/f16-vsub-minmax.cc", |
| 5405 | "test/vbinary-microkernel-tester.h", |
| 5406 | ] + MICROKERNEL_TEST_HDRS, |
| 5407 | deps = MICROKERNEL_TEST_DEPS, |
| 5408 | ) |
| 5409 | |
| 5410 | xnnpack_unit_test( |
| 5411 | name = "f16_vsubc_minmax_test", |
| 5412 | srcs = [ |
| 5413 | "test/f16-vsubc-minmax.cc", |
| 5414 | "test/vbinaryc-microkernel-tester.h", |
| 5415 | ] + MICROKERNEL_TEST_HDRS, |
| 5416 | deps = MICROKERNEL_TEST_DEPS, |
| 5417 | ) |
| 5418 | |
| 5419 | xnnpack_unit_test( |
| 5420 | name = "f16_vrsubc_minmax_test", |
| 5421 | srcs = [ |
| 5422 | "test/f16-vrsubc-minmax.cc", |
| 5423 | "test/vbinaryc-microkernel-tester.h", |
| 5424 | ] + MICROKERNEL_TEST_HDRS, |
| 5425 | deps = MICROKERNEL_TEST_DEPS, |
| 5426 | ) |
| 5427 | |
| 5428 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5429 | name = "f32_vabs_test", |
| 5430 | srcs = [ |
| 5431 | "test/f32-vabs.cc", |
| 5432 | "test/vunary-microkernel-tester.h", |
| 5433 | ] + MICROKERNEL_TEST_HDRS, |
| 5434 | deps = MICROKERNEL_TEST_DEPS, |
| 5435 | ) |
| 5436 | |
| 5437 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5438 | name = "f32_vadd_test", |
| 5439 | srcs = [ |
| 5440 | "test/f32-vadd.cc", |
| 5441 | "test/vbinary-microkernel-tester.h", |
| 5442 | ] + MICROKERNEL_TEST_HDRS, |
| 5443 | deps = MICROKERNEL_TEST_DEPS, |
| 5444 | ) |
| 5445 | |
| 5446 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5447 | name = "f32_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5448 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5449 | "test/f32-vadd-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5450 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5451 | ] + MICROKERNEL_TEST_HDRS, |
| 5452 | deps = MICROKERNEL_TEST_DEPS, |
| 5453 | ) |
| 5454 | |
| 5455 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5456 | name = "f32_vadd_relu_test", |
| 5457 | srcs = [ |
| 5458 | "test/f32-vadd-relu.cc", |
| 5459 | "test/vbinary-microkernel-tester.h", |
| 5460 | ] + MICROKERNEL_TEST_HDRS, |
| 5461 | deps = MICROKERNEL_TEST_DEPS, |
| 5462 | ) |
| 5463 | |
| 5464 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5465 | name = "f32_vaddc_test", |
| 5466 | srcs = [ |
| 5467 | "test/f32-vaddc.cc", |
| 5468 | "test/vbinaryc-microkernel-tester.h", |
| 5469 | ] + MICROKERNEL_TEST_HDRS, |
| 5470 | deps = MICROKERNEL_TEST_DEPS, |
| 5471 | ) |
| 5472 | |
| 5473 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5474 | name = "f32_vaddc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5475 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5476 | "test/f32-vaddc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5477 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5478 | ] + MICROKERNEL_TEST_HDRS, |
| 5479 | deps = MICROKERNEL_TEST_DEPS, |
| 5480 | ) |
| 5481 | |
| 5482 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5483 | name = "f32_vaddc_relu_test", |
| 5484 | srcs = [ |
| 5485 | "test/f32-vaddc-relu.cc", |
| 5486 | "test/vbinaryc-microkernel-tester.h", |
| 5487 | ] + MICROKERNEL_TEST_HDRS, |
| 5488 | deps = MICROKERNEL_TEST_DEPS, |
| 5489 | ) |
| 5490 | |
| 5491 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5492 | name = "f32_vdiv_test", |
| 5493 | srcs = [ |
| 5494 | "test/f32-vdiv.cc", |
| 5495 | "test/vbinary-microkernel-tester.h", |
| 5496 | ] + MICROKERNEL_TEST_HDRS, |
| 5497 | deps = MICROKERNEL_TEST_DEPS, |
| 5498 | ) |
| 5499 | |
| 5500 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5501 | name = "f32_vdiv_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5502 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5503 | "test/f32-vdiv-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5504 | "test/vbinary-microkernel-tester.h", |
| 5505 | ] + MICROKERNEL_TEST_HDRS, |
| 5506 | deps = MICROKERNEL_TEST_DEPS, |
| 5507 | ) |
| 5508 | |
| 5509 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5510 | name = "f32_vdiv_relu_test", |
| 5511 | srcs = [ |
| 5512 | "test/f32-vdiv-relu.cc", |
| 5513 | "test/vbinary-microkernel-tester.h", |
| 5514 | ] + MICROKERNEL_TEST_HDRS, |
| 5515 | deps = MICROKERNEL_TEST_DEPS, |
| 5516 | ) |
| 5517 | |
| 5518 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5519 | name = "f32_vdivc_test", |
| 5520 | srcs = [ |
| 5521 | "test/f32-vdivc.cc", |
| 5522 | "test/vbinaryc-microkernel-tester.h", |
| 5523 | ] + MICROKERNEL_TEST_HDRS, |
| 5524 | deps = MICROKERNEL_TEST_DEPS, |
| 5525 | ) |
| 5526 | |
| 5527 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5528 | name = "f32_vdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5529 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5530 | "test/f32-vdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5531 | "test/vbinaryc-microkernel-tester.h", |
| 5532 | ] + MICROKERNEL_TEST_HDRS, |
| 5533 | deps = MICROKERNEL_TEST_DEPS, |
| 5534 | ) |
| 5535 | |
| 5536 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5537 | name = "f32_vdivc_relu_test", |
| 5538 | srcs = [ |
| 5539 | "test/f32-vdivc-relu.cc", |
| 5540 | "test/vbinaryc-microkernel-tester.h", |
| 5541 | ] + MICROKERNEL_TEST_HDRS, |
| 5542 | deps = MICROKERNEL_TEST_DEPS, |
| 5543 | ) |
| 5544 | |
| 5545 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5546 | name = "f32_vrdivc_test", |
| 5547 | srcs = [ |
| 5548 | "test/f32-vrdivc.cc", |
| 5549 | "test/vbinaryc-microkernel-tester.h", |
| 5550 | ] + MICROKERNEL_TEST_HDRS, |
| 5551 | deps = MICROKERNEL_TEST_DEPS, |
| 5552 | ) |
| 5553 | |
| 5554 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5555 | name = "f32_vrdivc_minmax_test", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5556 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5557 | "test/f32-vrdivc-minmax.cc", |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 5558 | "test/vbinaryc-microkernel-tester.h", |
| 5559 | ] + MICROKERNEL_TEST_HDRS, |
| 5560 | deps = MICROKERNEL_TEST_DEPS, |
| 5561 | ) |
| 5562 | |
| 5563 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5564 | name = "f32_vrdivc_relu_test", |
| 5565 | srcs = [ |
| 5566 | "test/f32-vrdivc-relu.cc", |
| 5567 | "test/vbinaryc-microkernel-tester.h", |
| 5568 | ] + MICROKERNEL_TEST_HDRS, |
| 5569 | deps = MICROKERNEL_TEST_DEPS, |
| 5570 | ) |
| 5571 | |
| 5572 | xnnpack_unit_test( |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 5573 | name = "f32_vmax_test", |
| 5574 | srcs = [ |
| 5575 | "test/f32-vmax.cc", |
| 5576 | "test/vbinary-microkernel-tester.h", |
| 5577 | ] + MICROKERNEL_TEST_HDRS, |
| 5578 | deps = MICROKERNEL_TEST_DEPS, |
| 5579 | ) |
| 5580 | |
| 5581 | xnnpack_unit_test( |
| 5582 | name = "f32_vmaxc_test", |
| 5583 | srcs = [ |
| 5584 | "test/f32-vmaxc.cc", |
| 5585 | "test/vbinaryc-microkernel-tester.h", |
| 5586 | ] + MICROKERNEL_TEST_HDRS, |
| 5587 | deps = MICROKERNEL_TEST_DEPS, |
| 5588 | ) |
| 5589 | |
| 5590 | xnnpack_unit_test( |
| 5591 | name = "f32_vmin_test", |
| 5592 | srcs = [ |
| 5593 | "test/f32-vmin.cc", |
| 5594 | "test/vbinary-microkernel-tester.h", |
| 5595 | ] + MICROKERNEL_TEST_HDRS, |
| 5596 | deps = MICROKERNEL_TEST_DEPS, |
| 5597 | ) |
| 5598 | |
| 5599 | xnnpack_unit_test( |
| 5600 | name = "f32_vminc_test", |
| 5601 | srcs = [ |
| 5602 | "test/f32-vminc.cc", |
| 5603 | "test/vbinaryc-microkernel-tester.h", |
| 5604 | ] + MICROKERNEL_TEST_HDRS, |
| 5605 | deps = MICROKERNEL_TEST_DEPS, |
| 5606 | ) |
| 5607 | |
| 5608 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5609 | name = "f32_vmul_test", |
| 5610 | srcs = [ |
| 5611 | "test/f32-vmul.cc", |
| 5612 | "test/vbinary-microkernel-tester.h", |
| 5613 | ] + MICROKERNEL_TEST_HDRS, |
| 5614 | deps = MICROKERNEL_TEST_DEPS, |
| 5615 | ) |
| 5616 | |
| 5617 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5618 | name = "f32_vmul_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5619 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5620 | "test/f32-vmul-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5621 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5622 | ] + MICROKERNEL_TEST_HDRS, |
| 5623 | deps = MICROKERNEL_TEST_DEPS, |
| 5624 | ) |
| 5625 | |
| 5626 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5627 | name = "f32_vmul_relu_test", |
| 5628 | srcs = [ |
| 5629 | "test/f32-vmul-relu.cc", |
| 5630 | "test/vbinary-microkernel-tester.h", |
| 5631 | ] + MICROKERNEL_TEST_HDRS, |
| 5632 | deps = MICROKERNEL_TEST_DEPS, |
| 5633 | ) |
| 5634 | |
| 5635 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5636 | name = "f32_vmulc_test", |
| 5637 | srcs = [ |
| 5638 | "test/f32-vmulc.cc", |
| 5639 | "test/vbinaryc-microkernel-tester.h", |
| 5640 | ] + MICROKERNEL_TEST_HDRS, |
| 5641 | deps = MICROKERNEL_TEST_DEPS, |
| 5642 | ) |
| 5643 | |
| 5644 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5645 | name = "f32_vmulc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5646 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5647 | "test/f32-vmulc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5648 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5649 | ] + MICROKERNEL_TEST_HDRS, |
| 5650 | deps = MICROKERNEL_TEST_DEPS, |
| 5651 | ) |
| 5652 | |
| 5653 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5654 | name = "f32_vmulc_relu_test", |
| 5655 | srcs = [ |
| 5656 | "test/f32-vmulc-relu.cc", |
| 5657 | "test/vbinaryc-microkernel-tester.h", |
| 5658 | ] + MICROKERNEL_TEST_HDRS, |
| 5659 | deps = MICROKERNEL_TEST_DEPS, |
| 5660 | ) |
| 5661 | |
| 5662 | xnnpack_unit_test( |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 5663 | name = "f16_vmulcaddc_minmax_test", |
| 5664 | srcs = [ |
| 5665 | "test/f16-vmulcaddc-minmax.cc", |
| 5666 | "test/vmulcaddc-microkernel-tester.h", |
| 5667 | "src/xnnpack/AlignedAllocator.h", |
| 5668 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5669 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Frank Barchard | 2a1049c | 2020-06-03 02:31:27 -0700 | [diff] [blame] | 5670 | ) |
| 5671 | |
| 5672 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5673 | name = "f32_vmulcaddc_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5674 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 5675 | "test/f32-vmulcaddc-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5676 | "test/vmulcaddc-microkernel-tester.h", |
| 5677 | "src/xnnpack/AlignedAllocator.h", |
| 5678 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5679 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5680 | ) |
| 5681 | |
| 5682 | xnnpack_unit_test( |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 5683 | name = "f32_vlrelu_test", |
| 5684 | srcs = [ |
| 5685 | "test/f32-vlrelu.cc", |
| 5686 | "test/vunary-microkernel-tester.h", |
| 5687 | ] + MICROKERNEL_TEST_HDRS, |
| 5688 | deps = MICROKERNEL_TEST_DEPS, |
| 5689 | ) |
| 5690 | |
| 5691 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5692 | name = "f32_vneg_test", |
| 5693 | srcs = [ |
| 5694 | "test/f32-vneg.cc", |
| 5695 | "test/vunary-microkernel-tester.h", |
| 5696 | ] + MICROKERNEL_TEST_HDRS, |
| 5697 | deps = MICROKERNEL_TEST_DEPS, |
| 5698 | ) |
| 5699 | |
| 5700 | xnnpack_unit_test( |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 5701 | name = "f32_vrndne_test", |
| 5702 | srcs = [ |
| 5703 | "test/f32-vrndne.cc", |
| 5704 | "test/vunary-microkernel-tester.h", |
| 5705 | ] + MICROKERNEL_TEST_HDRS, |
| 5706 | deps = MICROKERNEL_TEST_DEPS, |
| 5707 | ) |
| 5708 | |
| 5709 | xnnpack_unit_test( |
| 5710 | name = "f32_vrndz_test", |
| 5711 | srcs = [ |
| 5712 | "test/f32-vrndz.cc", |
| 5713 | "test/vunary-microkernel-tester.h", |
| 5714 | ] + MICROKERNEL_TEST_HDRS, |
| 5715 | deps = MICROKERNEL_TEST_DEPS, |
| 5716 | ) |
| 5717 | |
| 5718 | xnnpack_unit_test( |
| 5719 | name = "f32_vrndu_test", |
| 5720 | srcs = [ |
| 5721 | "test/f32-vrndu.cc", |
| 5722 | "test/vunary-microkernel-tester.h", |
| 5723 | ] + MICROKERNEL_TEST_HDRS, |
| 5724 | deps = MICROKERNEL_TEST_DEPS, |
| 5725 | ) |
| 5726 | |
| 5727 | xnnpack_unit_test( |
| 5728 | name = "f32_vrndd_test", |
| 5729 | srcs = [ |
| 5730 | "test/f32-vrndd.cc", |
| 5731 | "test/vunary-microkernel-tester.h", |
| 5732 | ] + MICROKERNEL_TEST_HDRS, |
| 5733 | deps = MICROKERNEL_TEST_DEPS, |
| 5734 | ) |
| 5735 | |
| 5736 | xnnpack_unit_test( |
Marat Dukhan | 05ac8e3 | 2019-10-21 15:39:33 -0700 | [diff] [blame] | 5737 | name = "f32_vscale_test", |
| 5738 | srcs = [ |
| 5739 | "test/f32-vscale.cc", |
| 5740 | "test/vscale-microkernel-tester.h", |
| 5741 | ] + MICROKERNEL_TEST_HDRS, |
| 5742 | deps = MICROKERNEL_TEST_DEPS, |
| 5743 | ) |
| 5744 | |
| 5745 | xnnpack_unit_test( |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5746 | name = "f32_vscaleexpminusmax_test", |
| 5747 | srcs = [ |
| 5748 | "test/f32-vscaleexpminusmax.cc", |
| 5749 | "test/vscaleexpminusmax-microkernel-tester.h", |
| 5750 | ] + MICROKERNEL_TEST_HDRS, |
| 5751 | deps = MICROKERNEL_TEST_DEPS, |
| 5752 | ) |
| 5753 | |
| 5754 | xnnpack_unit_test( |
Marat Dukhan | 6f8d4d3 | 2019-10-25 17:07:09 -0700 | [diff] [blame] | 5755 | name = "f32_vscaleextexp_test", |
| 5756 | srcs = [ |
| 5757 | "test/f32-vscaleextexp.cc", |
| 5758 | "test/vscaleextexp-microkernel-tester.h", |
| 5759 | ] + MICROKERNEL_TEST_HDRS, |
| 5760 | deps = MICROKERNEL_TEST_DEPS, |
| 5761 | ) |
| 5762 | |
| 5763 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 5764 | name = "f32_vsqr_test", |
| 5765 | srcs = [ |
| 5766 | "test/f32-vsqr.cc", |
| 5767 | "test/vunary-microkernel-tester.h", |
| 5768 | ] + MICROKERNEL_TEST_HDRS, |
| 5769 | deps = MICROKERNEL_TEST_DEPS, |
| 5770 | ) |
| 5771 | |
| 5772 | xnnpack_unit_test( |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 5773 | name = "f32_vsqrdiff_test", |
| 5774 | srcs = [ |
| 5775 | "test/f32-vsqrdiff.cc", |
| 5776 | "test/vbinary-microkernel-tester.h", |
| 5777 | ] + MICROKERNEL_TEST_HDRS, |
| 5778 | deps = MICROKERNEL_TEST_DEPS, |
| 5779 | ) |
| 5780 | |
| 5781 | xnnpack_unit_test( |
| 5782 | name = "f32_vsqrdiffc_test", |
| 5783 | srcs = [ |
| 5784 | "test/f32-vsqrdiffc.cc", |
| 5785 | "test/vbinaryc-microkernel-tester.h", |
| 5786 | ] + MICROKERNEL_TEST_HDRS, |
| 5787 | deps = MICROKERNEL_TEST_DEPS, |
| 5788 | ) |
| 5789 | |
| 5790 | xnnpack_unit_test( |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 5791 | name = "f32_vsqrt_test", |
| 5792 | srcs = [ |
| 5793 | "test/f32-vsqrt.cc", |
| 5794 | "test/vunary-microkernel-tester.h", |
| 5795 | ] + MICROKERNEL_TEST_HDRS, |
| 5796 | deps = MICROKERNEL_TEST_DEPS, |
| 5797 | ) |
| 5798 | |
| 5799 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5800 | name = "f32_vsub_test", |
| 5801 | srcs = [ |
| 5802 | "test/f32-vsub.cc", |
| 5803 | "test/vbinary-microkernel-tester.h", |
| 5804 | ] + MICROKERNEL_TEST_HDRS, |
| 5805 | deps = MICROKERNEL_TEST_DEPS, |
| 5806 | ) |
| 5807 | |
| 5808 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5809 | name = "f32_vsub_minmax_test", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5810 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5811 | "test/f32-vsub-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5812 | "test/vbinary-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5813 | ] + MICROKERNEL_TEST_HDRS, |
| 5814 | deps = MICROKERNEL_TEST_DEPS, |
| 5815 | ) |
| 5816 | |
| 5817 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5818 | name = "f32_vsub_relu_test", |
| 5819 | srcs = [ |
| 5820 | "test/f32-vsub-relu.cc", |
| 5821 | "test/vbinary-microkernel-tester.h", |
| 5822 | ] + MICROKERNEL_TEST_HDRS, |
| 5823 | deps = MICROKERNEL_TEST_DEPS, |
| 5824 | ) |
| 5825 | |
| 5826 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5827 | name = "f32_vsubc_test", |
| 5828 | srcs = [ |
| 5829 | "test/f32-vsubc.cc", |
| 5830 | "test/vbinaryc-microkernel-tester.h", |
| 5831 | ] + MICROKERNEL_TEST_HDRS, |
| 5832 | deps = MICROKERNEL_TEST_DEPS, |
| 5833 | ) |
| 5834 | |
| 5835 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5836 | name = "f32_vsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5837 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5838 | "test/f32-vsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5839 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5840 | ] + MICROKERNEL_TEST_HDRS, |
| 5841 | deps = MICROKERNEL_TEST_DEPS, |
| 5842 | ) |
| 5843 | |
| 5844 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5845 | name = "f32_vsubc_relu_test", |
| 5846 | srcs = [ |
| 5847 | "test/f32-vsubc-relu.cc", |
| 5848 | "test/vbinaryc-microkernel-tester.h", |
| 5849 | ] + MICROKERNEL_TEST_HDRS, |
| 5850 | deps = MICROKERNEL_TEST_DEPS, |
| 5851 | ) |
| 5852 | |
| 5853 | xnnpack_unit_test( |
Frank Barchard | d5b9f1c | 2020-07-01 15:00:19 -0700 | [diff] [blame] | 5854 | name = "f32_vrsubc_test", |
| 5855 | srcs = [ |
| 5856 | "test/f32-vrsubc.cc", |
| 5857 | "test/vbinaryc-microkernel-tester.h", |
| 5858 | ] + MICROKERNEL_TEST_HDRS, |
| 5859 | deps = MICROKERNEL_TEST_DEPS, |
| 5860 | ) |
| 5861 | |
| 5862 | xnnpack_unit_test( |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5863 | name = "f32_vrsubc_minmax_test", |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 5864 | srcs = [ |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 5865 | "test/f32-vrsubc-minmax.cc", |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 5866 | "test/vbinaryc-microkernel-tester.h", |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 5867 | ] + MICROKERNEL_TEST_HDRS, |
| 5868 | deps = MICROKERNEL_TEST_DEPS, |
| 5869 | ) |
| 5870 | |
| 5871 | xnnpack_unit_test( |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 5872 | name = "f32_vrsubc_relu_test", |
| 5873 | srcs = [ |
| 5874 | "test/f32-vrsubc-relu.cc", |
| 5875 | "test/vbinaryc-microkernel-tester.h", |
| 5876 | ] + MICROKERNEL_TEST_HDRS, |
| 5877 | deps = MICROKERNEL_TEST_DEPS, |
| 5878 | ) |
| 5879 | |
| 5880 | xnnpack_unit_test( |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 5881 | name = "qs8_dwconv_minmax_test", |
| 5882 | srcs = [ |
| 5883 | "test/qs8-dwconv-minmax.cc", |
| 5884 | "test/dwconv-microkernel-tester.h", |
| 5885 | "src/xnnpack/AlignedAllocator.h", |
| 5886 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 5887 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 5888 | ) |
| 5889 | |
| 5890 | xnnpack_unit_test( |
Marat Dukhan | 4ed53f4 | 2020-08-06 01:12:55 -0700 | [diff] [blame] | 5891 | name = "qs8_gavgpool_minmax_test", |
| 5892 | srcs = [ |
| 5893 | "test/qs8-gavgpool-minmax.cc", |
| 5894 | "test/gavgpool-microkernel-tester.h", |
| 5895 | "src/xnnpack/AlignedAllocator.h", |
| 5896 | ] + MICROKERNEL_TEST_HDRS, |
| 5897 | deps = MICROKERNEL_TEST_DEPS, |
| 5898 | ) |
| 5899 | |
| 5900 | xnnpack_unit_test( |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 5901 | name = "qs8_gemm_minmax_test", |
| 5902 | srcs = [ |
| 5903 | "test/qs8-gemm-minmax.cc", |
| 5904 | "test/gemm-microkernel-tester.h", |
| 5905 | "src/xnnpack/AlignedAllocator.h", |
| 5906 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 5907 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 5908 | ) |
| 5909 | |
| 5910 | xnnpack_unit_test( |
Marat Dukhan | f948068 | 2020-07-31 14:50:24 -0700 | [diff] [blame] | 5911 | name = "qs8_igemm_minmax_test", |
| 5912 | srcs = [ |
| 5913 | "test/qs8-igemm-minmax.cc", |
| 5914 | "test/gemm-microkernel-tester.h", |
| 5915 | "src/xnnpack/AlignedAllocator.h", |
| 5916 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
| 5917 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
| 5918 | ) |
| 5919 | |
| 5920 | xnnpack_unit_test( |
| 5921 | name = "qs8_requantization_test", |
| 5922 | srcs = [ |
| 5923 | "src/xnnpack/requantization-stubs.h", |
| 5924 | "test/qs8-requantization.cc", |
| 5925 | "test/requantization-tester.h", |
| 5926 | ] + MICROKERNEL_TEST_HDRS, |
| 5927 | deps = MICROKERNEL_TEST_DEPS, |
| 5928 | ) |
| 5929 | |
| 5930 | xnnpack_unit_test( |
Marat Dukhan | d9f3ad4 | 2020-08-10 12:30:58 -0700 | [diff] [blame] | 5931 | name = "qs8_vadd_minmax_test", |
| 5932 | srcs = [ |
| 5933 | "test/qs8-vadd-minmax.cc", |
| 5934 | "test/vadd-microkernel-tester.h", |
| 5935 | ] + MICROKERNEL_TEST_HDRS, |
| 5936 | deps = MICROKERNEL_TEST_DEPS, |
| 5937 | ) |
| 5938 | |
| 5939 | xnnpack_unit_test( |
Marat Dukhan | 0270d9f | 2020-08-11 00:56:46 -0700 | [diff] [blame] | 5940 | name = "qs8_vaddc_minmax_test", |
| 5941 | srcs = [ |
| 5942 | "test/qs8-vaddc-minmax.cc", |
| 5943 | "test/vaddc-microkernel-tester.h", |
| 5944 | ] + MICROKERNEL_TEST_HDRS, |
| 5945 | deps = MICROKERNEL_TEST_DEPS, |
| 5946 | ) |
| 5947 | |
| 5948 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 5949 | name = "qu8_avgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5950 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 5951 | "test/qu8-avgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5952 | "test/avgpool-microkernel-tester.h", |
| 5953 | "src/xnnpack/AlignedAllocator.h", |
| 5954 | ] + MICROKERNEL_TEST_HDRS, |
| 5955 | deps = MICROKERNEL_TEST_DEPS, |
| 5956 | ) |
| 5957 | |
| 5958 | xnnpack_unit_test( |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 5959 | name = "qu8_dwconv_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5960 | srcs = [ |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 5961 | "test/qu8-dwconv-minmax.cc", |
| 5962 | "test/dwconv-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5963 | "src/xnnpack/AlignedAllocator.h", |
| 5964 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5965 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5966 | ) |
| 5967 | |
| 5968 | xnnpack_unit_test( |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 5969 | name = "qu8_igemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5970 | srcs = [ |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 5971 | "test/qu8-igemm-minmax.cc", |
| 5972 | "test/gemm-microkernel-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5973 | "src/xnnpack/AlignedAllocator.h", |
| 5974 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5975 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5976 | ) |
| 5977 | |
| 5978 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 5979 | name = "qu8_gavgpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5980 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 5981 | "test/qu8-gavgpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5982 | "test/gavgpool-microkernel-tester.h", |
| 5983 | "src/xnnpack/AlignedAllocator.h", |
| 5984 | ] + MICROKERNEL_TEST_HDRS, |
| 5985 | deps = MICROKERNEL_TEST_DEPS, |
| 5986 | ) |
| 5987 | |
| 5988 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 5989 | name = "qu8_gemm_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5990 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 5991 | "test/qu8-gemm-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5992 | "test/gemm-microkernel-tester.h", |
| 5993 | "src/xnnpack/AlignedAllocator.h", |
| 5994 | ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 5995 | deps = MICROKERNEL_TEST_DEPS + [":packing"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 5996 | ) |
| 5997 | |
| 5998 | xnnpack_unit_test( |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 5999 | name = "qu8_requantization_test", |
| 6000 | srcs = [ |
| 6001 | "src/xnnpack/requantization-stubs.h", |
| 6002 | "test/qu8-requantization.cc", |
| 6003 | "test/requantization-tester.h", |
| 6004 | ] + MICROKERNEL_TEST_HDRS, |
| 6005 | deps = MICROKERNEL_TEST_DEPS, |
| 6006 | ) |
| 6007 | |
| 6008 | xnnpack_unit_test( |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6009 | name = "qu8_vadd_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6010 | srcs = [ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 6011 | "test/qu8-vadd-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6012 | "test/vadd-microkernel-tester.h", |
| 6013 | ] + MICROKERNEL_TEST_HDRS, |
| 6014 | deps = MICROKERNEL_TEST_DEPS, |
| 6015 | ) |
| 6016 | |
| 6017 | xnnpack_unit_test( |
| 6018 | name = "u8_clamp_test", |
| 6019 | srcs = [ |
| 6020 | "test/u8-clamp.cc", |
| 6021 | "test/clamp-microkernel-tester.h", |
| 6022 | ] + MICROKERNEL_TEST_HDRS, |
| 6023 | deps = MICROKERNEL_TEST_DEPS, |
| 6024 | ) |
| 6025 | |
| 6026 | xnnpack_unit_test( |
| 6027 | name = "u8_lut32norm_test", |
| 6028 | srcs = [ |
| 6029 | "test/u8-lut32norm.cc", |
| 6030 | "test/lut-norm-microkernel-tester.h", |
| 6031 | ] + MICROKERNEL_TEST_HDRS, |
| 6032 | deps = MICROKERNEL_TEST_DEPS, |
| 6033 | ) |
| 6034 | |
| 6035 | xnnpack_unit_test( |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 6036 | name = "u8_maxpool_minmax_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6037 | srcs = [ |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 6038 | "test/u8-maxpool-minmax.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6039 | "test/maxpool-microkernel-tester.h", |
| 6040 | ] + MICROKERNEL_TEST_HDRS, |
| 6041 | deps = MICROKERNEL_TEST_DEPS, |
| 6042 | ) |
| 6043 | |
| 6044 | xnnpack_unit_test( |
| 6045 | name = "u8_rmax_test", |
| 6046 | srcs = [ |
| 6047 | "test/u8-rmax.cc", |
| 6048 | "test/rmax-microkernel-tester.h", |
| 6049 | ] + MICROKERNEL_TEST_HDRS, |
| 6050 | deps = MICROKERNEL_TEST_DEPS, |
| 6051 | ) |
| 6052 | |
| 6053 | xnnpack_unit_test( |
Marat Dukhan | 3bb3bfc | 2020-05-19 17:42:46 -0700 | [diff] [blame] | 6054 | name = "x32_fill_test", |
| 6055 | srcs = [ |
| 6056 | "test/x32-fill.cc", |
| 6057 | "test/fill-microkernel-tester.h", |
| 6058 | ] + MICROKERNEL_TEST_HDRS, |
| 6059 | deps = MICROKERNEL_TEST_DEPS, |
| 6060 | ) |
| 6061 | |
| 6062 | xnnpack_unit_test( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6063 | name = "x32_packx_test", |
| 6064 | srcs = [ |
| 6065 | "test/x32-packx.cc", |
| 6066 | "test/pack-microkernel-tester.h", |
| 6067 | "src/xnnpack/AlignedAllocator.h", |
| 6068 | ] + MICROKERNEL_TEST_HDRS, |
| 6069 | deps = MICROKERNEL_TEST_DEPS, |
| 6070 | ) |
| 6071 | |
| 6072 | xnnpack_unit_test( |
| 6073 | name = "x32_pad_test", |
| 6074 | srcs = [ |
| 6075 | "test/x32-pad.cc", |
| 6076 | "test/pad-microkernel-tester.h", |
| 6077 | ] + MICROKERNEL_TEST_HDRS, |
| 6078 | deps = MICROKERNEL_TEST_DEPS, |
| 6079 | ) |
| 6080 | |
| 6081 | xnnpack_unit_test( |
| 6082 | name = "x32_unpool_test", |
| 6083 | srcs = [ |
| 6084 | "test/x32-unpool.cc", |
| 6085 | "test/unpool-microkernel-tester.h", |
| 6086 | ] + MICROKERNEL_TEST_HDRS, |
| 6087 | deps = MICROKERNEL_TEST_DEPS, |
| 6088 | ) |
| 6089 | |
| 6090 | xnnpack_unit_test( |
| 6091 | name = "x32_zip_test", |
| 6092 | srcs = [ |
| 6093 | "test/x32-zip.cc", |
| 6094 | "test/zip-microkernel-tester.h", |
| 6095 | ] + MICROKERNEL_TEST_HDRS, |
| 6096 | deps = MICROKERNEL_TEST_DEPS, |
| 6097 | ) |
| 6098 | |
| 6099 | xnnpack_unit_test( |
| 6100 | name = "x8_lut_test", |
| 6101 | srcs = [ |
| 6102 | "test/x8-lut.cc", |
| 6103 | "test/lut-microkernel-tester.h", |
| 6104 | ] + MICROKERNEL_TEST_HDRS, |
| 6105 | deps = MICROKERNEL_TEST_DEPS, |
| 6106 | ) |
| 6107 | |
| 6108 | xnnpack_unit_test( |
| 6109 | name = "x8_zip_test", |
| 6110 | srcs = [ |
| 6111 | "test/x8-zip.cc", |
| 6112 | "test/zip-microkernel-tester.h", |
| 6113 | ] + MICROKERNEL_TEST_HDRS, |
| 6114 | deps = MICROKERNEL_TEST_DEPS, |
| 6115 | ) |
| 6116 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6117 | ########################## Size tests for the library ######################### |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6118 | |
| 6119 | xnnpack_binary( |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6120 | name = "operator_size_test", |
| 6121 | srcs = ["test/operator-size.c"], |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 6122 | deps = [":xnnpack_operators_nhwc_f32"], |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6123 | ) |
| 6124 | |
Marat Dukhan | 20c3b92 | 2020-03-10 03:45:06 -0700 | [diff] [blame] | 6125 | xnnpack_binary( |
| 6126 | name = "subgraph_size_test", |
| 6127 | srcs = ["test/subgraph-size.c"], |
| 6128 | deps = [":XNNPACK"], |
| 6129 | ) |
| 6130 | |
| 6131 | ########################### Unit tests for operators ########################## |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6132 | |
| 6133 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6134 | name = "abs_nc_test", |
| 6135 | srcs = [ |
| 6136 | "test/abs-nc.cc", |
| 6137 | "test/abs-operator-tester.h", |
| 6138 | ], |
| 6139 | deps = OPERATOR_TEST_DEPS, |
| 6140 | ) |
| 6141 | |
| 6142 | xnnpack_unit_test( |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6143 | name = "add_nd_test", |
| 6144 | srcs = [ |
| 6145 | "test/add-nd.cc", |
| 6146 | "test/binary-elementwise-operator-tester.h", |
| 6147 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6148 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6149 | ) |
| 6150 | |
| 6151 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6152 | name = "argmax_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6153 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6154 | "test/argmax-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6155 | "test/argmax-pooling-operator-tester.h", |
| 6156 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6157 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6158 | ) |
| 6159 | |
| 6160 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6161 | name = "average_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6162 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6163 | "test/average-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6164 | "test/average-pooling-operator-tester.h", |
| 6165 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6166 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6167 | ) |
| 6168 | |
| 6169 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6170 | name = "bankers_rounding_nc_test", |
| 6171 | srcs = [ |
| 6172 | "test/bankers-rounding-nc.cc", |
| 6173 | "test/bankers-rounding-operator-tester.h", |
| 6174 | ], |
| 6175 | deps = OPERATOR_TEST_DEPS, |
| 6176 | ) |
| 6177 | |
| 6178 | xnnpack_unit_test( |
| 6179 | name = "ceiling_nc_test", |
| 6180 | srcs = [ |
| 6181 | "test/ceiling-nc.cc", |
| 6182 | "test/ceiling-operator-tester.h", |
| 6183 | ], |
| 6184 | deps = OPERATOR_TEST_DEPS, |
| 6185 | ) |
| 6186 | |
| 6187 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6188 | name = "channel_shuffle_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6189 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6190 | "test/channel-shuffle-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6191 | "test/channel-shuffle-operator-tester.h", |
| 6192 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6193 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6194 | ) |
| 6195 | |
| 6196 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6197 | name = "clamp_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6198 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6199 | "test/clamp-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6200 | "test/clamp-operator-tester.h", |
| 6201 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6202 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6203 | ) |
| 6204 | |
| 6205 | xnnpack_unit_test( |
Marat Dukhan | 065b11e | 2020-05-22 09:49:41 -0700 | [diff] [blame] | 6206 | name = "constant_pad_nd_test", |
| 6207 | srcs = [ |
| 6208 | "test/constant-pad-nd.cc", |
| 6209 | "test/constant-pad-operator-tester.h", |
| 6210 | ], |
| 6211 | deps = OPERATOR_TEST_DEPS, |
| 6212 | ) |
| 6213 | |
| 6214 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6215 | name = "convolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6216 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6217 | "test/convolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6218 | "test/convolution-operator-tester.h", |
| 6219 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6220 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6221 | ) |
| 6222 | |
| 6223 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6224 | name = "convolution_nchw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6225 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6226 | "test/convolution-nchw.cc", |
| 6227 | "test/convolution-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6228 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6229 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6230 | ) |
| 6231 | |
| 6232 | xnnpack_unit_test( |
Marat Dukhan | 4e21b27 | 2020-06-04 18:45:01 -0700 | [diff] [blame] | 6233 | name = "copy_nc_test", |
| 6234 | srcs = [ |
| 6235 | "test/copy-nc.cc", |
| 6236 | "test/copy-operator-tester.h", |
| 6237 | ], |
| 6238 | deps = OPERATOR_TEST_DEPS, |
| 6239 | ) |
| 6240 | |
| 6241 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6242 | name = "deconvolution_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6243 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6244 | "test/deconvolution-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6245 | "test/deconvolution-operator-tester.h", |
| 6246 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6247 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6248 | ) |
| 6249 | |
| 6250 | xnnpack_unit_test( |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 6251 | name = "divide_nd_test", |
| 6252 | srcs = [ |
| 6253 | "test/binary-elementwise-operator-tester.h", |
| 6254 | "test/divide-nd.cc", |
| 6255 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6256 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 6257 | ) |
| 6258 | |
| 6259 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6260 | name = "fully_connected_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6261 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6262 | "test/fully-connected-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6263 | "test/fully-connected-operator-tester.h", |
| 6264 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6265 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6266 | ) |
| 6267 | |
| 6268 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6269 | name = "floor_nc_test", |
| 6270 | srcs = [ |
| 6271 | "test/floor-nc.cc", |
| 6272 | "test/floor-operator-tester.h", |
| 6273 | ], |
| 6274 | deps = OPERATOR_TEST_DEPS, |
| 6275 | ) |
| 6276 | |
| 6277 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6278 | name = "global_average_pooling_nwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6279 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6280 | "test/global-average-pooling-nwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6281 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | ef61d02 | 2020-06-19 13:54:49 -0700 | [diff] [blame] | 6282 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6283 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6284 | ) |
| 6285 | |
| 6286 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6287 | name = "global_average_pooling_ncw_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6288 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6289 | "test/global-average-pooling-ncw.cc", |
| 6290 | "test/global-average-pooling-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6291 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6292 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6293 | ) |
| 6294 | |
| 6295 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6296 | name = "hardswish_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6297 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6298 | "test/hardswish-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6299 | "test/hardswish-operator-tester.h", |
| 6300 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6301 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6302 | ) |
| 6303 | |
| 6304 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6305 | name = "leaky_relu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6306 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6307 | "test/leaky-relu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6308 | "test/leaky-relu-operator-tester.h", |
| 6309 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6310 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6311 | ) |
| 6312 | |
| 6313 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6314 | name = "max_pooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6315 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6316 | "test/max-pooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6317 | "test/max-pooling-operator-tester.h", |
| 6318 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6319 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6320 | ) |
| 6321 | |
| 6322 | xnnpack_unit_test( |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6323 | name = "maximum_nd_test", |
| 6324 | srcs = [ |
| 6325 | "test/binary-elementwise-operator-tester.h", |
| 6326 | "test/maximum-nd.cc", |
| 6327 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6328 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6329 | ) |
| 6330 | |
| 6331 | xnnpack_unit_test( |
| 6332 | name = "minimum_nd_test", |
| 6333 | srcs = [ |
| 6334 | "test/binary-elementwise-operator-tester.h", |
| 6335 | "test/minimum-nd.cc", |
| 6336 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6337 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 6338 | ) |
| 6339 | |
| 6340 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6341 | name = "multiply_nd_test", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6342 | srcs = [ |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 6343 | "test/binary-elementwise-operator-tester.h", |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6344 | "test/multiply-nd.cc", |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6345 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6346 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 6347 | ) |
| 6348 | |
| 6349 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6350 | name = "negate_nc_test", |
| 6351 | srcs = [ |
| 6352 | "test/negate-nc.cc", |
| 6353 | "test/negate-operator-tester.h", |
| 6354 | ], |
| 6355 | deps = OPERATOR_TEST_DEPS, |
| 6356 | ) |
| 6357 | |
| 6358 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6359 | name = "prelu_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6360 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6361 | "test/prelu-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6362 | "test/prelu-operator-tester.h", |
| 6363 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6364 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6365 | ) |
| 6366 | |
| 6367 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6368 | name = "resize_bilinear_nhwc_test", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6369 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6370 | "test/resize-bilinear-nhwc.cc", |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6371 | "test/resize-bilinear-operator-tester.h", |
| 6372 | ] + OPERATOR_TEST_PARAMS_HDRS, |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6373 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 6374 | ) |
| 6375 | |
| 6376 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6377 | name = "sigmoid_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6378 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6379 | "test/sigmoid-nc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6380 | "test/sigmoid-operator-tester.h", |
| 6381 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6382 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6383 | ) |
| 6384 | |
| 6385 | xnnpack_unit_test( |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 6386 | name = "softmax_nc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6387 | srcs = [ |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 6388 | "test/softmax-nc.cc", |
| 6389 | "test/softmax-operator-tester.h", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6390 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6391 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6392 | ) |
| 6393 | |
| 6394 | xnnpack_unit_test( |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 6395 | name = "square_nc_test", |
| 6396 | srcs = [ |
| 6397 | "test/square-nc.cc", |
| 6398 | "test/square-operator-tester.h", |
| 6399 | ], |
| 6400 | deps = OPERATOR_TEST_DEPS, |
| 6401 | ) |
| 6402 | |
| 6403 | xnnpack_unit_test( |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 6404 | name = "square_root_nc_test", |
| 6405 | srcs = [ |
| 6406 | "test/square-root-nc.cc", |
| 6407 | "test/square-root-operator-tester.h", |
| 6408 | ], |
| 6409 | deps = OPERATOR_TEST_DEPS, |
| 6410 | ) |
| 6411 | |
| 6412 | xnnpack_unit_test( |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 6413 | name = "squared_difference_nd_test", |
| 6414 | srcs = [ |
| 6415 | "test/binary-elementwise-operator-tester.h", |
| 6416 | "test/squared-difference-nd.cc", |
| 6417 | ], |
| 6418 | deps = OPERATOR_TEST_DEPS, |
| 6419 | ) |
| 6420 | |
| 6421 | xnnpack_unit_test( |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 6422 | name = "subtract_nd_test", |
| 6423 | srcs = [ |
| 6424 | "test/binary-elementwise-operator-tester.h", |
| 6425 | "test/subtract-nd.cc", |
| 6426 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6427 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 6428 | ) |
| 6429 | |
| 6430 | xnnpack_unit_test( |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 6431 | name = "truncation_nc_test", |
| 6432 | srcs = [ |
| 6433 | "test/truncation-nc.cc", |
| 6434 | "test/truncation-operator-tester.h", |
| 6435 | ], |
| 6436 | deps = OPERATOR_TEST_DEPS, |
| 6437 | ) |
| 6438 | |
| 6439 | xnnpack_unit_test( |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6440 | name = "unpooling_nhwc_test", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6441 | srcs = [ |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 6442 | "test/unpooling-nhwc.cc", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6443 | "test/unpooling-operator-tester.h", |
| 6444 | ], |
Marat Dukhan | 1b35463 | 2020-03-23 12:50:22 -0700 | [diff] [blame] | 6445 | deps = OPERATOR_TEST_DEPS, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6446 | ) |
| 6447 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 6448 | ############################### Misc unit tests ############################### |
| 6449 | |
| 6450 | xnnpack_unit_test( |
| 6451 | name = "memory_planner_test", |
| 6452 | srcs = [ |
| 6453 | "test/memory-planner-test.cc", |
| 6454 | ], |
| 6455 | deps = [ |
| 6456 | ":XNNPACK", |
| 6457 | ":memory_planner", |
| 6458 | ], |
| 6459 | ) |
| 6460 | |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6461 | ############################# Build configurations ############################# |
| 6462 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6463 | # Enables usage of assembly kernels. |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6464 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6465 | name = "xnn_enable_assembly_explicit_true", |
| 6466 | define_values = {"xnn_enable_assembly": "true"}, |
| 6467 | ) |
| 6468 | |
| 6469 | # Disables usage of assembly kernels. |
| 6470 | config_setting( |
| 6471 | name = "xnn_enable_assembly_explicit_false", |
| 6472 | define_values = {"xnn_enable_assembly": "false"}, |
| 6473 | ) |
| 6474 | |
Marat Dukhan | 9de90e0 | 2020-06-18 16:04:12 -0700 | [diff] [blame] | 6475 | # Enables usage of sparse inference. |
| 6476 | config_setting( |
| 6477 | name = "xnn_enable_sparse_explicit_true", |
| 6478 | define_values = {"xnn_enable_sparse": "true"}, |
| 6479 | ) |
| 6480 | |
| 6481 | # Disables usage of sparse inference. |
| 6482 | config_setting( |
| 6483 | name = "xnn_enable_sparse_explicit_false", |
| 6484 | define_values = {"xnn_enable_sparse": "false"}, |
| 6485 | ) |
| 6486 | |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 6487 | # Disables usage of HMP-aware optimizations. |
| 6488 | config_setting( |
| 6489 | name = "xnn_enable_hmp_explicit_false", |
| 6490 | define_values = {"xnn_enable_hmp": "false"}, |
| 6491 | ) |
| 6492 | |
Chao Mei | 6ddfc60 | 2020-05-13 22:29:36 -0700 | [diff] [blame] | 6493 | # Enable usage of optimized memory allocation |
| 6494 | config_setting( |
| 6495 | name = "xnn_enable_memopt_explicit_true", |
| 6496 | define_values = {"xnnpack_enable_memopt": "true"}, |
| 6497 | ) |
| 6498 | |
| 6499 | # Disable usage of optimized memory allocation |
| 6500 | config_setting( |
| 6501 | name = "xnn_enable_memopt_explicit_false", |
| 6502 | define_values = {"xnnpack_enable_memopt": "false"}, |
| 6503 | ) |
| 6504 | |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6505 | # Builds with -c dbg |
| 6506 | config_setting( |
| 6507 | name = "debug_build", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6508 | values = { |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6509 | "compilation_mode": "dbg", |
| 6510 | }, |
| 6511 | ) |
| 6512 | |
| 6513 | # Builds with -c opt |
| 6514 | config_setting( |
| 6515 | name = "optimized_build", |
| 6516 | values = { |
| 6517 | "compilation_mode": "opt", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6518 | }, |
| 6519 | ) |
| 6520 | |
| 6521 | config_setting( |
Marat Dukhan | b864235 | 2019-10-30 15:43:02 -0700 | [diff] [blame] | 6522 | name = "linux_k8", |
| 6523 | values = {"cpu": "k8"}, |
| 6524 | ) |
| 6525 | |
| 6526 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 6527 | name = "linux_arm", |
| 6528 | values = {"cpu": "arm"}, |
Marat Dukhan | 4e45e66 | 2019-10-03 15:40:24 -0700 | [diff] [blame] | 6529 | ) |
| 6530 | |
| 6531 | config_setting( |
Marat Dukhan | f0bd4de | 2020-06-15 15:53:19 -0700 | [diff] [blame] | 6532 | name = "linux_armeabi", |
| 6533 | values = {"cpu": "armeabi"}, |
| 6534 | ) |
| 6535 | |
| 6536 | config_setting( |
Terry Heo | 68eef3f | 2020-04-13 22:53:52 -0700 | [diff] [blame] | 6537 | name = "linux_armhf", |
| 6538 | values = {"cpu": "armhf"}, |
| 6539 | ) |
| 6540 | |
| 6541 | config_setting( |
Marat Dukhan | a720e93 | 2020-06-10 13:01:11 -0700 | [diff] [blame] | 6542 | name = "linux_armv7a", |
| 6543 | values = {"cpu": "armv7a"}, |
| 6544 | ) |
| 6545 | |
| 6546 | config_setting( |
Marat Dukhan | 582094e | 2020-04-30 17:21:25 -0700 | [diff] [blame] | 6547 | name = "linux_aarch64", |
| 6548 | values = {"cpu": "aarch64"}, |
| 6549 | ) |
| 6550 | |
| 6551 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6552 | name = "android", |
| 6553 | values = {"crosstool_top": "//external:android/crosstool"}, |
| 6554 | ) |
| 6555 | |
| 6556 | config_setting( |
| 6557 | name = "android_armv7", |
| 6558 | values = { |
| 6559 | "crosstool_top": "//external:android/crosstool", |
| 6560 | "cpu": "armeabi-v7a", |
| 6561 | }, |
| 6562 | ) |
| 6563 | |
| 6564 | config_setting( |
| 6565 | name = "android_arm64", |
| 6566 | values = { |
| 6567 | "crosstool_top": "//external:android/crosstool", |
| 6568 | "cpu": "arm64-v8a", |
| 6569 | }, |
| 6570 | ) |
| 6571 | |
| 6572 | config_setting( |
| 6573 | name = "android_x86", |
| 6574 | values = { |
| 6575 | "crosstool_top": "//external:android/crosstool", |
| 6576 | "cpu": "x86", |
| 6577 | }, |
| 6578 | ) |
| 6579 | |
| 6580 | config_setting( |
| 6581 | name = "android_x86_64", |
| 6582 | values = { |
| 6583 | "crosstool_top": "//external:android/crosstool", |
| 6584 | "cpu": "x86_64", |
| 6585 | }, |
| 6586 | ) |
| 6587 | |
| 6588 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 6589 | name = "windows_x86_64", |
| 6590 | values = {"cpu": "x64_windows"}, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 6591 | ) |
| 6592 | |
| 6593 | config_setting( |
Marat Dukhan | 10a3808 | 2020-04-17 03:58:35 -0700 | [diff] [blame] | 6594 | name = "windows_x86_64_clang", |
| 6595 | values = { |
| 6596 | "compiler": "clang-cl", |
| 6597 | "cpu": "x64_windows", |
| 6598 | }, |
| 6599 | ) |
| 6600 | |
| 6601 | config_setting( |
| 6602 | name = "windows_x86_64_mingw", |
| 6603 | values = { |
| 6604 | "compiler": "mingw-gcc", |
| 6605 | "cpu": "x64_windows", |
| 6606 | }, |
| 6607 | ) |
| 6608 | |
| 6609 | config_setting( |
| 6610 | name = "windows_x86_64_msys", |
| 6611 | values = { |
| 6612 | "compiler": "msys-gcc", |
| 6613 | "cpu": "x64_windows", |
| 6614 | }, |
Marat Dukhan | 9fe932e | 2020-04-11 17:14:15 -0700 | [diff] [blame] | 6615 | ) |
| 6616 | |
| 6617 | config_setting( |
Marat Dukhan | 885ca24 | 2019-10-07 09:17:32 -0700 | [diff] [blame] | 6618 | name = "macos_x86_64", |
| 6619 | values = { |
| 6620 | "apple_platform_type": "macos", |
| 6621 | "cpu": "darwin", |
| 6622 | }, |
| 6623 | ) |
| 6624 | |
| 6625 | config_setting( |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6626 | name = "emscripten", |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6627 | values = {"crosstool_top": "//toolchain:emscripten"}, |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6628 | ) |
| 6629 | |
| 6630 | config_setting( |
| 6631 | name = "emscripten_wasm", |
| 6632 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6633 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6634 | "cpu": "wasm", |
| 6635 | }, |
| 6636 | ) |
| 6637 | |
| 6638 | config_setting( |
| 6639 | name = "emscripten_wasmsimd", |
| 6640 | values = { |
Marat Dukhan | 1a729ec | 2019-10-07 09:31:44 -0700 | [diff] [blame] | 6641 | "crosstool_top": "//toolchain:emscripten", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6642 | "cpu": "wasm", |
Marat Dukhan | 81c6260 | 2020-05-29 13:22:49 -0700 | [diff] [blame] | 6643 | "copt": "-msimd128", |
Marat Dukhan | 08c4a43 | 2019-10-03 09:29:21 -0700 | [diff] [blame] | 6644 | }, |
| 6645 | ) |
| 6646 | |
| 6647 | config_setting( |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6648 | name = "ios_armv7", |
| 6649 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6650 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6651 | "cpu": "ios_armv7", |
| 6652 | }, |
| 6653 | ) |
| 6654 | |
| 6655 | config_setting( |
| 6656 | name = "ios_arm64", |
| 6657 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6658 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6659 | "cpu": "ios_arm64", |
| 6660 | }, |
| 6661 | ) |
| 6662 | |
| 6663 | config_setting( |
| 6664 | name = "ios_arm64e", |
| 6665 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6666 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6667 | "cpu": "ios_arm64e", |
| 6668 | }, |
| 6669 | ) |
| 6670 | |
| 6671 | config_setting( |
| 6672 | name = "ios_x86", |
| 6673 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6674 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6675 | "cpu": "ios_i386", |
| 6676 | }, |
| 6677 | ) |
| 6678 | |
| 6679 | config_setting( |
| 6680 | name = "ios_x86_64", |
| 6681 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6682 | "apple_platform_type": "ios", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6683 | "cpu": "ios_x86_64", |
| 6684 | }, |
| 6685 | ) |
| 6686 | |
| 6687 | config_setting( |
| 6688 | name = "watchos_armv7k", |
| 6689 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6690 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6691 | "cpu": "watchos_armv7k", |
| 6692 | }, |
| 6693 | ) |
| 6694 | |
| 6695 | config_setting( |
| 6696 | name = "watchos_arm64_32", |
| 6697 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6698 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6699 | "cpu": "watchos_arm64_32", |
| 6700 | }, |
| 6701 | ) |
| 6702 | |
| 6703 | config_setting( |
| 6704 | name = "watchos_x86", |
| 6705 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6706 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6707 | "cpu": "watchos_i386", |
| 6708 | }, |
| 6709 | ) |
| 6710 | |
| 6711 | config_setting( |
| 6712 | name = "watchos_x86_64", |
| 6713 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6714 | "apple_platform_type": "watchos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6715 | "cpu": "watchos_x86_64", |
| 6716 | }, |
| 6717 | ) |
| 6718 | |
| 6719 | config_setting( |
| 6720 | name = "tvos_arm64", |
| 6721 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6722 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6723 | "cpu": "tvos_arm64", |
| 6724 | }, |
| 6725 | ) |
| 6726 | |
| 6727 | config_setting( |
| 6728 | name = "tvos_x86_64", |
| 6729 | values = { |
Marat Dukhan | f85fc33 | 2020-02-13 00:05:20 -0800 | [diff] [blame] | 6730 | "apple_platform_type": "tvos", |
Marat Dukhan | 1498d1d | 2020-02-11 20:00:05 -0800 | [diff] [blame] | 6731 | "cpu": "tvos_x86_64", |
| 6732 | }, |
| 6733 | ) |