blob: 3a54fd3d9bf9359a4870a7ca2e94af79c583dc81 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# 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 Dukhan10a38082020-04-17 03:58:35 -07009load(":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 Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "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 Dukhane8265432020-04-28 18:42:59 -070065 "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 Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070087 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070088 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700100 "src/subgraph/square-root.c",
101 "src/subgraph/square.c",
102 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700103 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700104 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700105 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700113 "src/tables/exp2minus-k-over-64.c",
114 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115]
116
Marat Dukhan08c4a432019-10-03 09:29:21 -0700117SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800119 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800120 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700121 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
122 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700123 "src/f32-clamp/gen/scalar-x1.c",
124 "src/f32-clamp/gen/scalar-x2.c",
125 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700126 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700127 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700128 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700129 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700130 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700133 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700134 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
135 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
143 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
144 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
147 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
148 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -0700153 "src/f32-dwconv2d-chw/3x3p1-scalar-1x1-acc3.c",
154 "src/f32-dwconv2d-chw/3x3s2p1-scalar-1x1-acc3.c",
155 "src/f32-dwconv2d-chw/5x5p2-scalar-1x1-acc5.c",
156 "src/f32-dwconv2d-chw/5x5s2p2-scalar-1x1-acc5.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700157 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700158 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
159 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700160 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
161 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
162 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700163 "src/f32-gemm/gen/1x4-minmax-scalar.c",
164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/1x4-scalar.c",
166 "src/f32-gemm/gen/2x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-relu-scalar.c",
168 "src/f32-gemm/gen/2x4-scalar.c",
169 "src/f32-gemm/gen/4x2-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-relu-scalar.c",
171 "src/f32-gemm/gen/4x2-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
173 "src/f32-gemm/gen/4x4-relu-scalar.c",
174 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700175 "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",
XNNPACK Team21432672020-10-19 19:58:48 -0700178 "src/f32-ibilinear-chw/gen/scalar-p1.c",
179 "src/f32-ibilinear-chw/gen/scalar-p2.c",
180 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700181 "src/f32-ibilinear/gen/scalar-c1.c",
182 "src/f32-ibilinear/gen/scalar-c2.c",
183 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700184 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700185 "src/f32-igemm/gen/1x4-relu-scalar.c",
186 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700188 "src/f32-igemm/gen/2x4-relu-scalar.c",
189 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700190 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700191 "src/f32-igemm/gen/4x2-relu-scalar.c",
192 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700193 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700194 "src/f32-igemm/gen/4x4-relu-scalar.c",
195 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700196 "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 Dukhan1c587112020-04-08 20:04:28 -0700199 "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 Dukhan40a672f2019-11-25 03:08:22 -0800203 "src/f32-prelu/gen/scalar-2x1.c",
204 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700210 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
215 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700216 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700217 "src/f32-relu/gen/scalar-x1.c",
218 "src/f32-relu/gen/scalar-x2.c",
219 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700220 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700221 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800222 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700225 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800228 "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 Dukhan355ab432020-04-09 19:01:52 -0700231 "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 Dukhan91cd2b72020-04-09 23:57:31 -0700241 "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 Barchard9c7308f2020-08-31 17:03:01 -0700244 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700245 "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 Barchard9c7308f2020-08-31 17:03:01 -0700248 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700249 "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 Barchard9c7308f2020-08-31 17:03:01 -0700252 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "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 Barchard9c7308f2020-08-31 17:03:01 -0700256 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700257 "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 Barchard9c7308f2020-08-31 17:03:01 -0700260 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700261 "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 Barchard9c7308f2020-08-31 17:03:01 -0700264 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "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 Barchard9c7308f2020-08-31 17:03:01 -0700268 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700269 "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 Barchard9c7308f2020-08-31 17:03:01 -0700272 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700273 "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 Barchard9c7308f2020-08-31 17:03:01 -0700276 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "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 Barchard9c7308f2020-08-31 17:03:01 -0700280 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700281 "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 Barchard9c7308f2020-08-31 17:03:01 -0700284 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700285 "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 Barchard9c7308f2020-08-31 17:03:01 -0700288 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800289 "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 Barchard9c7308f2020-08-31 17:03:01 -0700292 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800293 "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 Barchard9c7308f2020-08-31 17:03:01 -0700296 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800297 "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 Barchard9c7308f2020-08-31 17:03:01 -0700300 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800301 "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 Barchard9c7308f2020-08-31 17:03:01 -0700304 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700305 "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 Barchard9c7308f2020-08-31 17:03:01 -0700308 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700309 "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 Barchard9c7308f2020-08-31 17:03:01 -0700312 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700313 "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 Barchard9c7308f2020-08-31 17:03:01 -0700316 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700317 "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 Barchard9c7308f2020-08-31 17:03:01 -0700320 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700321 "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 Barchard9c7308f2020-08-31 17:03:01 -0700324 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700325 "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 Barchard9c7308f2020-08-31 17:03:01 -0700328 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700329 "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 Barchard9c7308f2020-08-31 17:03:01 -0700332 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700333 "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 Barchard9c7308f2020-08-31 17:03:01 -0700336 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700337 "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 Barchard9c7308f2020-08-31 17:03:01 -0700340 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700341 "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 Barchard9c7308f2020-08-31 17:03:01 -0700344 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700345 "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 Barchard9c7308f2020-08-31 17:03:01 -0700348 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700349 "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 Barchard9c7308f2020-08-31 17:03:01 -0700352 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700353 "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 Barchard9c7308f2020-08-31 17:03:01 -0700356 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700357 "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 Barchard9c7308f2020-08-31 17:03:01 -0700360 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700361 "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 Barchard9c7308f2020-08-31 17:03:01 -0700364 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700365 "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 Barchard9c7308f2020-08-31 17:03:01 -0700368 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700369 "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 Barchard9c7308f2020-08-31 17:03:01 -0700372 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700373 "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 Barchard9c7308f2020-08-31 17:03:01 -0700376 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700377 "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 Barchard9c7308f2020-08-31 17:03:01 -0700380 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700381 "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 Barchard9c7308f2020-08-31 17:03:01 -0700384 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700385 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
386 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
387 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700388 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
389 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
390 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700391 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
392 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
393 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700394 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
395 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
396 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
397 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
398 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
399 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
400 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
401 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
402 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700403 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
404 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
405 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700406 "src/f32-vunary/gen/vabs-scalar-x1.c",
407 "src/f32-vunary/gen/vabs-scalar-x2.c",
408 "src/f32-vunary/gen/vabs-scalar-x4.c",
409 "src/f32-vunary/gen/vneg-scalar-x1.c",
410 "src/f32-vunary/gen/vneg-scalar-x2.c",
411 "src/f32-vunary/gen/vneg-scalar-x4.c",
412 "src/f32-vunary/gen/vsqr-scalar-x1.c",
413 "src/f32-vunary/gen/vsqr-scalar-x2.c",
414 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800415 "src/math/expminus-scalar-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700416 "src/math/expminus-scalar-lut2048-p1.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800417 "src/math/expminus-scalar-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700418 "src/math/roundd-scalar-addsub.c",
419 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700421 "src/math/roundne-scalar-addsub.c",
422 "src/math/roundne-scalar-nearbyint.c",
423 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700424 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700425 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700426 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700427 "src/math/roundz-scalar-addsub.c",
428 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700429 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700430 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700431 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700432 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700433 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
434 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700435 "src/qs8-requantization/fp32-scalar-lrintf.c",
436 "src/qs8-requantization/fp32-scalar-magic.c",
437 "src/qs8-requantization/precise-scalar-signed64.c",
438 "src/qs8-requantization/precise-scalar-unsigned32.c",
439 "src/qs8-requantization/precise-scalar-unsigned64.c",
440 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700441 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
442 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
443 "src/qu8-dwconv/up1x9-minmax-scalar.c",
444 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
445 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700446 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700447 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
448 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700449 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700450 "src/qu8-requantization/fp32-scalar-lrintf.c",
451 "src/qu8-requantization/fp32-scalar-magic.c",
452 "src/qu8-requantization/precise-scalar-signed64.c",
453 "src/qu8-requantization/precise-scalar-unsigned32.c",
454 "src/qu8-requantization/precise-scalar-unsigned64.c",
455 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700456 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700457 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700459 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700461 "src/x8-lut/scalar.c",
462 "src/x8-zip/x2-scalar.c",
463 "src/x8-zip/x3-scalar.c",
464 "src/x8-zip/x4-scalar.c",
465 "src/x8-zip/xm-scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700466 "src/x32-fill/scalar-float.c",
467 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700468 "src/x32-packx/x2-scalar.c",
469 "src/x32-packx/x3-scalar.c",
470 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700471 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700472 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700473 "src/x32-unpool/scalar.c",
474 "src/x32-zip/x2-scalar.c",
475 "src/x32-zip/x3-scalar.c",
476 "src/x32-zip/x4-scalar.c",
477 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700478]
479
Marat Dukhan436ebe62019-12-04 15:10:12 -0800480WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700481 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
482 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700483 "src/f32-clamp/gen/wasm-x1.c",
484 "src/f32-clamp/gen/wasm-x2.c",
485 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700486 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
487 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700488 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
489 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700490 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
491 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
493 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700494 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
495 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700496 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
497 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700498 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
499 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700500 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
501 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700502 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
503 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700504 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
505 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700506 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
507 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700508 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
509 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700510 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
511 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700512 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
513 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
514 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
515 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700516 "src/f32-gemm/gen/1x4-relu-wasm.c",
517 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700519 "src/f32-gemm/gen/2x4-relu-wasm.c",
520 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700521 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700522 "src/f32-gemm/gen/4x2-relu-wasm.c",
523 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700524 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700525 "src/f32-gemm/gen/4x4-relu-wasm.c",
526 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700527 "src/f32-hswish/gen/hswish-wasm-x1.c",
528 "src/f32-hswish/gen/hswish-wasm-x2.c",
529 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700530 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700531 "src/f32-igemm/gen/1x4-relu-wasm.c",
532 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700534 "src/f32-igemm/gen/2x4-relu-wasm.c",
535 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700536 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700537 "src/f32-igemm/gen/4x2-relu-wasm.c",
538 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700539 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700540 "src/f32-igemm/gen/4x4-relu-wasm.c",
541 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700542 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
543 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
544 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700545 "src/f32-prelu/gen/wasm-2x1.c",
546 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700547 "src/f32-relu/gen/wasm-x1.c",
548 "src/f32-relu/gen/wasm-x2.c",
549 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700550 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700551 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
552 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
553 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700554 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700555 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
556 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
557 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700558 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
560 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
561 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
562 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700563 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
564 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
565 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
568 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
569 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
570 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700571 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
573 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700574 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
576 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
577 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
578 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700579 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
580 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
581 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700582 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800583 "src/f32-vbinary/gen/vmax-wasm-x1.c",
584 "src/f32-vbinary/gen/vmax-wasm-x2.c",
585 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700586 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800587 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
588 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
589 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700590 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800591 "src/f32-vbinary/gen/vmin-wasm-x1.c",
592 "src/f32-vbinary/gen/vmin-wasm-x2.c",
593 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700594 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800595 "src/f32-vbinary/gen/vminc-wasm-x1.c",
596 "src/f32-vbinary/gen/vminc-wasm-x2.c",
597 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700598 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700599 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
600 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
601 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700602 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700603 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
604 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
605 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700606 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700607 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
608 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
609 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
610 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700611 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
612 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
613 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700614 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700615 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
616 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
617 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
618 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700619 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
620 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
621 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700622 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700623 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
624 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
625 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
626 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700627 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
628 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
629 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700631 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
632 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
633 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
634 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700635 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
637 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700638 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700639 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
640 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
641 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
642 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700643 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
644 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
645 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700646 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700647 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
648 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
649 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700650 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
651 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
652 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800653]
654
Marat Dukhan290055c2020-06-09 12:24:29 -0700655WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700656 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
657 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
658 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700659 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
660 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
661 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
662 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700663 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
664 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
665 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
666 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700667 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700669 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700671 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700675 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700676 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
677 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
678 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
679 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
681 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
682 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700683 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700684 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700685 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700686 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
687 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700688 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700689 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700691 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
692 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
693 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
694 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
695 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
696 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
697 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
698 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700699 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
700 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
701 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
702 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700703 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
705 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
706 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
707 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
708 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
709 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
711 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
712 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
713 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
714 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
715 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
716 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
717 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
718 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
719 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
720 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
722 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
723 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
724 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
725 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
726 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
728 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
729 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
730 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
731 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
732 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
733 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
734 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
735 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
736 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
737 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
738 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
739 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
740 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
741 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
743 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
744 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
745 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
746 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
747 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
748 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
749 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
750 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
751 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
752 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
753 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
754 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
756 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
757 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
758 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
760 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
761 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
762 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
763 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
764 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
765 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
766 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
767 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
768 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
769 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
770 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
771 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
772 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700773 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
774 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
775 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700776 "src/f32-ibilinear/gen/wasmsimd-c4.c",
777 "src/f32-ibilinear/gen/wasmsimd-c8.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700778 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
779 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700780 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
781 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
782 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
784 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
785 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
786 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
788 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
789 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
793 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
794 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
795 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
797 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
798 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
799 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
801 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
802 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
803 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
804 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
805 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
806 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
807 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
808 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
809 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
810 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
811 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
812 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
813 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
814 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
815 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
816 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
817 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
818 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
819 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700820 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
821 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700822 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
823 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
824 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
825 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700826 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
827 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700828 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
830 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700831 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
832 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700833 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
834 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
835 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
836 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
839 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700840 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
841 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700842 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
843 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
844 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
845 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700851 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700852 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
853 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700854 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700855 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
856 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700857 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700858 "src/f32-relu/gen/wasmsimd-x4.c",
859 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700860 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700861 "src/f32-rmax/wasmsimd-arm.c",
862 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
868 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700869 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
870 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
871 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
872 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
873 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
874 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
875 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll2.c",
876 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll4.c",
877 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
878 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll2.c",
879 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll4.c",
880 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
881 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll2.c",
882 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll4.c",
883 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
884 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll2.c",
885 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll4.c",
886 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
887 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll2.c",
888 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll4.c",
889 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
890 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll2.c",
891 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll4.c",
892 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700893 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
894 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700895 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700896 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
897 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700898 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700899 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
900 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700901 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700902 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
903 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700904 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700905 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
906 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700907 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700908 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
909 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700910 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700911 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
912 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700913 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700914 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
915 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700916 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700917 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
918 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700919 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700920 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
921 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700922 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700923 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
924 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700925 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700926 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
927 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700928 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700929 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
930 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700931 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700932 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
933 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700934 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700935 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
936 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700937 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700938 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
939 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700940 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700941 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
942 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700943 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700944 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
945 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700946 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700947 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
948 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700949 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700950 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
951 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700952 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700953 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
954 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700955 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700956 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
957 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700958 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700959 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
960 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700961 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700962 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
963 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700964 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700965 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
966 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700967 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700968 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
969 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700970 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700971 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
972 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700973 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700974 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
975 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700976 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700977 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
978 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700979 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700980 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
981 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700982 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700983 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
984 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700985 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700986 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
987 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700988 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700989 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
990 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700991 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700992 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
993 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700994 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700995 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
996 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700997 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700998 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
999 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001000 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001001 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1002 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001003 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001004 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1005 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001006 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001007 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1008 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001009 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001010 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1011 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001012 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001013 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1014 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001015 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001016 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1017 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001018 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001019 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1020 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001021 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001022 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1023 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001024 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001025 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1026 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001027 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001028 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1029 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001030 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001031 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1032 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001033 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001034 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1035 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001036 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001037 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1038 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001039 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001040 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1041 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001042 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001043 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1044 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1045 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1046 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001047 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001048 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001049 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001050 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001051 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1052 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1053 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1054 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001055 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1056 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1057 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1058 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1059 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1060 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1061 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1062 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1063 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1064 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001065 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1066 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001067 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1068 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1069 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1070 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1071 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1072 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001073 "src/math/roundd-wasmsimd-addsub.c",
1074 "src/math/roundd-wasmsimd-cvt.c",
1075 "src/math/roundne-wasmsimd-addsub.c",
1076 "src/math/roundu-wasmsimd-addsub.c",
1077 "src/math/roundu-wasmsimd-cvt.c",
1078 "src/math/roundz-wasmsimd-addsub.c",
1079 "src/math/roundz-wasmsimd-cvt.c",
1080 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1081 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001082 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1083 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1084 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001085 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1086 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1087 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001088 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1089 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1090 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001091 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001092 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001093 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001094 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1095 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001096 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001097 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1098 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001099 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1100 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001101 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001102 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001103 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001104 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001105 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001106 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001107 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001108 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1109 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1110 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1111 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1112 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1113 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1114 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1115 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001116 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001117 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001118 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001119 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001120 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001121 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001122 "src/x32-zip/x2-wasmsimd.c",
1123 "src/x32-zip/x3-wasmsimd.c",
1124 "src/x32-zip/x4-wasmsimd.c",
1125 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001126]
1127
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001128PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001129 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001130 "src/f32-dwconv2d-chw/3x3p1-psimd-1x4-acc3.c",
1131 "src/f32-dwconv2d-chw/3x3s2p1-psimd-1x4-acc3.c",
1132 "src/f32-dwconv2d-chw/5x5p2-psimd-3x4.c",
1133 "src/f32-dwconv2d-chw/5x5s2p2-psimd-1x4-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001134]
1135
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001136PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001137 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001138 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001139 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001141]
1142
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143# ISA-specific micro-kernels
1144NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001145 "src/f32-argmaxpool/4x-neon-c4.c",
1146 "src/f32-argmaxpool/9p8x-neon-c4.c",
1147 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001148 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1149 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001150 "src/f32-clamp/gen/neon-x4.c",
1151 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001152 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001153 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001154 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001155 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001156 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001157 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001158 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001159 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001160 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001161 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001162 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001163 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001164 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001165 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1166 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1167 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1168 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1169 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001170 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001171 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001172 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
1173 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1174 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1175 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1176 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1177 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
1178 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1179 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1180 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
1181 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001182 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001183 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1184 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001185 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001186 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1187 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001188 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001189 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1190 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1191 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1192 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1193 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001194 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1195 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001196 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1197 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001198 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1199 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001200 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1201 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1202 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1203 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1204 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1205 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1206 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1207 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1208 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1209 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1210 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1211 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1212 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1213 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1214 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1215 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001216 "src/f32-hswish/gen/hswish-neon-x4.c",
1217 "src/f32-hswish/gen/hswish-neon-x8.c",
1218 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001219 "src/f32-ibilinear/gen/neon-c4.c",
1220 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001221 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001222 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001223 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001224 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1225 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001226 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001227 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1228 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1229 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1230 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001231 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1232 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001233 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1234 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001235 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1236 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001237 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1238 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1239 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001240 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1241 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001242 "src/f32-prelu/gen/neon-1x4.c",
1243 "src/f32-prelu/gen/neon-1x8.c",
1244 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001245 "src/f32-prelu/gen/neon-2x4.c",
1246 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001247 "src/f32-prelu/gen/neon-2x16.c",
1248 "src/f32-prelu/gen/neon-4x4.c",
1249 "src/f32-prelu/gen/neon-4x8.c",
1250 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001251 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001252 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001253 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001254 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1255 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001256 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001257 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1258 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001259 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001260 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1261 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001262 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1263 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1264 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1265 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1266 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1267 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1268 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1269 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1270 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1271 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1272 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1273 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1274 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001275 "src/f32-relu/gen/neon-x4.c",
1276 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001277 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001278 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001279 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1280 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1281 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1282 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1283 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1284 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1285 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1286 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1287 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1288 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1289 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1290 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001291 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1292 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1293 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1294 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1295 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1296 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001297 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1298 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1299 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1300 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001301 "src/f32-vbinary/gen/vmax-neon-x4.c",
1302 "src/f32-vbinary/gen/vmax-neon-x8.c",
1303 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1304 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1305 "src/f32-vbinary/gen/vmin-neon-x4.c",
1306 "src/f32-vbinary/gen/vmin-neon-x8.c",
1307 "src/f32-vbinary/gen/vminc-neon-x4.c",
1308 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001309 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1310 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1311 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1312 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1313 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1314 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001315 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1316 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1317 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1318 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001319 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1320 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1321 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1322 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001323 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1324 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001325 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1326 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001327 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1328 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001329 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1330 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1331 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1332 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1333 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1334 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001335 "src/f32-vunary/gen/vabs-neon-x4.c",
1336 "src/f32-vunary/gen/vabs-neon-x8.c",
1337 "src/f32-vunary/gen/vneg-neon-x4.c",
1338 "src/f32-vunary/gen/vneg-neon-x8.c",
1339 "src/f32-vunary/gen/vsqr-neon-x4.c",
1340 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001341 "src/math/roundd-neon-addsub.c",
1342 "src/math/roundd-neon-cvt.c",
1343 "src/math/roundne-neon-addsub.c",
1344 "src/math/roundu-neon-addsub.c",
1345 "src/math/roundu-neon-cvt.c",
1346 "src/math/roundz-neon-addsub.c",
1347 "src/math/roundz-neon-cvt.c",
1348 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1349 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1350 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1351 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1352 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1353 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1354 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1355 "src/math/sqrt-neon-nr1rsqrts.c",
1356 "src/math/sqrt-neon-nr2rsqrts.c",
1357 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001358 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1359 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1360 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1361 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001362 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1363 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1364 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1365 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001366 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1367 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1368 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1369 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001370 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001371 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001372 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001373 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1374 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001375 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001376 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001377 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001378 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001379 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001380 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001381 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1382 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1383 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1384 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1385 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1386 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1387 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1388 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001389 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1390 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1391 "src/qu8-dwconv/up8x9-minmax-neon.c",
1392 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1393 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1394 "src/qu8-gemm/4x8-minmax-neon.c",
1395 "src/qu8-gemm/8x8-minmax-neon.c",
1396 "src/qu8-igemm/4x8-minmax-neon.c",
1397 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001398 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001399 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001400 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001401 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001402 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001403 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001404 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001405 "src/x8-zip/x2-neon.c",
1406 "src/x8-zip/x3-neon.c",
1407 "src/x8-zip/x4-neon.c",
1408 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001409 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001410 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001411 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001412 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001413 "src/x32-zip/x2-neon.c",
1414 "src/x32-zip/x3-neon.c",
1415 "src/x32-zip/x4-neon.c",
1416 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001417]
1418
1419NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001420 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1421 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1422 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1423 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1424 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1425 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1426 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1427 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1428 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1429 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1430 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1431 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1432 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1433 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1434 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1435 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1436 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1437 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1438 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1439 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1440 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1441 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1442 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1443 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1444 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1445 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1446 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1447 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1448 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1449 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001450 "src/f32-ibilinear/gen/neonfma-c4.c",
1451 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001452 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001453 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001454 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001455 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1456 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001457 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1458 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001459 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1460 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001461 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1462 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001463 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001464 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001465 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001466 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1467 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001468 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001469 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1470 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001471 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001472 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1473 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001474 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1475 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1476 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1477 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1478 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1479 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1480 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1481 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1482 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1483 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1484 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1485 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1486 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001493 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001499 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1509 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1510 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001511 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1512 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1513 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1514 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1515 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1516 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001517 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1518 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1519 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1520 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1521 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1522 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001523 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1524 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1525 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1526 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1527 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1528 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1529 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1530 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1531 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1532 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1533 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1534 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1535 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1536 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1537 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1538 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1539 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1540 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001541 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1542 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001543 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1544 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1545 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1546 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1547 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1548 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1549 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1550 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1551 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1552 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1553 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1554 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1555 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1556 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1557 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1558 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1559 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1560 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1561 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1562 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001563 "src/math/exp-neonfma-lut64-p2.c",
1564 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001565 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001566 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001567 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001568 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1569 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1570 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001571 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1572 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1573 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001574 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1575 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1576 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001577 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1578 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1579 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001580 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1581 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1582 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001583 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1584 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1585 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001586 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001587 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001588 "src/math/sqrt-neonfma-nr2fma.c",
1589 "src/math/sqrt-neonfma-nr2fma1adj.c",
1590 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001591]
1592
1593AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001594 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001595 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001596 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001597 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001598 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001599 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001600 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001601 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001602 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001603 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
1604 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1605 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1606 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1607 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1608 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
1609 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1610 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1611 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
1612 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001613 "src/f32-dwconv2d-chw/3x3s2p1-neonfma-1x4-acc3.c",
1614 "src/f32-dwconv2d-chw/5x5p2-neonfma-3x4.c",
1615 "src/f32-dwconv2d-chw/5x5s2p2-neonfma-1x4-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001616 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1617 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1618 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1619 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1620 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1621 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1622 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1623 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1624 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1625 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1626 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1627 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1628 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1629 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1630 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1631 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1632 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1633 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1634 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1635 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001636 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1637 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1638 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1639 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1640 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1641 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1642 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1643 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1644 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1645 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001648 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1651 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001654 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1655 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1656 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1657 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1658 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1659 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1660 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1661 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1662 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1663 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001664 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1665 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1666 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1667 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1668 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1669 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1670 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1671 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1672 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1673 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1674 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1675 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1676 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1677 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001678 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1679 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001680 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001681 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001682 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001683 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001684 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001685 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686]
1687
Marat Dukhan8853b822020-05-07 12:19:01 -07001688NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001689 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1690 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001691 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1692 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1693 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1694 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1695 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1696 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001697 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001698 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001699 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001700 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001701]
1702
Marat Dukhan08c4a432019-10-03 09:29:21 -07001703AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001704 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001705 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001706 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1707 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1708 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1709 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001710 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1711 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1712 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1713 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1714 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1715 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1716 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1717 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001718 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1719 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001720 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1723 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1724 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1725 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1726 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1727 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1735 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001736 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001737 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1738 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1739 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1740 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1741 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1742 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1743 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1744 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1745 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001746 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001747 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001748 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001749 "src/f16-relu/gen/neonfp16arith-x16.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001750 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001751 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001752 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001753 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001754 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001755 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001756 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001757 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1758 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1759 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1760 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1761 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1762 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1763 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1764 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1765 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1766 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1767 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1768 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1769 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1770 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1771 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1772 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1773 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1774 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1775 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1776 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1777 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1778 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1779 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1780 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1781 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1782 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1783 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1784 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1785 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1786 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1787 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001788]
1789
Benoit Jacoba9644732020-08-13 12:48:55 -07001790NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001791 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001792 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001793 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001794 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001795 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001796 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001797 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001798 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001799 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001800 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001801 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001802 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001803 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001804 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001805 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001806 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001807 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001808 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001809]
1810
Marat Dukhan08c4a432019-10-03 09:29:21 -07001811SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001812 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1813 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001814 "src/f32-clamp/gen/sse-x4.c",
1815 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001816 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1817 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001818 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1819 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1820 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1821 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001822 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1823 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001824 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1825 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1826 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1827 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001828 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1829 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001830 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
1831 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1832 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1833 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1834 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1835 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
1836 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1837 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1838 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
1839 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001840 "src/f32-dwconv2d-chw/3x3s2p1-sse-1x4-acc3.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001841 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001842 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1843 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001844 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1845 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1846 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1847 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1848 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1849 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001850 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1851 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1852 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1853 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1854 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1855 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1856 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001857 "src/f32-hswish/gen/hswish-sse-x4.c",
1858 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001859 "src/f32-ibilinear/gen/sse-c4.c",
1860 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001861 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1862 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1863 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1864 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1865 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1866 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1867 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001868 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1869 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1870 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001871 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001872 "src/f32-prelu/gen/sse-2x4.c",
1873 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001874 "src/f32-relu/gen/sse-x4.c",
1875 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001877 "src/f32-spmm/gen/4x1-minmax-sse.c",
1878 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001879 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001880 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1881 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1882 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1883 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1884 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1885 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1886 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1887 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001888 "src/f32-vbinary/gen/vmax-sse-x4.c",
1889 "src/f32-vbinary/gen/vmax-sse-x8.c",
1890 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1891 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1892 "src/f32-vbinary/gen/vmin-sse-x4.c",
1893 "src/f32-vbinary/gen/vmin-sse-x8.c",
1894 "src/f32-vbinary/gen/vminc-sse-x4.c",
1895 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001896 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1897 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1898 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1899 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1900 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1901 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1902 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1903 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001904 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1905 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1906 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1907 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001908 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1909 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1910 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1911 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001912 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1913 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001914 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1915 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001916 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1917 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001918 "src/f32-vunary/gen/vabs-sse-x4.c",
1919 "src/f32-vunary/gen/vabs-sse-x8.c",
1920 "src/f32-vunary/gen/vneg-sse-x4.c",
1921 "src/f32-vunary/gen/vneg-sse-x8.c",
1922 "src/f32-vunary/gen/vsqr-sse-x4.c",
1923 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001924 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001925 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001926 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001927 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001928 "src/math/sqrt-sse-hh1mac.c",
1929 "src/math/sqrt-sse-nr1mac.c",
1930 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001931 "src/x32-fill/sse.c",
1932 "src/x32-packx/x4-sse.c",
1933 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001934]
1935
1936SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001937 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001938 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08001939 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001940 "src/f32-prelu/gen/sse2-2x4.c",
1941 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001942 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001943 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001945 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1946 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001947 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001948 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1949 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001950 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001951 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1952 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001953 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001954 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1955 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1956 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1957 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1958 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1959 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001960 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001961 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001962 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001963 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001964 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1965 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001966 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1967 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001968 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1969 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001970 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1971 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1972 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1973 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1974 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1975 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1976 "src/math/exp-sse2-lut64-p2.c",
1977 "src/math/exp-sse2-p5.c",
1978 "src/math/expminus-sse2-p5.c",
1979 "src/math/roundd-sse2-cvt.c",
1980 "src/math/roundne-sse2-cvt.c",
1981 "src/math/roundu-sse2-cvt.c",
1982 "src/math/roundz-sse2-cvt.c",
1983 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
1984 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
1985 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
1986 "src/math/sigmoid-sse2-rr2-p5-div.c",
1987 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
1988 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001989 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1990 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1991 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001992 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1993 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1994 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001995 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1996 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1997 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001998 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001999 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002000 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002001 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002002 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002003 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002004 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2005 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002006 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002007 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2008 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002009 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002010 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2011 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2012 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002013 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002014 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002015 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002016 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002017 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002018 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002019 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002020 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002021 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2022 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002023 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002024 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002025 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002026 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2027 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2028 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2029 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002030 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2031 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2032 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2033 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002034 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2035 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002036 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2037 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2038 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2039 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2040 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002041 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002042 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002043 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002044 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002045 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002046 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002047 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002048 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002049 "src/x8-zip/x2-sse2.c",
2050 "src/x8-zip/x3-sse2.c",
2051 "src/x8-zip/x4-sse2.c",
2052 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002053 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 "src/x32-zip/x2-sse2.c",
2055 "src/x32-zip/x3-sse2.c",
2056 "src/x32-zip/x4-sse2.c",
2057 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002058]
2059
2060SSSE3_UKERNELS = [
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002061 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
2062 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2063 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2064 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2065 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2066 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
2067 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
2070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002071 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2072 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2073 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002074 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2075 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2076 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002077 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2078 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2079 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002080 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002081 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002082 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002083 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002084 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002085 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002086 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2087 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002088 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002089 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2090 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002091 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002092 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2093 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2094 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002095 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002096 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002097 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002098 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002099 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002100 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002101 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002102 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002103 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2104 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002105 "src/qs8-requantization/precise-ssse3.c",
2106 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002107 "src/qu8-requantization/precise-ssse3.c",
2108 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002109]
2110
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002111SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002112 "src/f32-prelu/gen/sse41-2x4.c",
2113 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002114 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2115 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2116 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2117 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2118 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2119 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002120 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002121 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002122 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002123 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002124 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2125 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002126 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2127 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002128 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2129 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002130 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2131 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2132 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2133 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2134 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2135 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2136 "src/math/roundd-sse41.c",
2137 "src/math/roundne-sse41.c",
2138 "src/math/roundu-sse41.c",
2139 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002140 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2141 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2142 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002143 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2144 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2145 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2147 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2148 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002149 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002150 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002151 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002152 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002153 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002154 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002155 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2156 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002157 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002158 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2159 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002160 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2162 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2163 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002164 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002165 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002166 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002167 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002168 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002169 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002170 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002171 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002172 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2173 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002174 "src/qs8-requantization/fp32-sse4.c",
2175 "src/qs8-requantization/precise-sse4.c",
2176 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002177 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2178 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2179 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2180 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002181 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2182 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2183 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2184 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002185 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2186 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2187 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2188 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002189 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2190 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2191 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2192 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002193 "src/qu8-requantization/precise-sse4.c",
2194 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002195]
2196
Marat Dukhan08c4a432019-10-03 09:29:21 -07002197AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002198 "src/f32-clamp/gen/avx-x8.c",
2199 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002200 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2201 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002202 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2203 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002204 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2205 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002206 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2207 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2208 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2209 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2210 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2211 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002212 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002213 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2214 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002215 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002216 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002218 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002219 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2220 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2221 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2222 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2223 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2224 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2225 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2226 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2227 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2228 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2229 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002230 "src/f32-hswish/gen/hswish-avx-x8.c",
2231 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002232 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002233 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2234 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002236 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002238 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002239 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2240 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002241 "src/f32-prelu/gen/avx-2x8.c",
2242 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002243 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002244 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002245 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002246 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2247 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2248 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2249 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2250 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2251 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2252 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2253 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2254 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2255 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2256 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2257 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2258 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2259 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2260 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2261 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2262 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2263 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2264 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2265 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002266 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2267 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2268 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2269 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2270 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2271 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2272 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2273 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002274 "src/f32-vbinary/gen/vmax-avx-x8.c",
2275 "src/f32-vbinary/gen/vmax-avx-x16.c",
2276 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2277 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2278 "src/f32-vbinary/gen/vmin-avx-x8.c",
2279 "src/f32-vbinary/gen/vmin-avx-x16.c",
2280 "src/f32-vbinary/gen/vminc-avx-x8.c",
2281 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002282 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2283 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2284 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2285 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2286 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2287 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2288 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2289 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002290 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2291 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2292 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2293 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002294 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2295 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2296 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2297 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002298 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2299 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002300 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2301 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002302 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2303 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2304 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2305 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2306 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2307 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2308 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002309 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2310 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002311 "src/f32-vunary/gen/vabs-avx-x8.c",
2312 "src/f32-vunary/gen/vabs-avx-x16.c",
2313 "src/f32-vunary/gen/vneg-avx-x8.c",
2314 "src/f32-vunary/gen/vneg-avx-x16.c",
2315 "src/f32-vunary/gen/vsqr-avx-x8.c",
2316 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002317 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002318 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002319 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002320 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2321 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002322]
2323
Marat Dukhan1566fee2020-08-02 21:55:41 -07002324XOP_UKERNELS = [
2325 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002326 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002327 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002328 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002329 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002330 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002331 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2332 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002333 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002334 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2335 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002336 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002337 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2338 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2339 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002340 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002341 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002342 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002343 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002344 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002345 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002346 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002347 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002348 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2349 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002350 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2351 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2352 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2353 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2354 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2355 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2356 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2357 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002358]
2359
Marat Dukhanfda12b82019-11-21 12:27:59 -08002360FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002361 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2362 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002363 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2364 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2366 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002367 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2368 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2369 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2370 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2371 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2372 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002373 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002374 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2375 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2376 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2377 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002378 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002379 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2380 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002381 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002382 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2383 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002384 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2385 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2386 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002387 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2388 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2389 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2390 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2391 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2392 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2393 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2394 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2395 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2396 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2397 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2398 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2399 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2400 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002401 "src/f32-hswish/gen/hswish-fma3-x8.c",
2402 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002403 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002404 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2405 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2406 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2407 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002408 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002409 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2410 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002411 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002412 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2413 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002414 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2415 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2416 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002417 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2418 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2419 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2420 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2421 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2422 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2423 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2424 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002425 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002426 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002427 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002428]
2429
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002430AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002431 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2432 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002433 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002434 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002435 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002436 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2437 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002438 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002439 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2440 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2441 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002442 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002443 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2444 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002445 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002446 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002447 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002448 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2449 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002450 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002451 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2452 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2453 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002454 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002455 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2456 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002457 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002458 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002459 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002460 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2461 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002462 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002463 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2464 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2465 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002466 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002467 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2468 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2469 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2470 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2471 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2472 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2473 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2474 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2475 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2476 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2477 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2478 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2479 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2480 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2481 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2482 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2483 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2484 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2485 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2486 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2487 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2488 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2489 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2490 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2491 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2492 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2493 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2494 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2495 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2496 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002497 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2498 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2499 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2500 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2501 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2502 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2503 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2504 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2505 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2506 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2507 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2508 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2509 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2510 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2511 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2512 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2513 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2514 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2515 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2516 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2517 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2518 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2519 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2520 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/math/exp-avx2-lut8-p3-perm.c",
2522 "src/math/exp-avx2-lut8-p4-perm.c",
2523 "src/math/exp-avx2-p5.c",
2524 "src/math/expminus-avx2-p5.c",
2525 "src/math/extexp-avx2-p5.c",
2526 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2527 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2528 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2529 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2530 "src/math/sigmoid-avx2-rr1-p5-div.c",
2531 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2532 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2533 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2534 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2535 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2536 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2537 "src/math/sigmoid-avx2-rr2-p5-div.c",
2538 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2539 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002540 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002541 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002542 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2543 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002544 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002545 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002546 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002547 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002549 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002551 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002552 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2553 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2554 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002555 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2556 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2557 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2558 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2559 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2560 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2561 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2562 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002563]
2564
Marat Dukhan08c4a432019-10-03 09:29:21 -07002565AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002566 "src/f32-clamp/gen/avx512f-x16.c",
2567 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002568 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2569 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002570 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2571 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002572 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2573 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2575 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2576 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2577 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2578 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2579 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002580 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2581 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2582 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2583 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2584 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2585 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2587 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2588 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2589 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2590 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2591 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002592 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2593 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002594 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2595 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2596 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2597 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2598 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2599 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002600 "src/f32-prelu/gen/avx512f-2x16.c",
2601 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002602 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2603 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002604 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002605 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002606 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002607 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2608 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002609 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002610 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2611 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2612 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002613 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002614 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2615 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002616 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002617 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002618 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002619 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2620 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002621 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002622 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2623 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2624 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002625 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002626 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2627 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002628 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002629 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002630 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002631 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2632 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002633 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002634 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2635 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2636 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002637 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002638 "src/f32-relu/gen/avx512f-x16.c",
2639 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002640 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002641 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2642 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2643 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2644 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2645 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2646 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2647 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2648 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2649 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2650 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2651 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2652 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2653 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2654 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2655 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2656 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002657 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2658 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2659 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2660 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2661 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2662 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2663 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2664 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2665 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2666 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2667 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2668 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2669 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2670 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2671 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2672 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002673 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2674 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2675 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2676 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2677 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2678 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2679 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2680 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2681 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2682 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2683 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2684 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2685 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2686 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2687 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2688 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002689 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2690 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2691 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2692 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2693 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2694 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2695 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2696 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002697 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2698 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2699 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2700 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2701 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2702 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2703 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2704 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002705 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2706 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2707 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2708 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2709 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2710 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2711 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2712 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002713 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2714 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2715 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2716 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002717 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2718 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2719 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2720 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002721 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2722 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002723 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2724 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2725 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2726 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2727 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2728 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2729 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2730 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002731 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002732 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2733 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2734 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2735 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2736 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2737 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2738 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2739 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2740 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2741 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2742 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2743 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2744 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2745 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2746 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2747 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2748 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2749 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2750 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2751 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2752 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2753 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2754 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2755 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002756 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2757 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2758 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2759 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2760 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2761 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2762 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2763 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002764 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2765 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2766 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2767 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2768 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2769 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002770 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2771 "src/math/exp-avx512f-lut16-p3-perm.c",
2772 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2773 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002774 "src/math/exp-avx512f-p5-scalef.c",
2775 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002776 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002777 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002778 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002779 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002780 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002781 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002782 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002783 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002784 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002785 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2786 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2787 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2788 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2789 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2790 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2791 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2792 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2793 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2794 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002795 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002796 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2798 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2799 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2800 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002801 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002802 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002803 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002804]
2805
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002806AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002807 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2808 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002809 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2810 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2811 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2812 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2813 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2814 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2815 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2816 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2817]
2818
Frank Barchardbcedc082020-08-17 18:00:51 -07002819WASM32_ASM_UKERNELS = [
2820 "src/f32-relu/wasm_shr_x1.S",
2821 "src/f32-relu/wasm_shr_x2.S",
2822 "src/f32-relu/wasm_shr_x4.S",
2823]
2824
Marat Dukhan08c4a432019-10-03 09:29:21 -07002825AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002826 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002827 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002828 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2829 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002830 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002831 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002832 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002833 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002834 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2835 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002836 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2837 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2838 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2839 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840]
2841
2842AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002843 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002844 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002845 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002846 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002847 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002848 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002849 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002850 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2851 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2852 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2853 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2854 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2855 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2856 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002857 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2858 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002859 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2860 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2861 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002862 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2863 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002864 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2865 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2866 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2867 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002868 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002869 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2870 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002871 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2872 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2873 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2874 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002875 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002876 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002877 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002878 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002879 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2880 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2881 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2882 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2883 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2884 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2885 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2886 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2887 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2888 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2889 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2890 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2891 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2892 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2893 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2894 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2895 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2896 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2897 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2898 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2899 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2900 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002901 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002902 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002903 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2904 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002905 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2906 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2907 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2908 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2909 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2910 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002911 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2912 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2913 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2914 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002915 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2916 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002917 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002918 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002919 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002920]
2921
Marat Dukhan1b354632020-03-23 12:50:22 -07002922INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002923 "src/xnnpack/argmaxpool.h",
2924 "src/xnnpack/avgpool.h",
2925 "src/xnnpack/clamp.h",
2926 "src/xnnpack/common.h",
2927 "src/xnnpack/conv.h",
2928 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002929 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 "src/xnnpack/gavgpool.h",
2931 "src/xnnpack/gemm.h",
2932 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002933 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002935 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "src/xnnpack/lut.h",
2937 "src/xnnpack/math.h",
2938 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002939 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002940 "src/xnnpack/packx.h",
2941 "src/xnnpack/pad.h",
2942 "src/xnnpack/params.h",
2943 "src/xnnpack/pavgpool.h",
2944 "src/xnnpack/ppmm.h",
2945 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002946 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002947 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002948 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 "src/xnnpack/rmax.h",
2950 "src/xnnpack/scalar-utils.h",
2951 "src/xnnpack/spmm.h",
2952 "src/xnnpack/unpool.h",
2953 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002954 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002955 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002956 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002957 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002958 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002959 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002961]
2962
2963INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 "include/xnnpack.h",
2965 "src/xnnpack/allocator.h",
2966 "src/xnnpack/compute.h",
2967 "src/xnnpack/im2col.h",
2968 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002969 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002970 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 "src/xnnpack/operator.h",
2972 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002973 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002974 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002975 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002976 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002977]
2978
Marat Dukhan1b354632020-03-23 12:50:22 -07002979ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002980 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002981]
2982
Marat Dukhan1b354632020-03-23 12:50:22 -07002983MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002984 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 "include/xnnpack.h",
2986]
2987
Marat Dukhan1b354632020-03-23 12:50:22 -07002988MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002989 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002990 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 "src/xnnpack/requantization.h",
2992 "include/xnnpack.h",
2993]
2994
2995OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002996 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07002997 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002998]
2999
3000WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003002 "src/xnnpack/operator.h",
3003 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004]
3005
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003006LOGGING_COPTS = select({
3007 # No logging in optimized mode
3008 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3009 # Full logging in debug mode
3010 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3011 # Error-only logging in default (fastbuild) mode
3012 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3013})
3014
Marat Dukhan3b59de22020-06-03 20:15:19 -07003015LOGGING_SRCS = select({
3016 # No logging in optimized mode
3017 ":optimized_build": [],
3018 "//conditions:default": [
3019 "src/operator-strings.c",
3020 "src/subgraph-strings.c",
3021 ],
3022})
3023
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003024LOGGING_HDRS = [
3025 "src/xnnpack/log.h",
3026]
3027
Marat Dukhan08c4a432019-10-03 09:29:21 -07003028xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003029 name = "tables",
3030 srcs = TABLE_SRCS,
3031 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003032 gcc_copts = xnnpack_gcc_std_copts(),
3033 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003034)
3035
3036xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003037 name = "scalar_ukernels",
3038 srcs = SCALAR_UKERNELS,
3039 hdrs = INTERNAL_HDRS,
3040 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003041 gcc_copts = xnnpack_gcc_std_copts(),
3042 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003044 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045 "@FP16",
3046 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003047 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003048 ],
3049)
3050
3051xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003052 name = "scalar_ukernels_test_mode",
3053 srcs = SCALAR_UKERNELS,
3054 hdrs = INTERNAL_HDRS,
3055 aarch32_copts = ["-marm"],
3056 copts = [
3057 "-UNDEBUG",
3058 "-DXNN_TEST_MODE=1",
3059 ],
3060 gcc_copts = xnnpack_gcc_std_copts(),
3061 msvc_copts = xnnpack_msvc_std_copts(),
3062 deps = [
3063 ":tables",
3064 "@FP16",
3065 "@FXdiv",
3066 "@pthreadpool",
3067 ],
3068)
3069
3070xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003071 name = "wasm_ukernels",
3072 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003073 gcc_copts = xnnpack_gcc_std_copts(),
3074 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003075 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003076 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003077 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003078 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003079 "@FP16",
3080 "@FXdiv",
3081 "@pthreadpool",
3082 ],
3083)
3084
3085xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003086 name = "wasm_ukernels_test_mode",
3087 hdrs = INTERNAL_HDRS,
3088 copts = [
3089 "-UNDEBUG",
3090 "-DXNN_TEST_MODE=1",
3091 ],
3092 gcc_copts = xnnpack_gcc_std_copts(),
3093 msvc_copts = xnnpack_msvc_std_copts(),
3094 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003095 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003096 deps = [
3097 ":tables",
3098 "@FP16",
3099 "@FXdiv",
3100 "@pthreadpool",
3101 ],
3102)
3103
3104xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003105 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 hdrs = INTERNAL_HDRS,
3107 aarch32_copts = [
3108 "-marm",
3109 "-mfpu=neon",
3110 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003111 gcc_copts = xnnpack_gcc_std_copts(),
3112 gcc_x86_copts = ["-msse2"],
3113 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114 optimized_copts = [
3115 "-O3",
3116 "-ffast-math",
3117 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003118 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003120 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003121 "@FP16",
3122 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003123 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003124 ],
3125)
3126
3127xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003128 name = "psimd_fastmath_ukernels_test_mode",
3129 hdrs = INTERNAL_HDRS,
3130 aarch32_copts = [
3131 "-marm",
3132 "-mfpu=neon",
3133 ],
3134 copts = [
3135 "-UNDEBUG",
3136 "-DXNN_TEST_MODE=1",
3137 ],
3138 gcc_copts = xnnpack_gcc_std_copts(),
3139 gcc_x86_copts = ["-msse2"],
3140 msvc_copts = xnnpack_msvc_std_copts(),
3141 optimized_copts = [
3142 "-O3",
3143 "-ffast-math",
3144 ],
3145 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3146 deps = [
3147 ":tables",
3148 "@FP16",
3149 "@psimd",
3150 "@pthreadpool",
3151 ],
3152)
3153
3154xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003155 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003156 hdrs = INTERNAL_HDRS,
3157 aarch32_copts = [
3158 "-marm",
3159 "-mfpu=neon",
3160 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003161 gcc_copts = xnnpack_gcc_std_copts(),
3162 gcc_x86_copts = ["-msse2"],
3163 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003164 optimized_copts = [
3165 "-O3",
3166 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003167 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003168 deps = [
3169 ":tables",
3170 "@FP16",
3171 "@psimd",
3172 "@pthreadpool",
3173 ],
3174)
3175
3176xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003177 name = "psimd_accmath_ukernels_test_mode",
3178 hdrs = INTERNAL_HDRS,
3179 aarch32_copts = [
3180 "-marm",
3181 "-mfpu=neon",
3182 ],
3183 copts = [
3184 "-UNDEBUG",
3185 "-DXNN_TEST_MODE=1",
3186 ],
3187 gcc_copts = xnnpack_gcc_std_copts(),
3188 gcc_x86_copts = ["-msse2"],
3189 msvc_copts = xnnpack_msvc_std_copts(),
3190 optimized_copts = [
3191 "-O3",
3192 ],
3193 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3194 deps = [
3195 ":tables",
3196 "@FP16",
3197 "@psimd",
3198 "@pthreadpool",
3199 ],
3200)
3201
3202xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 name = "neon_ukernels",
3204 hdrs = INTERNAL_HDRS,
3205 aarch32_copts = [
3206 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003207 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208 "-mfpu=neon",
3209 ],
3210 aarch32_srcs = NEON_UKERNELS,
3211 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003212 gcc_copts = xnnpack_gcc_std_copts(),
3213 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003214 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003215 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003216 "@FP16",
3217 "@pthreadpool",
3218 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219)
3220
3221xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003222 name = "neon_ukernels_test_mode",
3223 hdrs = INTERNAL_HDRS,
3224 aarch32_copts = [
3225 "-marm",
3226 "-march=armv7-a",
3227 "-mfpu=neon",
3228 ],
3229 aarch32_srcs = NEON_UKERNELS,
3230 aarch64_srcs = NEON_UKERNELS,
3231 copts = [
3232 "-UNDEBUG",
3233 "-DXNN_TEST_MODE=1",
3234 ],
3235 gcc_copts = xnnpack_gcc_std_copts(),
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 deps = [
3238 ":tables",
3239 "@FP16",
3240 "@pthreadpool",
3241 ],
3242)
3243
3244xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 name = "neonfma_ukernels",
3246 hdrs = INTERNAL_HDRS,
3247 aarch32_copts = [
3248 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003249 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 "-mfpu=neon-vfpv4",
3251 ],
3252 aarch32_srcs = NEONFMA_UKERNELS,
3253 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003254 apple_aarch32_copts = [
3255 "-mcpu=swift",
3256 "-mtune=generic",
3257 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003258 gcc_copts = xnnpack_gcc_std_copts(),
3259 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003260 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003261 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003262 "@FP16",
3263 "@pthreadpool",
3264 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265)
3266
3267xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003268 name = "neonfma_ukernels_test_mode",
3269 hdrs = INTERNAL_HDRS,
3270 aarch32_copts = [
3271 "-marm",
3272 "-march=armv7-a",
3273 "-mfpu=neon-vfpv4",
3274 ],
3275 aarch32_srcs = NEONFMA_UKERNELS,
3276 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003277 apple_aarch32_copts = [
3278 "-mcpu=swift",
3279 "-mtune=generic",
3280 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003281 copts = [
3282 "-UNDEBUG",
3283 "-DXNN_TEST_MODE=1",
3284 ],
3285 gcc_copts = xnnpack_gcc_std_copts(),
3286 msvc_copts = xnnpack_msvc_std_copts(),
3287 deps = [
3288 ":tables",
3289 "@FP16",
3290 "@pthreadpool",
3291 ],
3292)
3293
3294xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003295 name = "neonv8_ukernels",
3296 hdrs = INTERNAL_HDRS,
3297 aarch32_copts = [
3298 "-marm",
3299 "-march=armv8-a",
3300 "-mfpu=neon-fp-armv8",
3301 ],
3302 aarch32_srcs = NEONV8_UKERNELS,
3303 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003304 apple_aarch32_copts = [
3305 "-mcpu=cyclone",
3306 "-mtune=generic",
3307 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003308 gcc_copts = xnnpack_gcc_std_copts(),
3309 msvc_copts = xnnpack_msvc_std_copts(),
3310 deps = [
3311 ":tables",
3312 "@FP16",
3313 "@pthreadpool",
3314 ],
3315)
3316
3317xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003318 name = "neonv8_ukernels_test_mode",
3319 hdrs = INTERNAL_HDRS,
3320 aarch32_copts = [
3321 "-marm",
3322 "-march=armv8-a",
3323 "-mfpu=neon-fp-armv8",
3324 ],
3325 aarch32_srcs = NEONV8_UKERNELS,
3326 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003327 apple_aarch32_copts = [
3328 "-mcpu=cyclone",
3329 "-mtune=generic",
3330 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003331 copts = [
3332 "-UNDEBUG",
3333 "-DXNN_TEST_MODE=1",
3334 ],
3335 gcc_copts = xnnpack_gcc_std_copts(),
3336 msvc_copts = xnnpack_msvc_std_copts(),
3337 deps = [
3338 ":tables",
3339 "@FP16",
3340 "@pthreadpool",
3341 ],
3342)
3343
3344xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345 name = "neonfp16arith_ukernels",
3346 hdrs = INTERNAL_HDRS,
3347 aarch64_copts = ["-march=armv8.2-a+fp16"],
3348 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003349 gcc_copts = xnnpack_gcc_std_copts(),
3350 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003351 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003352 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003353 "@FP16",
3354 "@pthreadpool",
3355 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003356)
3357
3358xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003359 name = "neonfp16arith_ukernels_test_mode",
3360 hdrs = INTERNAL_HDRS,
3361 aarch64_copts = ["-march=armv8.2-a+fp16"],
3362 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3363 copts = [
3364 "-UNDEBUG",
3365 "-DXNN_TEST_MODE=1",
3366 ],
3367 gcc_copts = xnnpack_gcc_std_copts(),
3368 msvc_copts = xnnpack_msvc_std_copts(),
3369 deps = [
3370 ":tables",
3371 "@FP16",
3372 "@pthreadpool",
3373 ],
3374)
3375
3376xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003377 name = "neondot_ukernels",
3378 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003379 aarch32_copts = [
3380 "-marm",
3381 "-march=armv8.2-a+dotprod",
3382 "-mfpu=neon-fp-armv8",
3383 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003384 aarch32_srcs = NEONDOT_UKERNELS,
3385 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3386 aarch64_srcs = NEONDOT_UKERNELS,
3387 gcc_copts = xnnpack_gcc_std_copts(),
3388 msvc_copts = xnnpack_msvc_std_copts(),
3389 deps = [
3390 ":tables",
3391 "@FP16",
3392 "@pthreadpool",
3393 ],
3394)
3395
3396xnnpack_cc_library(
3397 name = "neondot_ukernels_test_mode",
3398 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003399 aarch32_copts = [
3400 "-marm",
3401 "-march=armv8.2-a+dotprod",
3402 "-mfpu=neon-fp-armv8",
3403 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003404 aarch32_srcs = NEONDOT_UKERNELS,
3405 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3406 aarch64_srcs = NEONDOT_UKERNELS,
3407 copts = [
3408 "-UNDEBUG",
3409 "-DXNN_TEST_MODE=1",
3410 ],
3411 gcc_copts = xnnpack_gcc_std_copts(),
3412 msvc_copts = xnnpack_msvc_std_copts(),
3413 deps = [
3414 ":tables",
3415 "@FP16",
3416 "@pthreadpool",
3417 ],
3418)
3419
3420xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421 name = "sse2_ukernels",
3422 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003423 gcc_copts = xnnpack_gcc_std_copts(),
3424 gcc_x86_copts = ["-msse2"],
3425 msvc_copts = xnnpack_msvc_std_copts(),
3426 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003428 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003429 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003430 "@FP16",
3431 "@pthreadpool",
3432 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433)
3434
3435xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003436 name = "sse2_ukernels_test_mode",
3437 hdrs = INTERNAL_HDRS,
3438 copts = [
3439 "-UNDEBUG",
3440 "-DXNN_TEST_MODE=1",
3441 ],
3442 gcc_copts = xnnpack_gcc_std_copts(),
3443 gcc_x86_copts = ["-msse2"],
3444 msvc_copts = xnnpack_msvc_std_copts(),
3445 msvc_x86_32_copts = ["/arch:SSE2"],
3446 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3447 deps = [
3448 ":tables",
3449 "@FP16",
3450 "@pthreadpool",
3451 ],
3452)
3453
3454xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003455 name = "ssse3_ukernels",
3456 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003457 gcc_copts = xnnpack_gcc_std_copts(),
3458 gcc_x86_copts = ["-mssse3"],
3459 msvc_copts = xnnpack_msvc_std_copts(),
3460 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003461 x86_srcs = SSSE3_UKERNELS,
3462 deps = [
3463 ":tables",
3464 "@FP16",
3465 "@pthreadpool",
3466 ],
3467)
3468
3469xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003470 name = "ssse3_ukernels_test_mode",
3471 hdrs = INTERNAL_HDRS,
3472 copts = [
3473 "-UNDEBUG",
3474 "-DXNN_TEST_MODE=1",
3475 ],
3476 gcc_copts = xnnpack_gcc_std_copts(),
3477 gcc_x86_copts = ["-mssse3"],
3478 msvc_copts = xnnpack_msvc_std_copts(),
3479 msvc_x86_32_copts = ["/arch:SSE2"],
3480 x86_srcs = SSSE3_UKERNELS,
3481 deps = [
3482 ":tables",
3483 "@FP16",
3484 "@pthreadpool",
3485 ],
3486)
3487
3488xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003489 name = "sse41_ukernels",
3490 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003491 gcc_copts = xnnpack_gcc_std_copts(),
3492 gcc_x86_copts = ["-msse4.1"],
3493 msvc_copts = xnnpack_msvc_std_copts(),
3494 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003495 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003496 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003497 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003498 "@FP16",
3499 "@pthreadpool",
3500 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003501)
3502
3503xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003504 name = "sse41_ukernels_test_mode",
3505 hdrs = INTERNAL_HDRS,
3506 copts = [
3507 "-UNDEBUG",
3508 "-DXNN_TEST_MODE=1",
3509 ],
3510 gcc_copts = xnnpack_gcc_std_copts(),
3511 gcc_x86_copts = ["-msse4.1"],
3512 msvc_copts = xnnpack_msvc_std_copts(),
3513 msvc_x86_32_copts = ["/arch:SSE2"],
3514 x86_srcs = SSE41_UKERNELS,
3515 deps = [
3516 ":tables",
3517 "@FP16",
3518 "@pthreadpool",
3519 ],
3520)
3521
3522xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003523 name = "avx_ukernels",
3524 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003525 gcc_copts = xnnpack_gcc_std_copts(),
3526 gcc_x86_copts = ["-mavx"],
3527 msvc_copts = xnnpack_msvc_std_copts(),
3528 msvc_x86_32_copts = ["/arch:AVX"],
3529 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003531 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003532 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003533 "@FP16",
3534 "@pthreadpool",
3535 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536)
3537
3538xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003539 name = "avx_ukernels_test_mode",
3540 hdrs = INTERNAL_HDRS,
3541 copts = [
3542 "-UNDEBUG",
3543 "-DXNN_TEST_MODE=1",
3544 ],
3545 gcc_copts = xnnpack_gcc_std_copts(),
3546 gcc_x86_copts = ["-mavx"],
3547 msvc_copts = xnnpack_msvc_std_copts(),
3548 msvc_x86_32_copts = ["/arch:AVX"],
3549 msvc_x86_64_copts = ["/arch:AVX"],
3550 x86_srcs = AVX_UKERNELS,
3551 deps = [
3552 ":tables",
3553 "@FP16",
3554 "@pthreadpool",
3555 ],
3556)
3557
3558xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003559 name = "xop_ukernels",
3560 hdrs = INTERNAL_HDRS,
3561 gcc_copts = xnnpack_gcc_std_copts(),
3562 gcc_x86_copts = ["-mxop"],
3563 msvc_copts = xnnpack_msvc_std_copts(),
3564 msvc_x86_32_copts = ["/arch:AVX"],
3565 msvc_x86_64_copts = ["/arch:AVX"],
3566 x86_srcs = XOP_UKERNELS,
3567 deps = [
3568 ":tables",
3569 "@FP16",
3570 "@pthreadpool",
3571 ],
3572)
3573
3574xnnpack_cc_library(
3575 name = "xop_ukernels_test_mode",
3576 hdrs = INTERNAL_HDRS,
3577 copts = [
3578 "-UNDEBUG",
3579 "-DXNN_TEST_MODE=1",
3580 ],
3581 gcc_copts = xnnpack_gcc_std_copts(),
3582 gcc_x86_copts = ["-mxop"],
3583 msvc_copts = xnnpack_msvc_std_copts(),
3584 msvc_x86_32_copts = ["/arch:AVX"],
3585 msvc_x86_64_copts = ["/arch:AVX"],
3586 x86_srcs = XOP_UKERNELS,
3587 deps = [
3588 ":tables",
3589 "@FP16",
3590 "@pthreadpool",
3591 ],
3592)
3593
3594xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003595 name = "fma3_ukernels",
3596 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003597 gcc_copts = xnnpack_gcc_std_copts(),
3598 gcc_x86_copts = ["-mfma"],
3599 msvc_copts = xnnpack_msvc_std_copts(),
3600 msvc_x86_32_copts = ["/arch:AVX"],
3601 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003602 x86_srcs = FMA3_UKERNELS,
3603 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003604 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003605 "@FP16",
3606 "@pthreadpool",
3607 ],
3608)
3609
3610xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003611 name = "fma3_ukernels_test_mode",
3612 hdrs = INTERNAL_HDRS,
3613 copts = [
3614 "-UNDEBUG",
3615 "-DXNN_TEST_MODE=1",
3616 ],
3617 gcc_copts = xnnpack_gcc_std_copts(),
3618 gcc_x86_copts = ["-mfma"],
3619 msvc_copts = xnnpack_msvc_std_copts(),
3620 msvc_x86_32_copts = ["/arch:AVX"],
3621 msvc_x86_64_copts = ["/arch:AVX"],
3622 x86_srcs = FMA3_UKERNELS,
3623 deps = [
3624 ":tables",
3625 "@FP16",
3626 "@pthreadpool",
3627 ],
3628)
3629
3630xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003631 name = "avx2_ukernels",
3632 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003633 gcc_copts = xnnpack_gcc_std_copts(),
3634 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003635 "-mfma",
3636 "-mavx2",
3637 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003638 msvc_copts = xnnpack_msvc_std_copts(),
3639 msvc_x86_32_copts = ["/arch:AVX2"],
3640 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003641 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003642 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003643 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003644 "@FP16",
3645 "@pthreadpool",
3646 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003647)
3648
3649xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003650 name = "avx2_ukernels_test_mode",
3651 hdrs = INTERNAL_HDRS,
3652 copts = [
3653 "-UNDEBUG",
3654 "-DXNN_TEST_MODE=1",
3655 ],
3656 gcc_copts = xnnpack_gcc_std_copts(),
3657 gcc_x86_copts = [
3658 "-mfma",
3659 "-mavx2",
3660 ],
3661 msvc_copts = xnnpack_msvc_std_copts(),
3662 msvc_x86_32_copts = ["/arch:AVX2"],
3663 msvc_x86_64_copts = ["/arch:AVX2"],
3664 x86_srcs = AVX2_UKERNELS,
3665 deps = [
3666 ":tables",
3667 "@FP16",
3668 "@pthreadpool",
3669 ],
3670)
3671
3672xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003673 name = "avx512f_ukernels",
3674 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003675 gcc_copts = xnnpack_gcc_std_copts(),
3676 gcc_x86_copts = ["-mavx512f"],
3677 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3678 msvc_copts = xnnpack_msvc_std_copts(),
3679 msvc_x86_32_copts = ["/arch:AVX512"],
3680 msvc_x86_64_copts = ["/arch:AVX512"],
3681 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003682 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003683 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003684 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003685 "@FP16",
3686 "@pthreadpool",
3687 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003688)
3689
3690xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003691 name = "avx512f_ukernels_test_mode",
3692 hdrs = INTERNAL_HDRS,
3693 copts = [
3694 "-UNDEBUG",
3695 "-DXNN_TEST_MODE=1",
3696 ],
3697 gcc_copts = xnnpack_gcc_std_copts(),
3698 gcc_x86_copts = ["-mavx512f"],
3699 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3700 msvc_copts = xnnpack_msvc_std_copts(),
3701 msvc_x86_32_copts = ["/arch:AVX512"],
3702 msvc_x86_64_copts = ["/arch:AVX512"],
3703 msys_copts = ["-fno-asynchronous-unwind-tables"],
3704 x86_srcs = AVX512F_UKERNELS,
3705 deps = [
3706 ":tables",
3707 "@FP16",
3708 "@pthreadpool",
3709 ],
3710)
3711
3712xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003713 name = "avx512skx_ukernels",
3714 hdrs = INTERNAL_HDRS,
3715 gcc_copts = xnnpack_gcc_std_copts(),
3716 gcc_x86_copts = [
3717 "-mavx512f",
3718 "-mavx512cd",
3719 "-mavx512bw",
3720 "-mavx512dq",
3721 "-mavx512vl",
3722 ],
3723 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3724 msvc_copts = xnnpack_msvc_std_copts(),
3725 msvc_x86_32_copts = ["/arch:AVX512"],
3726 msvc_x86_64_copts = ["/arch:AVX512"],
3727 msys_copts = ["-fno-asynchronous-unwind-tables"],
3728 x86_srcs = AVX512SKX_UKERNELS,
3729 deps = [
3730 ":tables",
3731 "@FP16",
3732 "@pthreadpool",
3733 ],
3734)
3735
3736xnnpack_cc_library(
3737 name = "avx512skx_ukernels_test_mode",
3738 hdrs = INTERNAL_HDRS,
3739 copts = [
3740 "-UNDEBUG",
3741 "-DXNN_TEST_MODE=1",
3742 ],
3743 gcc_copts = xnnpack_gcc_std_copts(),
3744 gcc_x86_copts = [
3745 "-mavx512f",
3746 "-mavx512cd",
3747 "-mavx512bw",
3748 "-mavx512dq",
3749 "-mavx512vl",
3750 ],
3751 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3752 msvc_copts = xnnpack_msvc_std_copts(),
3753 msvc_x86_32_copts = ["/arch:AVX512"],
3754 msvc_x86_64_copts = ["/arch:AVX512"],
3755 msys_copts = ["-fno-asynchronous-unwind-tables"],
3756 x86_srcs = AVX512SKX_UKERNELS,
3757 deps = [
3758 ":tables",
3759 "@FP16",
3760 "@pthreadpool",
3761 ],
3762)
3763
3764xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003765 name = "asm_ukernels",
3766 hdrs = ["src/xnnpack/assembly.h"],
3767 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003768 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003770 wasm_srcs = WASM32_ASM_UKERNELS,
3771 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772)
3773
Marat Dukhan3b59de22020-06-03 20:15:19 -07003774xnnpack_cc_library(
3775 name = "logging_utils",
3776 srcs = LOGGING_SRCS,
3777 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3778 copts = LOGGING_COPTS + [
3779 "-Isrc",
3780 "-Iinclude",
3781 ] + select({
3782 ":debug_build": [],
3783 "//conditions:default": xnnpack_min_size_copts(),
3784 }),
3785 gcc_copts = xnnpack_gcc_std_copts(),
3786 msvc_copts = xnnpack_msvc_std_copts(),
3787 visibility = xnnpack_visibility(),
3788 deps = [
3789 "@FP16",
3790 "@clog",
3791 "@pthreadpool",
3792 ],
3793)
3794
Marat Dukhan08c4a432019-10-03 09:29:21 -07003795xnnpack_aggregate_library(
3796 name = "ukernels",
3797 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003798 ":neon_ukernels",
3799 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003800 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003801 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802 ":asm_ukernels",
3803 ],
3804 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805 ":neon_ukernels",
3806 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003807 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003809 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810 ":asm_ukernels",
3811 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003812 generic_deps = [
3813 ":scalar_ukernels",
3814 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003815 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003816 ":psimd_fastmath_ukernels",
3817 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003818 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003819 wasm_deps = [
3820 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003821 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003822 ],
3823 wasmsimd_deps = [
3824 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003825 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003826 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003829 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003830 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003831 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003832 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003833 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003834 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003835 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003836 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003837 ],
3838)
3839
Marat Dukhan33fcf782020-05-24 14:27:15 -07003840xnnpack_aggregate_library(
3841 name = "ukernels_test_mode",
3842 aarch32_deps = [
3843 ":neon_ukernels_test_mode",
3844 ":neonfma_ukernels_test_mode",
3845 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003846 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003847 ":asm_ukernels",
3848 ],
3849 aarch64_deps = [
3850 ":neon_ukernels_test_mode",
3851 ":neonfma_ukernels_test_mode",
3852 ":neonv8_ukernels_test_mode",
3853 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003854 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003855 ":asm_ukernels",
3856 ],
3857 generic_deps = [
3858 ":scalar_ukernels_test_mode",
3859 ],
3860 psimd_deps = [
3861 ":psimd_fastmath_ukernels_test_mode",
3862 ":psimd_accmath_ukernels_test_mode",
3863 ],
3864 wasm_deps = [
3865 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003866 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003867 ],
3868 wasmsimd_deps = [
3869 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003870 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003871 ],
3872 x86_deps = [
3873 ":sse2_ukernels_test_mode",
3874 ":ssse3_ukernels_test_mode",
3875 ":sse41_ukernels_test_mode",
3876 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003877 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003878 ":fma3_ukernels_test_mode",
3879 ":avx2_ukernels_test_mode",
3880 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003881 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003882 ],
3883)
3884
Marat Dukhan86df3842020-06-17 14:03:31 -07003885# This target must be compatible with Android NDK r17 and compile with Clang 6.
3886# It is similar to :ukernels target, but excludes microkernels which require
3887# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003888xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003889 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003890 aarch32_deps = [
3891 ":neon_ukernels",
3892 ":neonfma_ukernels",
3893 ":neonv8_ukernels",
3894 ":asm_ukernels",
3895 ],
3896 aarch64_deps = [
3897 ":neon_ukernels",
3898 ":neonfma_ukernels",
3899 ":neonv8_ukernels",
3900 ":asm_ukernels",
3901 ],
3902 generic_deps = [
3903 ":scalar_ukernels",
3904 ],
3905 psimd_deps = [
3906 ":psimd_fastmath_ukernels",
3907 ":psimd_accmath_ukernels",
3908 ],
3909 wasm_deps = [
3910 ":wasm_ukernels",
3911 ],
3912 wasmsimd_deps = [
3913 ":wasm_ukernels",
3914 ],
3915 x86_deps = [
3916 ":sse2_ukernels",
3917 ":ssse3_ukernels",
3918 ":sse41_ukernels",
3919 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003920 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003921 ":fma3_ukernels",
3922 ":avx2_ukernels",
3923 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003924 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003925 ],
3926)
3927
Marat Dukhan08c4a432019-10-03 09:29:21 -07003928xnnpack_cc_library(
3929 name = "im2col",
3930 srcs = ["src/im2col.c"],
3931 hdrs = [
3932 "src/xnnpack/common.h",
3933 "src/xnnpack/im2col.h",
3934 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003935 gcc_copts = xnnpack_gcc_std_copts(),
3936 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003937)
3938
3939xnnpack_cc_library(
3940 name = "indirection",
3941 srcs = ["src/indirection.c"],
3942 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003943 gcc_copts = xnnpack_gcc_std_copts(),
3944 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003945 deps = [
3946 "@FP16",
3947 "@FXdiv",
3948 "@pthreadpool",
3949 ],
3950)
3951
3952xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003953 name = "indirection_test_mode",
3954 srcs = ["src/indirection.c"],
3955 hdrs = INTERNAL_HDRS,
3956 copts = [
3957 "-UNDEBUG",
3958 "-DXNN_TEST_MODE=1",
3959 ],
3960 gcc_copts = xnnpack_gcc_std_copts(),
3961 msvc_copts = xnnpack_msvc_std_copts(),
3962 deps = [
3963 "@FP16",
3964 "@FXdiv",
3965 "@pthreadpool",
3966 ],
3967)
3968
3969xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003970 name = "packing",
3971 srcs = ["src/packing.c"],
3972 hdrs = INTERNAL_HDRS,
3973 gcc_copts = xnnpack_gcc_std_copts(),
3974 msvc_copts = xnnpack_msvc_std_copts(),
3975 deps = [
3976 "@FP16",
3977 "@FXdiv",
3978 "@pthreadpool",
3979 ],
3980)
3981
3982xnnpack_cc_library(
3983 name = "packing_test_mode",
3984 srcs = ["src/packing.c"],
3985 hdrs = INTERNAL_HDRS,
3986 copts = [
3987 "-UNDEBUG",
3988 "-DXNN_TEST_MODE=1",
3989 ],
3990 gcc_copts = xnnpack_gcc_std_copts(),
3991 msvc_copts = xnnpack_msvc_std_copts(),
3992 deps = [
3993 "@FP16",
3994 "@FXdiv",
3995 "@pthreadpool",
3996 ],
3997)
3998
3999xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004000 name = "operator_run",
4001 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004002 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004003 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004004 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4005 "//conditions:default": [],
4006 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004007 gcc_copts = xnnpack_gcc_std_copts(),
4008 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004009 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004010 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004011 "@FP16",
4012 "@FXdiv",
4013 "@clog",
4014 "@pthreadpool",
4015 ],
4016)
4017
Chao Mei6ddfc602020-05-13 22:29:36 -07004018xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004019 name = "operator_run_test_mode",
4020 srcs = ["src/operator-run.c"],
4021 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4022 copts = LOGGING_COPTS + [
4023 "-UNDEBUG",
4024 "-DXNN_TEST_MODE=1",
4025 ] + select({
4026 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4027 "//conditions:default": [],
4028 }),
4029 gcc_copts = xnnpack_gcc_std_copts(),
4030 msvc_copts = xnnpack_msvc_std_copts(),
4031 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004032 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004033 "@FP16",
4034 "@FXdiv",
4035 "@clog",
4036 "@pthreadpool",
4037 ],
4038)
4039
4040xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004041 name = "memory_planner",
4042 srcs = ["src/memory-planner.c"],
4043 hdrs = INTERNAL_HDRS,
4044 defines = select({
4045 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4046 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4047 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4048 }),
4049 gcc_copts = xnnpack_gcc_std_copts(),
4050 msvc_copts = xnnpack_msvc_std_copts(),
4051 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004052 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004053 "@pthreadpool",
4054 ],
4055)
4056
Marat Dukhan33fcf782020-05-24 14:27:15 -07004057xnnpack_cc_library(
4058 name = "memory_planner_test_mode",
4059 srcs = ["src/memory-planner.c"],
4060 hdrs = INTERNAL_HDRS,
4061 copts = [
4062 "-UNDEBUG",
4063 "-DXNN_TEST_MODE=1",
4064 ],
4065 defines = select({
4066 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4067 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4068 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4069 }),
4070 gcc_copts = xnnpack_gcc_std_copts(),
4071 msvc_copts = xnnpack_msvc_std_copts(),
4072 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004073 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004074 "@pthreadpool",
4075 ],
4076)
4077
Marat Dukhan08c4a432019-10-03 09:29:21 -07004078cc_library(
4079 name = "enable_assembly",
4080 defines = select({
4081 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4082 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004083 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004084 }),
4085)
4086
Marat Dukhan9de90e02020-06-18 16:04:12 -07004087cc_library(
4088 name = "enable_sparse",
4089 defines = select({
4090 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4091 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004092 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004093 }),
4094)
4095
Marat Dukhancf056b22019-10-07 10:26:29 -07004096xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004097 name = "operators",
4098 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004099 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004100 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004101 ],
4102 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004103 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004104 "-Isrc",
4105 "-Iinclude",
4106 ] + select({
4107 ":debug_build": [],
4108 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004109 }) + select({
4110 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4111 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004112 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004113 gcc_copts = xnnpack_gcc_std_copts(),
4114 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004115 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004116 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004117 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004118 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004119 "@FP16",
4120 "@FXdiv",
4121 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004122 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004123 ],
4124)
4125
Marat Dukhan10a38082020-04-17 03:58:35 -07004126xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004127 name = "operators_test_mode",
4128 srcs = OPERATOR_SRCS + [
4129 "src/memory.c",
4130 "src/operator-delete.c",
4131 ],
4132 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4133 copts = LOGGING_COPTS + [
4134 "-Isrc",
4135 "-Iinclude",
4136 "-UNDEBUG",
4137 "-DXNN_TEST_MODE=1",
4138 ] + select({
4139 ":debug_build": [],
4140 "//conditions:default": xnnpack_min_size_copts(),
4141 }) + select({
4142 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4143 "//conditions:default": [],
4144 }),
4145 gcc_copts = xnnpack_gcc_std_copts(),
4146 msvc_copts = xnnpack_msvc_std_copts(),
4147 deps = [
4148 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004149 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004150 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004151 "@FP16",
4152 "@FXdiv",
4153 "@clog",
4154 "@pthreadpool",
4155 ],
4156)
4157
4158xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004159 name = "XNNPACK",
4160 srcs = [
4161 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004162 "src/runtime.c",
4163 "src/subgraph.c",
4164 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004165 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004166 hdrs = ["include/xnnpack.h"],
4167 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004168 "-Isrc",
4169 "-Iinclude",
4170 ] + select({
4171 ":debug_build": [],
4172 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004173 }) + select({
4174 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4175 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004176 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004177 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004178 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004179 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004180 visibility = xnnpack_visibility(),
4181 deps = [
4182 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004183 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004184 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004185 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004186 ":operator_run",
4187 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004188 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004189 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004190 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004191 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004192 ] + select({
4193 ":emscripten": [],
4194 "//conditions:default": ["@cpuinfo"],
4195 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196)
4197
Marat Dukhan10a38082020-04-17 03:58:35 -07004198xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004199 name = "XNNPACK_test_mode",
4200 srcs = [
4201 "src/init.c",
4202 "src/runtime.c",
4203 "src/subgraph.c",
4204 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004205 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004206 hdrs = ["include/xnnpack.h"],
4207 copts = LOGGING_COPTS + [
4208 "-Isrc",
4209 "-Iinclude",
4210 "-UNDEBUG",
4211 "-DXNN_TEST_MODE=1",
4212 ] + select({
4213 ":debug_build": [],
4214 "//conditions:default": xnnpack_min_size_copts(),
4215 }) + select({
4216 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4217 "//conditions:default": [],
4218 }),
4219 gcc_copts = xnnpack_gcc_std_copts(),
4220 includes = ["include"],
4221 msvc_copts = xnnpack_msvc_std_copts(),
4222 visibility = xnnpack_visibility(),
4223 deps = [
4224 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004225 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004226 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004227 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004228 ":operator_run_test_mode",
4229 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004230 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004231 "@clog",
4232 "@FP16",
4233 "@pthreadpool",
4234 ] + select({
4235 ":emscripten": [],
4236 "//conditions:default": ["@cpuinfo"],
4237 }),
4238)
4239
4240xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004241 name = "xnnpack_f32",
4242 srcs = [
4243 "src/init.c",
4244 "src/runtime.c",
4245 "src/subgraph.c",
4246 "src/tensor.c",
4247 ] + SUBGRAPH_SRCS,
4248 hdrs = ["include/xnnpack.h"],
4249 copts = LOGGING_COPTS + [
4250 "-Isrc",
4251 "-Iinclude",
4252 ] + select({
4253 ":debug_build": [],
4254 "//conditions:default": xnnpack_min_size_copts(),
4255 }) + select({
4256 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4257 "//conditions:default": [],
4258 }),
4259 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004260 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004261 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004262 "XNN_NO_U8_OPERATORS",
4263 "XNN_NO_X8_OPERATORS",
4264 "XNN_NO_F16_OPERATORS",
4265 "XNN_NO_X16_OPERATORS",
4266 ],
4267 gcc_copts = xnnpack_gcc_std_copts(),
4268 includes = ["include"],
4269 msvc_copts = xnnpack_msvc_std_copts(),
4270 visibility = xnnpack_visibility(),
4271 deps = [
4272 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004273 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004274 ":logging_utils",
4275 ":memory_planner",
4276 ":operator_run",
4277 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004278 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004279 "@clog",
4280 "@FP16",
4281 "@pthreadpool",
4282 ] + select({
4283 ":emscripten": [],
4284 "//conditions:default": ["@cpuinfo"],
4285 }),
4286)
4287
4288xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004289 name = "xnnpack_operators_nhwc_f32",
4290 srcs = [
4291 "src/init.c",
4292 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004293 hdrs = ["include/xnnpack.h"],
4294 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004295 "-Isrc",
4296 "-Iinclude",
4297 ] + select({
4298 ":debug_build": [],
4299 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004300 }) + select({
4301 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4302 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004303 }),
4304 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004305 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004306 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004307 "XNN_NO_U8_OPERATORS",
4308 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004309 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004310 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004311 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004313 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004314 visibility = xnnpack_visibility(),
4315 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004316 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004317 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318 ":operator_run",
4319 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004320 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004321 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004322 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004323 ] + select({
4324 ":emscripten": [],
4325 "//conditions:default": ["@cpuinfo"],
4326 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327)
4328
Marat Dukhancf056b22019-10-07 10:26:29 -07004329xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004330 name = "bench_utils",
4331 srcs = ["bench/utils.cc"],
4332 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004333 deps = [
4334 "@com_google_benchmark//:benchmark",
4335 "@cpuinfo",
4336 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337)
4338
Frank Barchard7e955972019-10-11 10:34:25 -07004339######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004340
4341xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004342 name = "qs8_gemm_bench",
4343 srcs = [
4344 "bench/gemm.h",
4345 "bench/qs8-gemm.cc",
4346 "src/xnnpack/AlignedAllocator.h",
4347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004348 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4349 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004350)
4351
4352xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004353 name = "qs8_requantization_bench",
4354 srcs = [
4355 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004356 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004357 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004358 ] + MICROKERNEL_BENCHMARK_HDRS,
4359 deps = MICROKERNEL_BENCHMARK_DEPS,
4360)
4361
4362xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004363 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 srcs = [
4365 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004366 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004367 "src/xnnpack/AlignedAllocator.h",
4368 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004369 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004370 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004371)
4372
4373xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004374 name = "qu8_requantization_bench",
4375 srcs = [
4376 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004377 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004378 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004379 ] + MICROKERNEL_BENCHMARK_HDRS,
4380 deps = MICROKERNEL_BENCHMARK_DEPS,
4381)
4382
4383xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004384 name = "f16_igemm_bench",
4385 srcs = [
4386 "bench/f16-igemm.cc",
4387 "bench/conv.h",
4388 "bench/google/conv.h",
4389 "src/xnnpack/AlignedAllocator.h",
4390 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004391 deps = MICROKERNEL_BENCHMARK_DEPS + [
4392 ":indirection",
4393 ":packing",
4394 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004395)
4396
4397xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 name = "f16_gemm_bench",
4399 srcs = [
4400 "bench/f16-gemm.cc",
4401 "bench/gemm.h",
4402 "src/xnnpack/AlignedAllocator.h",
4403 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004404 deps = MICROKERNEL_BENCHMARK_DEPS + [
4405 ":packing",
4406 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004407)
4408
4409xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004410 name = "f16_spmm_bench",
4411 srcs = [
4412 "bench/f16-spmm.cc",
4413 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004414 "src/xnnpack/AlignedAllocator.h",
4415 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004416 deps = MICROKERNEL_BENCHMARK_DEPS,
4417)
4418
4419xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004420 name = "f32_igemm_bench",
4421 srcs = [
4422 "bench/f32-igemm.cc",
4423 "bench/conv.h",
4424 "src/xnnpack/AlignedAllocator.h",
4425 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004426 deps = MICROKERNEL_BENCHMARK_DEPS + [
4427 ":indirection",
4428 ":packing",
4429 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430)
4431
4432xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004433 name = "f16_relu_bench",
4434 srcs = [
4435 "bench/f16-relu.cc",
4436 "src/xnnpack/AlignedAllocator.h",
4437 ] + MICROKERNEL_BENCHMARK_HDRS,
4438 deps = MICROKERNEL_BENCHMARK_DEPS,
4439)
4440
4441xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 name = "f32_conv_hwc_bench",
4443 srcs = [
4444 "bench/f32-conv-hwc.cc",
4445 "bench/dconv.h",
4446 "src/xnnpack/AlignedAllocator.h",
4447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004448 deps = MICROKERNEL_BENCHMARK_DEPS + [
4449 ":packing",
4450 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451)
4452
4453xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004454 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004455 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004456 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004457 "bench/dconv.h",
4458 "src/xnnpack/AlignedAllocator.h",
4459 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004460 deps = MICROKERNEL_BENCHMARK_DEPS + [
4461 ":packing",
4462 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004463)
4464
4465xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004466 name = "f16_dwconv_bench",
4467 srcs = [
4468 "bench/f16-dwconv.cc",
4469 "bench/dwconv.h",
4470 "bench/google/dwconv.h",
4471 "src/xnnpack/AlignedAllocator.h",
4472 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004473 deps = MICROKERNEL_BENCHMARK_DEPS + [
4474 ":indirection",
4475 ":packing",
4476 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004477)
4478
4479xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 name = "f32_dwconv_bench",
4481 srcs = [
4482 "bench/f32-dwconv.cc",
4483 "bench/dwconv.h",
4484 "src/xnnpack/AlignedAllocator.h",
4485 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004486 deps = MICROKERNEL_BENCHMARK_DEPS + [
4487 ":indirection",
4488 ":packing",
4489 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004490)
4491
4492xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004493 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004495 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496 "bench/dwconv.h",
4497 "src/xnnpack/AlignedAllocator.h",
4498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004499 deps = MICROKERNEL_BENCHMARK_DEPS + [
4500 ":indirection",
4501 ":packing",
4502 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503)
4504
4505xnnpack_benchmark(
4506 name = "f32_gemm_bench",
4507 srcs = [
4508 "bench/f32-gemm.cc",
4509 "bench/gemm.h",
4510 "src/xnnpack/AlignedAllocator.h",
4511 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004512 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004513 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004514)
4515
4516xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004517 name = "f32_hswish_bench",
4518 srcs = [
4519 "bench/f32-hswish.cc",
4520 "src/xnnpack/AlignedAllocator.h",
4521 ] + MICROKERNEL_BENCHMARK_HDRS,
4522 deps = MICROKERNEL_BENCHMARK_DEPS,
4523)
4524
4525xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004526 name = "f32_raddexpminusmax_bench",
4527 srcs = [
4528 "bench/f32-raddexpminusmax.cc",
4529 "src/xnnpack/AlignedAllocator.h",
4530 ] + MICROKERNEL_BENCHMARK_HDRS,
4531 deps = MICROKERNEL_BENCHMARK_DEPS,
4532)
4533
4534xnnpack_benchmark(
4535 name = "f32_raddextexp_bench",
4536 srcs = [
4537 "bench/f32-raddextexp.cc",
4538 "src/xnnpack/AlignedAllocator.h",
4539 ] + MICROKERNEL_BENCHMARK_HDRS,
4540 deps = MICROKERNEL_BENCHMARK_DEPS,
4541)
4542
4543xnnpack_benchmark(
4544 name = "f32_raddstoreexpminusmax_bench",
4545 srcs = [
4546 "bench/f32-raddstoreexpminusmax.cc",
4547 "src/xnnpack/AlignedAllocator.h",
4548 ] + MICROKERNEL_BENCHMARK_HDRS,
4549 deps = MICROKERNEL_BENCHMARK_DEPS,
4550)
4551
4552xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004553 name = "f32_relu_bench",
4554 srcs = [
4555 "bench/f32-relu.cc",
4556 "src/xnnpack/AlignedAllocator.h",
4557 ] + MICROKERNEL_BENCHMARK_HDRS,
4558 deps = MICROKERNEL_BENCHMARK_DEPS,
4559)
4560
4561xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004562 name = "f32_rmax_bench",
4563 srcs = [
4564 "bench/f32-rmax.cc",
4565 "src/xnnpack/AlignedAllocator.h",
4566 ] + MICROKERNEL_BENCHMARK_HDRS,
4567 deps = MICROKERNEL_BENCHMARK_DEPS,
4568)
4569
4570xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004571 name = "f32_sigmoid_bench",
4572 srcs = [
4573 "bench/f32-sigmoid.cc",
4574 "src/xnnpack/AlignedAllocator.h",
4575 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004576 deps = MICROKERNEL_BENCHMARK_DEPS,
4577)
4578
4579xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004580 name = "f32_spmm_bench",
4581 srcs = [
4582 "bench/f32-spmm.cc",
4583 "bench/gemm.h",
4584 "src/xnnpack/AlignedAllocator.h",
4585 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004586 deps = MICROKERNEL_BENCHMARK_DEPS,
4587)
4588
4589xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004590 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004591 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004592 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004593 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004594 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004595 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004596)
4597
4598xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004599 name = "f32_vscaleexpminusmax_bench",
4600 srcs = [
4601 "bench/f32-vscaleexpminusmax.cc",
4602 "src/xnnpack/AlignedAllocator.h",
4603 ] + MICROKERNEL_BENCHMARK_HDRS,
4604 deps = MICROKERNEL_BENCHMARK_DEPS,
4605)
4606
4607xnnpack_benchmark(
4608 name = "f32_vscaleextexp_bench",
4609 srcs = [
4610 "bench/f32-vscaleextexp.cc",
4611 "src/xnnpack/AlignedAllocator.h",
4612 ] + MICROKERNEL_BENCHMARK_HDRS,
4613 deps = MICROKERNEL_BENCHMARK_DEPS,
4614)
4615
4616xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004617 name = "f32_vsqrt_bench",
4618 srcs = [
4619 "bench/f32-vsqrt.cc",
4620 "src/xnnpack/AlignedAllocator.h",
4621 ] + MICROKERNEL_BENCHMARK_HDRS,
4622 deps = MICROKERNEL_BENCHMARK_DEPS,
4623)
4624
4625xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 name = "f32_im2col_gemm_bench",
4627 srcs = [
4628 "bench/f32-im2col-gemm.cc",
4629 "bench/conv.h",
4630 "src/xnnpack/AlignedAllocator.h",
4631 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004632 deps = MICROKERNEL_BENCHMARK_DEPS + [
4633 ":im2col",
4634 ":packing",
4635 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004636)
4637
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004638xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004639 name = "rounding_bench",
4640 srcs = [
4641 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004642 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004643 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004644 ] + MICROKERNEL_BENCHMARK_HDRS,
4645 deps = MICROKERNEL_BENCHMARK_DEPS,
4646)
4647
Marat Dukhan08c4a432019-10-03 09:29:21 -07004648########################### Benchmarks for operators ###########################
4649
4650xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004651 name = "average_pooling_bench",
4652 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004653 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004654 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004655 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656)
4657
4658xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004659 name = "bankers_rounding_bench",
4660 srcs = ["bench/bankers-rounding.cc"],
4661 copts = xnnpack_optional_tflite_copts(),
4662 tags = ["nowin32"],
4663 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4664)
4665
4666xnnpack_benchmark(
4667 name = "ceiling_bench",
4668 srcs = ["bench/ceiling.cc"],
4669 copts = xnnpack_optional_tflite_copts(),
4670 tags = ["nowin32"],
4671 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4672)
4673
4674xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675 name = "channel_shuffle_bench",
4676 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004677 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004678)
4679
4680xnnpack_benchmark(
4681 name = "convolution_bench",
4682 srcs = ["bench/convolution.cc"],
4683 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004684 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004685 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004686)
4687
4688xnnpack_benchmark(
4689 name = "deconvolution_bench",
4690 srcs = ["bench/deconvolution.cc"],
4691 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004692 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004693 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004694)
4695
4696xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004697 name = "floor_bench",
4698 srcs = ["bench/floor.cc"],
4699 copts = xnnpack_optional_tflite_copts(),
4700 tags = ["nowin32"],
4701 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4702)
4703
4704xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004705 name = "global_average_pooling_bench",
4706 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004707 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004708)
4709
4710xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004711 name = "hardswish_bench",
4712 srcs = ["bench/hardswish.cc"],
4713 copts = xnnpack_optional_tflite_copts(),
4714 tags = ["nowin32"],
4715 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4716)
4717
4718xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 name = "max_pooling_bench",
4720 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004721 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004722)
4723
4724xnnpack_benchmark(
4725 name = "sigmoid_bench",
4726 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004727 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004728 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004729 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004730)
4731
4732xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004733 name = "prelu_bench",
4734 srcs = ["bench/prelu.cc"],
4735 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004736 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004737 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004738)
4739
4740xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004741 name = "softmax_bench",
4742 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004743 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004744 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004745 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004746)
4747
Marat Dukhan87727142020-06-24 15:24:10 -07004748xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004749 name = "square_root_bench",
4750 srcs = ["bench/square-root.cc"],
4751 copts = xnnpack_optional_tflite_copts(),
4752 tags = ["nowin32"],
4753 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4754)
4755
4756xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004757 name = "truncation_bench",
4758 srcs = ["bench/truncation.cc"],
4759 deps = OPERATOR_BENCHMARK_DEPS,
4760)
4761
Marat Dukhanc068bb62019-10-04 13:24:39 -07004762############################# End-to-end benchmarks ############################
4763
4764cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004765 name = "fp32_mobilenet_v1",
4766 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004767 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004768 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004769 linkstatic = True,
4770 deps = [
4771 ":XNNPACK",
4772 "@pthreadpool",
4773 ],
4774)
4775
4776cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004777 name = "fp16_mobilenet_v1",
4778 srcs = ["models/fp16-mobilenet-v1.cc"],
4779 hdrs = ["models/models.h"],
4780 copts = xnnpack_std_cxxopts(),
4781 linkstatic = True,
4782 deps = [
4783 ":XNNPACK",
4784 "@FP16",
4785 "@pthreadpool",
4786 ],
4787)
4788
4789cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004790 name = "qs8_mobilenet_v1",
4791 srcs = ["models/qs8-mobilenet-v1.cc"],
4792 hdrs = ["models/models.h"],
4793 copts = xnnpack_std_cxxopts(),
4794 linkstatic = True,
4795 deps = [
4796 ":XNNPACK",
4797 "@pthreadpool",
4798 ],
4799)
4800
4801cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004802 name = "qs8_mobilenet_v2",
4803 srcs = ["models/qs8-mobilenet-v2.cc"],
4804 hdrs = ["models/models.h"],
4805 copts = xnnpack_std_cxxopts(),
4806 linkstatic = True,
4807 deps = [
4808 ":XNNPACK",
4809 "@pthreadpool",
4810 ],
4811)
4812
4813cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004814 name = "fp32_mobilenet_v2",
4815 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004816 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004817 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004818 linkstatic = True,
4819 deps = [
4820 ":XNNPACK",
4821 "@pthreadpool",
4822 ],
4823)
4824
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004825cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004826 name = "fp16_mobilenet_v2",
4827 srcs = ["models/fp16-mobilenet-v2.cc"],
4828 hdrs = ["models/models.h"],
4829 copts = xnnpack_std_cxxopts(),
4830 linkstatic = True,
4831 deps = [
4832 ":XNNPACK",
4833 "@FP16",
4834 "@pthreadpool",
4835 ],
4836)
4837
4838cc_library(
4839 name = "fp32_mobilenet_v3_large",
4840 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004841 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004842 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004843 linkstatic = True,
4844 deps = [
4845 ":XNNPACK",
4846 "@pthreadpool",
4847 ],
4848)
4849
4850cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004851 name = "fp16_mobilenet_v3_large",
4852 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4853 hdrs = ["models/models.h"],
4854 copts = xnnpack_std_cxxopts(),
4855 linkstatic = True,
4856 deps = [
4857 ":XNNPACK",
4858 "@FP16",
4859 "@pthreadpool",
4860 ],
4861)
4862
4863cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004864 name = "fp32_mobilenet_v3_small",
4865 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004866 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004867 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004868 linkstatic = True,
4869 deps = [
4870 ":XNNPACK",
4871 "@pthreadpool",
4872 ],
4873)
4874
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004875cc_library(
4876 name = "fp16_mobilenet_v3_small",
4877 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4878 hdrs = ["models/models.h"],
4879 copts = xnnpack_std_cxxopts(),
4880 linkstatic = True,
4881 deps = [
4882 ":XNNPACK",
4883 "@FP16",
4884 "@pthreadpool",
4885 ],
4886)
4887
Marat Dukhanc068bb62019-10-04 13:24:39 -07004888xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004889 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004890 srcs = [
4891 "bench/f32-dwconv-e2e.cc",
4892 "bench/end2end.h",
4893 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004894 deps = MICROKERNEL_BENCHMARK_DEPS + [
4895 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004896 ":fp32_mobilenet_v1",
4897 ":fp32_mobilenet_v2",
4898 ":fp32_mobilenet_v3_large",
4899 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004900 ],
4901)
4902
4903xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004904 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004905 srcs = [
4906 "bench/f32-gemm-e2e.cc",
4907 "bench/end2end.h",
4908 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004909 deps = MICROKERNEL_BENCHMARK_DEPS + [
4910 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004911 ":fp32_mobilenet_v1",
4912 ":fp32_mobilenet_v2",
4913 ":fp32_mobilenet_v3_large",
4914 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004915 ],
4916)
4917
4918xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004919 name = "end2end_bench",
4920 srcs = ["bench/end2end.cc"],
4921 deps = [
4922 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004923 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004924 ":fp16_mobilenet_v1",
4925 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004926 ":fp16_mobilenet_v3_large",
4927 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004928 ":fp32_mobilenet_v1",
4929 ":fp32_mobilenet_v2",
4930 ":fp32_mobilenet_v3_large",
4931 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004932 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004933 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004934 "@pthreadpool",
4935 ],
4936)
4937
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004938#################### Accuracy evaluation for math functions ####################
4939
4940xnnpack_benchmark(
4941 name = "f32_exp_eval",
4942 srcs = [
4943 "eval/f32-exp.cc",
4944 "src/xnnpack/AlignedAllocator.h",
4945 ] + ACCURACY_EVAL_HDRS,
4946 deps = ACCURACY_EVAL_DEPS,
4947)
4948
Marat Dukhan515c9772019-10-17 18:07:57 -07004949xnnpack_benchmark(
4950 name = "f32_expminus_eval",
4951 srcs = [
4952 "eval/f32-expminus.cc",
4953 "src/xnnpack/AlignedAllocator.h",
4954 ] + ACCURACY_EVAL_HDRS,
4955 deps = ACCURACY_EVAL_DEPS,
4956)
4957
Marat Dukhan98ba4412019-10-23 02:14:28 -07004958xnnpack_benchmark(
4959 name = "f32_extexp_eval",
4960 srcs = [
4961 "eval/f32-extexp.cc",
4962 "src/xnnpack/AlignedAllocator.h",
4963 ] + ACCURACY_EVAL_HDRS,
4964 deps = ACCURACY_EVAL_DEPS,
4965)
4966
Marat Dukhan8853b822020-05-07 12:19:01 -07004967xnnpack_unit_test(
4968 name = "f32_roundne_eval",
4969 srcs = [
4970 "eval/f32-roundne.cc",
4971 "src/xnnpack/AlignedAllocator.h",
4972 "src/xnnpack/math-stubs.h",
4973 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004974 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004975 deps = MICROKERNEL_TEST_DEPS,
4976)
4977
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004978xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004979 name = "f32_roundd_eval",
4980 srcs = [
4981 "eval/f32-roundd.cc",
4982 "src/xnnpack/AlignedAllocator.h",
4983 "src/xnnpack/math-stubs.h",
4984 ] + MICROKERNEL_TEST_HDRS,
4985 automatic = False,
4986 deps = MICROKERNEL_TEST_DEPS,
4987)
4988
4989xnnpack_unit_test(
4990 name = "f32_roundu_eval",
4991 srcs = [
4992 "eval/f32-roundu.cc",
4993 "src/xnnpack/AlignedAllocator.h",
4994 "src/xnnpack/math-stubs.h",
4995 ] + MICROKERNEL_TEST_HDRS,
4996 automatic = False,
4997 deps = MICROKERNEL_TEST_DEPS,
4998)
4999
5000xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005001 name = "f32_roundz_eval",
5002 srcs = [
5003 "eval/f32-roundz.cc",
5004 "src/xnnpack/AlignedAllocator.h",
5005 "src/xnnpack/math-stubs.h",
5006 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005007 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005008 deps = MICROKERNEL_TEST_DEPS,
5009)
5010
Marat Dukhan346a9e52019-11-15 09:06:30 -08005011xnnpack_benchmark(
5012 name = "f32_sigmoid_eval",
5013 srcs = [
5014 "eval/f32-sigmoid.cc",
5015 "src/xnnpack/AlignedAllocator.h",
5016 ] + ACCURACY_EVAL_HDRS,
5017 deps = ACCURACY_EVAL_DEPS,
5018)
5019
Marat Dukhan84000762020-06-29 18:38:43 -07005020xnnpack_benchmark(
5021 name = "f32_sqrt_eval",
5022 srcs = [
5023 "eval/f32-sqrt.cc",
5024 "src/xnnpack/AlignedAllocator.h",
5025 ] + ACCURACY_EVAL_HDRS,
5026 deps = ACCURACY_EVAL_DEPS + [
5027 ":bench_utils",
5028 ],
5029)
5030
Marat Dukhan08c4a432019-10-03 09:29:21 -07005031######################### Unit tests for micro-kernels #########################
5032
5033xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005034 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005036 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005037 "test/gemm-microkernel-tester.h",
5038 "src/xnnpack/AlignedAllocator.h",
5039 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005040 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005041)
5042
5043xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005044 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005045 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005046 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005047 "test/spmm-microkernel-tester.h",
5048 "src/xnnpack/AlignedAllocator.h",
5049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054 name = "f32_argmaxpool_test",
5055 srcs = [
5056 "test/f32-argmaxpool.cc",
5057 "test/argmaxpool-microkernel-tester.h",
5058 "src/xnnpack/AlignedAllocator.h",
5059 ] + MICROKERNEL_TEST_HDRS,
5060 deps = MICROKERNEL_TEST_DEPS,
5061)
5062
5063xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005064 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005066 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067 "test/avgpool-microkernel-tester.h",
5068 "src/xnnpack/AlignedAllocator.h",
5069 ] + MICROKERNEL_TEST_HDRS,
5070 deps = MICROKERNEL_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005074 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005075 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005076 "test/f32-ibilinear.cc",
5077 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005078 "src/xnnpack/AlignedAllocator.h",
5079 ] + MICROKERNEL_TEST_HDRS,
5080 deps = MICROKERNEL_TEST_DEPS,
5081)
5082
5083xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005084 name = "f32_ibilinear_chw_test",
5085 srcs = [
5086 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005087 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005088 "src/xnnpack/AlignedAllocator.h",
5089 ] + MICROKERNEL_TEST_HDRS,
5090 deps = MICROKERNEL_TEST_DEPS,
5091)
5092
5093xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005094 name = "f16_clamp_test",
5095 srcs = [
5096 "test/f16-clamp.cc",
5097 "test/clamp-microkernel-tester.h",
5098 ] + MICROKERNEL_TEST_HDRS,
5099 deps = MICROKERNEL_TEST_DEPS,
5100)
5101
5102xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005103 name = "f32_clamp_test",
5104 srcs = [
5105 "test/f32-clamp.cc",
5106 "test/clamp-microkernel-tester.h",
5107 ] + MICROKERNEL_TEST_HDRS,
5108 deps = MICROKERNEL_TEST_DEPS,
5109)
5110
5111xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005112 name = "f32_igemm_test",
5113 srcs = [
5114 "test/f32-igemm.cc",
5115 "test/gemm-microkernel-tester.h",
5116 "src/xnnpack/AlignedAllocator.h",
5117 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005118 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005119)
5120
5121xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005122 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005123 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005124 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 "test/gemm-microkernel-tester.h",
5126 "src/xnnpack/AlignedAllocator.h",
5127 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005128 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129)
5130
5131xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005132 name = "f32_igemm_minmax_test",
5133 srcs = [
5134 "test/f32-igemm-minmax.cc",
5135 "test/gemm-microkernel-tester.h",
5136 "src/xnnpack/AlignedAllocator.h",
5137 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005138 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005139)
5140
5141xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005142 name = "f16_igemm_minmax_test",
5143 srcs = [
5144 "test/f16-igemm-minmax.cc",
5145 "test/gemm-microkernel-tester.h",
5146 "src/xnnpack/AlignedAllocator.h",
5147 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005148 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005149)
5150
5151xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005152 name = "f32_conv_hwc_test",
5153 srcs = [
5154 "test/f32-conv-hwc.cc",
5155 "test/conv-hwc-microkernel-tester.h",
5156 "src/xnnpack/AlignedAllocator.h",
5157 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005158 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159)
5160
5161xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005162 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005164 "test/f32-conv-hwc2chw.cc",
5165 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005166 "src/xnnpack/AlignedAllocator.h",
5167 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005168 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169)
5170
5171xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005172 name = "f32_dwconv_test",
5173 srcs = [
5174 "test/f32-dwconv.cc",
5175 "test/dwconv-microkernel-tester.h",
5176 "src/xnnpack/AlignedAllocator.h",
5177 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005178 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005179)
5180
5181xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005182 name = "f16_dwconv_minmax_test",
5183 srcs = [
5184 "test/f16-dwconv-minmax.cc",
5185 "test/dwconv-microkernel-tester.h",
5186 "src/xnnpack/AlignedAllocator.h",
5187 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005188 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005189)
5190
5191xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005192 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005194 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 "test/dwconv-microkernel-tester.h",
5196 "src/xnnpack/AlignedAllocator.h",
5197 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005198 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199)
5200
5201xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005202 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005204 "test/f32-dwconv2d-chw.cc",
5205 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 "src/xnnpack/AlignedAllocator.h",
5207 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005208 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209)
5210
5211xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005212 name = "f16_gavgpool_minmax_test",
5213 srcs = [
5214 "test/f16-gavgpool-minmax.cc",
5215 "test/gavgpool-microkernel-tester.h",
5216 "src/xnnpack/AlignedAllocator.h",
5217 ] + MICROKERNEL_TEST_HDRS,
5218 deps = MICROKERNEL_TEST_DEPS,
5219)
5220
5221xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005222 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005224 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225 "test/gavgpool-microkernel-tester.h",
5226 "src/xnnpack/AlignedAllocator.h",
5227 ] + MICROKERNEL_TEST_HDRS,
5228 deps = MICROKERNEL_TEST_DEPS,
5229)
5230
5231xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005232 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005234 "test/f32-gavgpool-cw.cc",
5235 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005236 "src/xnnpack/AlignedAllocator.h",
5237 ] + MICROKERNEL_TEST_HDRS,
5238 deps = MICROKERNEL_TEST_DEPS,
5239)
5240
5241xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005242 name = "f32_gemm_test",
5243 srcs = [
5244 "test/f32-gemm.cc",
5245 "test/gemm-microkernel-tester.h",
5246 "src/xnnpack/AlignedAllocator.h",
5247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005248 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005249)
5250
5251xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005252 name = "f32_gemm_relu_test",
5253 srcs = [
5254 "test/f32-gemm-relu.cc",
5255 "test/gemm-microkernel-tester.h",
5256 "src/xnnpack/AlignedAllocator.h",
5257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005258 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005259)
5260
5261xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005262 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005264 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 "test/gemm-microkernel-tester.h",
5266 "src/xnnpack/AlignedAllocator.h",
5267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005268 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005269)
5270
5271xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005272 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005273 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005274 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 "test/gemm-microkernel-tester.h",
5276 "src/xnnpack/AlignedAllocator.h",
5277 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005278 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005279)
5280
5281xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005282 name = "f16_hswish_test",
5283 srcs = [
5284 "test/f16-hswish.cc",
5285 "test/hswish-microkernel-tester.h",
5286 ] + MICROKERNEL_TEST_HDRS,
5287 deps = MICROKERNEL_TEST_DEPS,
5288)
5289
5290xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 name = "f32_hswish_test",
5292 srcs = [
5293 "test/f32-hswish.cc",
5294 "test/hswish-microkernel-tester.h",
5295 ] + MICROKERNEL_TEST_HDRS,
5296 deps = MICROKERNEL_TEST_DEPS,
5297)
5298
5299xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005300 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005301 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005302 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303 "test/maxpool-microkernel-tester.h",
5304 ] + MICROKERNEL_TEST_HDRS,
5305 deps = MICROKERNEL_TEST_DEPS,
5306)
5307
5308xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005309 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005311 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312 "test/avgpool-microkernel-tester.h",
5313 "src/xnnpack/AlignedAllocator.h",
5314 ] + MICROKERNEL_TEST_HDRS,
5315 deps = MICROKERNEL_TEST_DEPS,
5316)
5317
5318xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005319 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005320 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005321 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 "test/gemm-microkernel-tester.h",
5323 "src/xnnpack/AlignedAllocator.h",
5324 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005325 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005326)
5327
5328xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005329 name = "f16_prelu_test",
5330 srcs = [
5331 "test/f16-prelu.cc",
5332 "test/prelu-microkernel-tester.h",
5333 "src/xnnpack/AlignedAllocator.h",
5334 ] + MICROKERNEL_TEST_HDRS,
5335 deps = MICROKERNEL_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339 name = "f32_prelu_test",
5340 srcs = [
5341 "test/f32-prelu.cc",
5342 "test/prelu-microkernel-tester.h",
5343 "src/xnnpack/AlignedAllocator.h",
5344 ] + MICROKERNEL_TEST_HDRS,
5345 deps = MICROKERNEL_TEST_DEPS,
5346)
5347
5348xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005349 name = "f32_raddexpminusmax_test",
5350 srcs = [
5351 "test/f32-raddexpminusmax.cc",
5352 "test/raddexpminusmax-microkernel-tester.h",
5353 ] + MICROKERNEL_TEST_HDRS,
5354 deps = MICROKERNEL_TEST_DEPS,
5355)
5356
5357xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005358 name = "f32_raddextexp_test",
5359 srcs = [
5360 "test/f32-raddextexp.cc",
5361 "test/raddextexp-microkernel-tester.h",
5362 ] + MICROKERNEL_TEST_HDRS,
5363 deps = MICROKERNEL_TEST_DEPS,
5364)
5365
5366xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005367 name = "f32_raddstoreexpminusmax_test",
5368 srcs = [
5369 "test/f32-raddstoreexpminusmax.cc",
5370 "test/raddstoreexpminusmax-microkernel-tester.h",
5371 ] + MICROKERNEL_TEST_HDRS,
5372 deps = MICROKERNEL_TEST_DEPS,
5373)
5374
5375xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005376 name = "f32_relu_test",
5377 srcs = [
5378 "test/f32-relu.cc",
5379 "test/vunary-microkernel-tester.h",
5380 ] + MICROKERNEL_TEST_HDRS,
5381 deps = MICROKERNEL_TEST_DEPS,
5382)
5383
5384xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005385 name = "f32_rmax_test",
5386 srcs = [
5387 "test/f32-rmax.cc",
5388 "test/rmax-microkernel-tester.h",
5389 ] + MICROKERNEL_TEST_HDRS,
5390 deps = MICROKERNEL_TEST_DEPS,
5391)
5392
5393xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005394 name = "f32_sigmoid_test",
5395 srcs = [
5396 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005397 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005398 ] + MICROKERNEL_TEST_HDRS,
5399 deps = MICROKERNEL_TEST_DEPS,
5400)
5401
5402xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005403 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005405 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 "test/spmm-microkernel-tester.h",
5407 "src/xnnpack/AlignedAllocator.h",
5408 ] + MICROKERNEL_TEST_HDRS,
5409 deps = MICROKERNEL_TEST_DEPS,
5410)
5411
5412xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005413 name = "f16_vadd_minmax_test",
5414 srcs = [
5415 "test/f16-vadd-minmax.cc",
5416 "test/vbinary-microkernel-tester.h",
5417 ] + MICROKERNEL_TEST_HDRS,
5418 deps = MICROKERNEL_TEST_DEPS,
5419)
5420
5421xnnpack_unit_test(
5422 name = "f16_vaddc_minmax_test",
5423 srcs = [
5424 "test/f16-vaddc-minmax.cc",
5425 "test/vbinaryc-microkernel-tester.h",
5426 ] + MICROKERNEL_TEST_HDRS,
5427 deps = MICROKERNEL_TEST_DEPS,
5428)
5429
5430xnnpack_unit_test(
5431 name = "f16_vdiv_minmax_test",
5432 srcs = [
5433 "test/f16-vdiv-minmax.cc",
5434 "test/vbinary-microkernel-tester.h",
5435 ] + MICROKERNEL_TEST_HDRS,
5436 deps = MICROKERNEL_TEST_DEPS,
5437)
5438
5439xnnpack_unit_test(
5440 name = "f16_vdivc_minmax_test",
5441 srcs = [
5442 "test/f16-vdivc-minmax.cc",
5443 "test/vbinaryc-microkernel-tester.h",
5444 ] + MICROKERNEL_TEST_HDRS,
5445 deps = MICROKERNEL_TEST_DEPS,
5446)
5447
5448xnnpack_unit_test(
5449 name = "f16_vrdivc_minmax_test",
5450 srcs = [
5451 "test/f16-vrdivc-minmax.cc",
5452 "test/vbinaryc-microkernel-tester.h",
5453 ] + MICROKERNEL_TEST_HDRS,
5454 deps = MICROKERNEL_TEST_DEPS,
5455)
5456
5457xnnpack_unit_test(
5458 name = "f16_vmax_test",
5459 srcs = [
5460 "test/f16-vmax.cc",
5461 "test/vbinary-microkernel-tester.h",
5462 ] + MICROKERNEL_TEST_HDRS,
5463 deps = MICROKERNEL_TEST_DEPS,
5464)
5465
5466xnnpack_unit_test(
5467 name = "f16_vmaxc_test",
5468 srcs = [
5469 "test/f16-vmaxc.cc",
5470 "test/vbinaryc-microkernel-tester.h",
5471 ] + MICROKERNEL_TEST_HDRS,
5472 deps = MICROKERNEL_TEST_DEPS,
5473)
5474
5475xnnpack_unit_test(
5476 name = "f16_vmin_test",
5477 srcs = [
5478 "test/f16-vmin.cc",
5479 "test/vbinary-microkernel-tester.h",
5480 ] + MICROKERNEL_TEST_HDRS,
5481 deps = MICROKERNEL_TEST_DEPS,
5482)
5483
5484xnnpack_unit_test(
5485 name = "f16_vminc_test",
5486 srcs = [
5487 "test/f16-vminc.cc",
5488 "test/vbinaryc-microkernel-tester.h",
5489 ] + MICROKERNEL_TEST_HDRS,
5490 deps = MICROKERNEL_TEST_DEPS,
5491)
5492
5493xnnpack_unit_test(
5494 name = "f16_vmul_minmax_test",
5495 srcs = [
5496 "test/f16-vmul-minmax.cc",
5497 "test/vbinary-microkernel-tester.h",
5498 ] + MICROKERNEL_TEST_HDRS,
5499 deps = MICROKERNEL_TEST_DEPS,
5500)
5501
5502xnnpack_unit_test(
5503 name = "f16_vmulc_minmax_test",
5504 srcs = [
5505 "test/f16-vmulc-minmax.cc",
5506 "test/vbinaryc-microkernel-tester.h",
5507 ] + MICROKERNEL_TEST_HDRS,
5508 deps = MICROKERNEL_TEST_DEPS,
5509)
5510
5511xnnpack_unit_test(
5512 name = "f16_vsub_minmax_test",
5513 srcs = [
5514 "test/f16-vsub-minmax.cc",
5515 "test/vbinary-microkernel-tester.h",
5516 ] + MICROKERNEL_TEST_HDRS,
5517 deps = MICROKERNEL_TEST_DEPS,
5518)
5519
5520xnnpack_unit_test(
5521 name = "f16_vsubc_minmax_test",
5522 srcs = [
5523 "test/f16-vsubc-minmax.cc",
5524 "test/vbinaryc-microkernel-tester.h",
5525 ] + MICROKERNEL_TEST_HDRS,
5526 deps = MICROKERNEL_TEST_DEPS,
5527)
5528
5529xnnpack_unit_test(
5530 name = "f16_vrsubc_minmax_test",
5531 srcs = [
5532 "test/f16-vrsubc-minmax.cc",
5533 "test/vbinaryc-microkernel-tester.h",
5534 ] + MICROKERNEL_TEST_HDRS,
5535 deps = MICROKERNEL_TEST_DEPS,
5536)
5537
5538xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005539 name = "f32_vabs_test",
5540 srcs = [
5541 "test/f32-vabs.cc",
5542 "test/vunary-microkernel-tester.h",
5543 ] + MICROKERNEL_TEST_HDRS,
5544 deps = MICROKERNEL_TEST_DEPS,
5545)
5546
5547xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005548 name = "f32_vadd_test",
5549 srcs = [
5550 "test/f32-vadd.cc",
5551 "test/vbinary-microkernel-tester.h",
5552 ] + MICROKERNEL_TEST_HDRS,
5553 deps = MICROKERNEL_TEST_DEPS,
5554)
5555
5556xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005557 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005559 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005560 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005561 ] + MICROKERNEL_TEST_HDRS,
5562 deps = MICROKERNEL_TEST_DEPS,
5563)
5564
5565xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005566 name = "f32_vadd_relu_test",
5567 srcs = [
5568 "test/f32-vadd-relu.cc",
5569 "test/vbinary-microkernel-tester.h",
5570 ] + MICROKERNEL_TEST_HDRS,
5571 deps = MICROKERNEL_TEST_DEPS,
5572)
5573
5574xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005575 name = "f32_vaddc_test",
5576 srcs = [
5577 "test/f32-vaddc.cc",
5578 "test/vbinaryc-microkernel-tester.h",
5579 ] + MICROKERNEL_TEST_HDRS,
5580 deps = MICROKERNEL_TEST_DEPS,
5581)
5582
5583xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005584 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005585 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005586 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005587 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005588 ] + MICROKERNEL_TEST_HDRS,
5589 deps = MICROKERNEL_TEST_DEPS,
5590)
5591
5592xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005593 name = "f32_vaddc_relu_test",
5594 srcs = [
5595 "test/f32-vaddc-relu.cc",
5596 "test/vbinaryc-microkernel-tester.h",
5597 ] + MICROKERNEL_TEST_HDRS,
5598 deps = MICROKERNEL_TEST_DEPS,
5599)
5600
5601xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005602 name = "f32_vdiv_test",
5603 srcs = [
5604 "test/f32-vdiv.cc",
5605 "test/vbinary-microkernel-tester.h",
5606 ] + MICROKERNEL_TEST_HDRS,
5607 deps = MICROKERNEL_TEST_DEPS,
5608)
5609
5610xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005611 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005612 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005613 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005614 "test/vbinary-microkernel-tester.h",
5615 ] + MICROKERNEL_TEST_HDRS,
5616 deps = MICROKERNEL_TEST_DEPS,
5617)
5618
5619xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005620 name = "f32_vdiv_relu_test",
5621 srcs = [
5622 "test/f32-vdiv-relu.cc",
5623 "test/vbinary-microkernel-tester.h",
5624 ] + MICROKERNEL_TEST_HDRS,
5625 deps = MICROKERNEL_TEST_DEPS,
5626)
5627
5628xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005629 name = "f32_vdivc_test",
5630 srcs = [
5631 "test/f32-vdivc.cc",
5632 "test/vbinaryc-microkernel-tester.h",
5633 ] + MICROKERNEL_TEST_HDRS,
5634 deps = MICROKERNEL_TEST_DEPS,
5635)
5636
5637xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005638 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005639 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005640 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005641 "test/vbinaryc-microkernel-tester.h",
5642 ] + MICROKERNEL_TEST_HDRS,
5643 deps = MICROKERNEL_TEST_DEPS,
5644)
5645
5646xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005647 name = "f32_vdivc_relu_test",
5648 srcs = [
5649 "test/f32-vdivc-relu.cc",
5650 "test/vbinaryc-microkernel-tester.h",
5651 ] + MICROKERNEL_TEST_HDRS,
5652 deps = MICROKERNEL_TEST_DEPS,
5653)
5654
5655xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005656 name = "f32_vrdivc_test",
5657 srcs = [
5658 "test/f32-vrdivc.cc",
5659 "test/vbinaryc-microkernel-tester.h",
5660 ] + MICROKERNEL_TEST_HDRS,
5661 deps = MICROKERNEL_TEST_DEPS,
5662)
5663
5664xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005665 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005666 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005667 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005668 "test/vbinaryc-microkernel-tester.h",
5669 ] + MICROKERNEL_TEST_HDRS,
5670 deps = MICROKERNEL_TEST_DEPS,
5671)
5672
5673xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005674 name = "f32_vrdivc_relu_test",
5675 srcs = [
5676 "test/f32-vrdivc-relu.cc",
5677 "test/vbinaryc-microkernel-tester.h",
5678 ] + MICROKERNEL_TEST_HDRS,
5679 deps = MICROKERNEL_TEST_DEPS,
5680)
5681
5682xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005683 name = "f32_vmax_test",
5684 srcs = [
5685 "test/f32-vmax.cc",
5686 "test/vbinary-microkernel-tester.h",
5687 ] + MICROKERNEL_TEST_HDRS,
5688 deps = MICROKERNEL_TEST_DEPS,
5689)
5690
5691xnnpack_unit_test(
5692 name = "f32_vmaxc_test",
5693 srcs = [
5694 "test/f32-vmaxc.cc",
5695 "test/vbinaryc-microkernel-tester.h",
5696 ] + MICROKERNEL_TEST_HDRS,
5697 deps = MICROKERNEL_TEST_DEPS,
5698)
5699
5700xnnpack_unit_test(
5701 name = "f32_vmin_test",
5702 srcs = [
5703 "test/f32-vmin.cc",
5704 "test/vbinary-microkernel-tester.h",
5705 ] + MICROKERNEL_TEST_HDRS,
5706 deps = MICROKERNEL_TEST_DEPS,
5707)
5708
5709xnnpack_unit_test(
5710 name = "f32_vminc_test",
5711 srcs = [
5712 "test/f32-vminc.cc",
5713 "test/vbinaryc-microkernel-tester.h",
5714 ] + MICROKERNEL_TEST_HDRS,
5715 deps = MICROKERNEL_TEST_DEPS,
5716)
5717
5718xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005719 name = "f32_vmul_test",
5720 srcs = [
5721 "test/f32-vmul.cc",
5722 "test/vbinary-microkernel-tester.h",
5723 ] + MICROKERNEL_TEST_HDRS,
5724 deps = MICROKERNEL_TEST_DEPS,
5725)
5726
5727xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005728 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005729 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005730 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005731 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005732 ] + MICROKERNEL_TEST_HDRS,
5733 deps = MICROKERNEL_TEST_DEPS,
5734)
5735
5736xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005737 name = "f32_vmul_relu_test",
5738 srcs = [
5739 "test/f32-vmul-relu.cc",
5740 "test/vbinary-microkernel-tester.h",
5741 ] + MICROKERNEL_TEST_HDRS,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005746 name = "f32_vmulc_test",
5747 srcs = [
5748 "test/f32-vmulc.cc",
5749 "test/vbinaryc-microkernel-tester.h",
5750 ] + MICROKERNEL_TEST_HDRS,
5751 deps = MICROKERNEL_TEST_DEPS,
5752)
5753
5754xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005755 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005756 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005757 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005758 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005759 ] + MICROKERNEL_TEST_HDRS,
5760 deps = MICROKERNEL_TEST_DEPS,
5761)
5762
5763xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005764 name = "f32_vmulc_relu_test",
5765 srcs = [
5766 "test/f32-vmulc-relu.cc",
5767 "test/vbinaryc-microkernel-tester.h",
5768 ] + MICROKERNEL_TEST_HDRS,
5769 deps = MICROKERNEL_TEST_DEPS,
5770)
5771
5772xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005773 name = "f16_vmulcaddc_minmax_test",
5774 srcs = [
5775 "test/f16-vmulcaddc-minmax.cc",
5776 "test/vmulcaddc-microkernel-tester.h",
5777 "src/xnnpack/AlignedAllocator.h",
5778 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005779 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005780)
5781
5782xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005783 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005784 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005785 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005786 "test/vmulcaddc-microkernel-tester.h",
5787 "src/xnnpack/AlignedAllocator.h",
5788 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005789 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005790)
5791
5792xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005793 name = "f32_vlrelu_test",
5794 srcs = [
5795 "test/f32-vlrelu.cc",
5796 "test/vunary-microkernel-tester.h",
5797 ] + MICROKERNEL_TEST_HDRS,
5798 deps = MICROKERNEL_TEST_DEPS,
5799)
5800
5801xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005802 name = "f32_vneg_test",
5803 srcs = [
5804 "test/f32-vneg.cc",
5805 "test/vunary-microkernel-tester.h",
5806 ] + MICROKERNEL_TEST_HDRS,
5807 deps = MICROKERNEL_TEST_DEPS,
5808)
5809
5810xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005811 name = "f32_vrndne_test",
5812 srcs = [
5813 "test/f32-vrndne.cc",
5814 "test/vunary-microkernel-tester.h",
5815 ] + MICROKERNEL_TEST_HDRS,
5816 deps = MICROKERNEL_TEST_DEPS,
5817)
5818
5819xnnpack_unit_test(
5820 name = "f32_vrndz_test",
5821 srcs = [
5822 "test/f32-vrndz.cc",
5823 "test/vunary-microkernel-tester.h",
5824 ] + MICROKERNEL_TEST_HDRS,
5825 deps = MICROKERNEL_TEST_DEPS,
5826)
5827
5828xnnpack_unit_test(
5829 name = "f32_vrndu_test",
5830 srcs = [
5831 "test/f32-vrndu.cc",
5832 "test/vunary-microkernel-tester.h",
5833 ] + MICROKERNEL_TEST_HDRS,
5834 deps = MICROKERNEL_TEST_DEPS,
5835)
5836
5837xnnpack_unit_test(
5838 name = "f32_vrndd_test",
5839 srcs = [
5840 "test/f32-vrndd.cc",
5841 "test/vunary-microkernel-tester.h",
5842 ] + MICROKERNEL_TEST_HDRS,
5843 deps = MICROKERNEL_TEST_DEPS,
5844)
5845
5846xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005847 name = "f32_vscale_test",
5848 srcs = [
5849 "test/f32-vscale.cc",
5850 "test/vscale-microkernel-tester.h",
5851 ] + MICROKERNEL_TEST_HDRS,
5852 deps = MICROKERNEL_TEST_DEPS,
5853)
5854
5855xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005856 name = "f32_vscaleexpminusmax_test",
5857 srcs = [
5858 "test/f32-vscaleexpminusmax.cc",
5859 "test/vscaleexpminusmax-microkernel-tester.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005865 name = "f32_vscaleextexp_test",
5866 srcs = [
5867 "test/f32-vscaleextexp.cc",
5868 "test/vscaleextexp-microkernel-tester.h",
5869 ] + MICROKERNEL_TEST_HDRS,
5870 deps = MICROKERNEL_TEST_DEPS,
5871)
5872
5873xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005874 name = "f32_vsqr_test",
5875 srcs = [
5876 "test/f32-vsqr.cc",
5877 "test/vunary-microkernel-tester.h",
5878 ] + MICROKERNEL_TEST_HDRS,
5879 deps = MICROKERNEL_TEST_DEPS,
5880)
5881
5882xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005883 name = "f32_vsqrdiff_test",
5884 srcs = [
5885 "test/f32-vsqrdiff.cc",
5886 "test/vbinary-microkernel-tester.h",
5887 ] + MICROKERNEL_TEST_HDRS,
5888 deps = MICROKERNEL_TEST_DEPS,
5889)
5890
5891xnnpack_unit_test(
5892 name = "f32_vsqrdiffc_test",
5893 srcs = [
5894 "test/f32-vsqrdiffc.cc",
5895 "test/vbinaryc-microkernel-tester.h",
5896 ] + MICROKERNEL_TEST_HDRS,
5897 deps = MICROKERNEL_TEST_DEPS,
5898)
5899
5900xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005901 name = "f32_vsqrt_test",
5902 srcs = [
5903 "test/f32-vsqrt.cc",
5904 "test/vunary-microkernel-tester.h",
5905 ] + MICROKERNEL_TEST_HDRS,
5906 deps = MICROKERNEL_TEST_DEPS,
5907)
5908
5909xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005910 name = "f32_vsub_test",
5911 srcs = [
5912 "test/f32-vsub.cc",
5913 "test/vbinary-microkernel-tester.h",
5914 ] + MICROKERNEL_TEST_HDRS,
5915 deps = MICROKERNEL_TEST_DEPS,
5916)
5917
5918xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005919 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005920 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005921 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005922 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005923 ] + MICROKERNEL_TEST_HDRS,
5924 deps = MICROKERNEL_TEST_DEPS,
5925)
5926
5927xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005928 name = "f32_vsub_relu_test",
5929 srcs = [
5930 "test/f32-vsub-relu.cc",
5931 "test/vbinary-microkernel-tester.h",
5932 ] + MICROKERNEL_TEST_HDRS,
5933 deps = MICROKERNEL_TEST_DEPS,
5934)
5935
5936xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005937 name = "f32_vsubc_test",
5938 srcs = [
5939 "test/f32-vsubc.cc",
5940 "test/vbinaryc-microkernel-tester.h",
5941 ] + MICROKERNEL_TEST_HDRS,
5942 deps = MICROKERNEL_TEST_DEPS,
5943)
5944
5945xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005946 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005947 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005948 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005949 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005950 ] + MICROKERNEL_TEST_HDRS,
5951 deps = MICROKERNEL_TEST_DEPS,
5952)
5953
5954xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005955 name = "f32_vsubc_relu_test",
5956 srcs = [
5957 "test/f32-vsubc-relu.cc",
5958 "test/vbinaryc-microkernel-tester.h",
5959 ] + MICROKERNEL_TEST_HDRS,
5960 deps = MICROKERNEL_TEST_DEPS,
5961)
5962
5963xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005964 name = "f32_vrsubc_test",
5965 srcs = [
5966 "test/f32-vrsubc.cc",
5967 "test/vbinaryc-microkernel-tester.h",
5968 ] + MICROKERNEL_TEST_HDRS,
5969 deps = MICROKERNEL_TEST_DEPS,
5970)
5971
5972xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005973 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005974 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005975 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005976 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005977 ] + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS,
5979)
5980
5981xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005982 name = "f32_vrsubc_relu_test",
5983 srcs = [
5984 "test/f32-vrsubc-relu.cc",
5985 "test/vbinaryc-microkernel-tester.h",
5986 ] + MICROKERNEL_TEST_HDRS,
5987 deps = MICROKERNEL_TEST_DEPS,
5988)
5989
5990xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005991 name = "qs8_dwconv_minmax_test",
5992 srcs = [
5993 "test/qs8-dwconv-minmax.cc",
5994 "test/dwconv-microkernel-tester.h",
5995 "src/xnnpack/AlignedAllocator.h",
5996 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5997 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5998)
5999
6000xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006001 name = "qs8_gavgpool_minmax_test",
6002 srcs = [
6003 "test/qs8-gavgpool-minmax.cc",
6004 "test/gavgpool-microkernel-tester.h",
6005 "src/xnnpack/AlignedAllocator.h",
6006 ] + MICROKERNEL_TEST_HDRS,
6007 deps = MICROKERNEL_TEST_DEPS,
6008)
6009
6010xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006011 name = "qs8_gemm_minmax_test",
6012 srcs = [
6013 "test/qs8-gemm-minmax.cc",
6014 "test/gemm-microkernel-tester.h",
6015 "src/xnnpack/AlignedAllocator.h",
6016 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6017 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6018)
6019
6020xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006021 name = "qs8_igemm_minmax_test",
6022 srcs = [
6023 "test/qs8-igemm-minmax.cc",
6024 "test/gemm-microkernel-tester.h",
6025 "src/xnnpack/AlignedAllocator.h",
6026 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6027 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6028)
6029
6030xnnpack_unit_test(
6031 name = "qs8_requantization_test",
6032 srcs = [
6033 "src/xnnpack/requantization-stubs.h",
6034 "test/qs8-requantization.cc",
6035 "test/requantization-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006041 name = "qs8_vadd_minmax_test",
6042 srcs = [
6043 "test/qs8-vadd-minmax.cc",
6044 "test/vadd-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006050 name = "qs8_vaddc_minmax_test",
6051 srcs = [
6052 "test/qs8-vaddc-minmax.cc",
6053 "test/vaddc-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006059 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006060 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006061 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006062 "test/avgpool-microkernel-tester.h",
6063 "src/xnnpack/AlignedAllocator.h",
6064 ] + MICROKERNEL_TEST_HDRS,
6065 deps = MICROKERNEL_TEST_DEPS,
6066)
6067
6068xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006069 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006070 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006071 "test/qu8-dwconv-minmax.cc",
6072 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006073 "src/xnnpack/AlignedAllocator.h",
6074 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006075 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006076)
6077
6078xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006079 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006080 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006081 "test/qu8-igemm-minmax.cc",
6082 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006083 "src/xnnpack/AlignedAllocator.h",
6084 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006085 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006086)
6087
6088xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006089 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006090 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006091 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006092 "test/gavgpool-microkernel-tester.h",
6093 "src/xnnpack/AlignedAllocator.h",
6094 ] + MICROKERNEL_TEST_HDRS,
6095 deps = MICROKERNEL_TEST_DEPS,
6096)
6097
6098xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006099 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006100 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006101 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006102 "test/gemm-microkernel-tester.h",
6103 "src/xnnpack/AlignedAllocator.h",
6104 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006105 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006106)
6107
6108xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006109 name = "qu8_requantization_test",
6110 srcs = [
6111 "src/xnnpack/requantization-stubs.h",
6112 "test/qu8-requantization.cc",
6113 "test/requantization-tester.h",
6114 ] + MICROKERNEL_TEST_HDRS,
6115 deps = MICROKERNEL_TEST_DEPS,
6116)
6117
6118xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006119 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006120 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006121 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006122 "test/vadd-microkernel-tester.h",
6123 ] + MICROKERNEL_TEST_HDRS,
6124 deps = MICROKERNEL_TEST_DEPS,
6125)
6126
6127xnnpack_unit_test(
6128 name = "u8_clamp_test",
6129 srcs = [
6130 "test/u8-clamp.cc",
6131 "test/clamp-microkernel-tester.h",
6132 ] + MICROKERNEL_TEST_HDRS,
6133 deps = MICROKERNEL_TEST_DEPS,
6134)
6135
6136xnnpack_unit_test(
6137 name = "u8_lut32norm_test",
6138 srcs = [
6139 "test/u8-lut32norm.cc",
6140 "test/lut-norm-microkernel-tester.h",
6141 ] + MICROKERNEL_TEST_HDRS,
6142 deps = MICROKERNEL_TEST_DEPS,
6143)
6144
6145xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006146 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006147 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006148 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006149 "test/maxpool-microkernel-tester.h",
6150 ] + MICROKERNEL_TEST_HDRS,
6151 deps = MICROKERNEL_TEST_DEPS,
6152)
6153
6154xnnpack_unit_test(
6155 name = "u8_rmax_test",
6156 srcs = [
6157 "test/u8-rmax.cc",
6158 "test/rmax-microkernel-tester.h",
6159 ] + MICROKERNEL_TEST_HDRS,
6160 deps = MICROKERNEL_TEST_DEPS,
6161)
6162
6163xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006164 name = "x32_fill_test",
6165 srcs = [
6166 "test/x32-fill.cc",
6167 "test/fill-microkernel-tester.h",
6168 ] + MICROKERNEL_TEST_HDRS,
6169 deps = MICROKERNEL_TEST_DEPS,
6170)
6171
6172xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006173 name = "x32_packx_test",
6174 srcs = [
6175 "test/x32-packx.cc",
6176 "test/pack-microkernel-tester.h",
6177 "src/xnnpack/AlignedAllocator.h",
6178 ] + MICROKERNEL_TEST_HDRS,
6179 deps = MICROKERNEL_TEST_DEPS,
6180)
6181
6182xnnpack_unit_test(
6183 name = "x32_pad_test",
6184 srcs = [
6185 "test/x32-pad.cc",
6186 "test/pad-microkernel-tester.h",
6187 ] + MICROKERNEL_TEST_HDRS,
6188 deps = MICROKERNEL_TEST_DEPS,
6189)
6190
6191xnnpack_unit_test(
6192 name = "x32_unpool_test",
6193 srcs = [
6194 "test/x32-unpool.cc",
6195 "test/unpool-microkernel-tester.h",
6196 ] + MICROKERNEL_TEST_HDRS,
6197 deps = MICROKERNEL_TEST_DEPS,
6198)
6199
6200xnnpack_unit_test(
6201 name = "x32_zip_test",
6202 srcs = [
6203 "test/x32-zip.cc",
6204 "test/zip-microkernel-tester.h",
6205 ] + MICROKERNEL_TEST_HDRS,
6206 deps = MICROKERNEL_TEST_DEPS,
6207)
6208
6209xnnpack_unit_test(
6210 name = "x8_lut_test",
6211 srcs = [
6212 "test/x8-lut.cc",
6213 "test/lut-microkernel-tester.h",
6214 ] + MICROKERNEL_TEST_HDRS,
6215 deps = MICROKERNEL_TEST_DEPS,
6216)
6217
6218xnnpack_unit_test(
6219 name = "x8_zip_test",
6220 srcs = [
6221 "test/x8-zip.cc",
6222 "test/zip-microkernel-tester.h",
6223 ] + MICROKERNEL_TEST_HDRS,
6224 deps = MICROKERNEL_TEST_DEPS,
6225)
6226
Marat Dukhan20c3b922020-03-10 03:45:06 -07006227########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006228
6229xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006230 name = "operator_size_test",
6231 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006232 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233)
6234
Marat Dukhan20c3b922020-03-10 03:45:06 -07006235xnnpack_binary(
6236 name = "subgraph_size_test",
6237 srcs = ["test/subgraph-size.c"],
6238 deps = [":XNNPACK"],
6239)
6240
6241########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242
6243xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006244 name = "abs_nc_test",
6245 srcs = [
6246 "test/abs-nc.cc",
6247 "test/abs-operator-tester.h",
6248 ],
6249 deps = OPERATOR_TEST_DEPS,
6250)
6251
6252xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006253 name = "add_nd_test",
6254 srcs = [
6255 "test/add-nd.cc",
6256 "test/binary-elementwise-operator-tester.h",
6257 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006258 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006259)
6260
6261xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006262 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006263 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006264 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006265 "test/argmax-pooling-operator-tester.h",
6266 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006267 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006268)
6269
6270xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006271 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006273 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006274 "test/average-pooling-operator-tester.h",
6275 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006276 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006277)
6278
6279xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006280 name = "bankers_rounding_nc_test",
6281 srcs = [
6282 "test/bankers-rounding-nc.cc",
6283 "test/bankers-rounding-operator-tester.h",
6284 ],
6285 deps = OPERATOR_TEST_DEPS,
6286)
6287
6288xnnpack_unit_test(
6289 name = "ceiling_nc_test",
6290 srcs = [
6291 "test/ceiling-nc.cc",
6292 "test/ceiling-operator-tester.h",
6293 ],
6294 deps = OPERATOR_TEST_DEPS,
6295)
6296
6297xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006298 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006300 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006301 "test/channel-shuffle-operator-tester.h",
6302 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006303 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006304)
6305
6306xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006307 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006309 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006310 "test/clamp-operator-tester.h",
6311 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006312 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006313)
6314
6315xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006316 name = "constant_pad_nd_test",
6317 srcs = [
6318 "test/constant-pad-nd.cc",
6319 "test/constant-pad-operator-tester.h",
6320 ],
6321 deps = OPERATOR_TEST_DEPS,
6322)
6323
6324xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006325 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006326 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006327 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006328 "test/convolution-operator-tester.h",
6329 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006330 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006331)
6332
6333xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006334 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006335 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006336 "test/convolution-nchw.cc",
6337 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006338 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006339 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006340)
6341
6342xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006343 name = "copy_nc_test",
6344 srcs = [
6345 "test/copy-nc.cc",
6346 "test/copy-operator-tester.h",
6347 ],
6348 deps = OPERATOR_TEST_DEPS,
6349)
6350
6351xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006352 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006353 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006354 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006355 "test/deconvolution-operator-tester.h",
6356 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006357 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006358)
6359
6360xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006361 name = "divide_nd_test",
6362 srcs = [
6363 "test/binary-elementwise-operator-tester.h",
6364 "test/divide-nd.cc",
6365 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006366 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006367)
6368
6369xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006370 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006371 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006372 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006373 "test/fully-connected-operator-tester.h",
6374 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006375 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006376)
6377
6378xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006379 name = "floor_nc_test",
6380 srcs = [
6381 "test/floor-nc.cc",
6382 "test/floor-operator-tester.h",
6383 ],
6384 deps = OPERATOR_TEST_DEPS,
6385)
6386
6387xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006388 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006389 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006390 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006391 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006392 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006393 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006394)
6395
6396xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006397 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006399 "test/global-average-pooling-ncw.cc",
6400 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006401 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006402 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006403)
6404
6405xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006406 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006408 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006409 "test/hardswish-operator-tester.h",
6410 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006411 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006412)
6413
6414xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006415 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006416 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006417 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006418 "test/leaky-relu-operator-tester.h",
6419 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006420 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006421)
6422
6423xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006424 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006425 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006426 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006427 "test/max-pooling-operator-tester.h",
6428 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006429 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006430)
6431
6432xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006433 name = "maximum_nd_test",
6434 srcs = [
6435 "test/binary-elementwise-operator-tester.h",
6436 "test/maximum-nd.cc",
6437 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006438 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006439)
6440
6441xnnpack_unit_test(
6442 name = "minimum_nd_test",
6443 srcs = [
6444 "test/binary-elementwise-operator-tester.h",
6445 "test/minimum-nd.cc",
6446 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006447 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006448)
6449
6450xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006451 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006452 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006453 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006454 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006455 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006456 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006457)
6458
6459xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006460 name = "negate_nc_test",
6461 srcs = [
6462 "test/negate-nc.cc",
6463 "test/negate-operator-tester.h",
6464 ],
6465 deps = OPERATOR_TEST_DEPS,
6466)
6467
6468xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006469 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006470 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006471 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006472 "test/prelu-operator-tester.h",
6473 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006474 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475)
6476
6477xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006478 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006479 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006480 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006481 "test/resize-bilinear-operator-tester.h",
6482 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006483 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006484)
6485
6486xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006487 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006488 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006489 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006490 "test/sigmoid-operator-tester.h",
6491 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006492 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006493)
6494
6495xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006496 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006497 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006498 "test/softmax-nc.cc",
6499 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006500 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006501 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502)
6503
6504xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006505 name = "square_nc_test",
6506 srcs = [
6507 "test/square-nc.cc",
6508 "test/square-operator-tester.h",
6509 ],
6510 deps = OPERATOR_TEST_DEPS,
6511)
6512
6513xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006514 name = "square_root_nc_test",
6515 srcs = [
6516 "test/square-root-nc.cc",
6517 "test/square-root-operator-tester.h",
6518 ],
6519 deps = OPERATOR_TEST_DEPS,
6520)
6521
6522xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006523 name = "squared_difference_nd_test",
6524 srcs = [
6525 "test/binary-elementwise-operator-tester.h",
6526 "test/squared-difference-nd.cc",
6527 ],
6528 deps = OPERATOR_TEST_DEPS,
6529)
6530
6531xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006532 name = "subtract_nd_test",
6533 srcs = [
6534 "test/binary-elementwise-operator-tester.h",
6535 "test/subtract-nd.cc",
6536 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006537 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006538)
6539
6540xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006541 name = "truncation_nc_test",
6542 srcs = [
6543 "test/truncation-nc.cc",
6544 "test/truncation-operator-tester.h",
6545 ],
6546 deps = OPERATOR_TEST_DEPS,
6547)
6548
6549xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006550 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006551 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006552 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006553 "test/unpooling-operator-tester.h",
6554 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006555 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006556)
6557
Chao Mei6ddfc602020-05-13 22:29:36 -07006558############################### Misc unit tests ###############################
6559
6560xnnpack_unit_test(
6561 name = "memory_planner_test",
6562 srcs = [
6563 "test/memory-planner-test.cc",
6564 ],
6565 deps = [
6566 ":XNNPACK",
6567 ":memory_planner",
6568 ],
6569)
6570
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006571xnnpack_unit_test(
6572 name = "subgraph_nchw_test",
6573 srcs = [
6574 "src/xnnpack/subgraph.h",
6575 "test/subgraph-nchw.cc",
6576 "test/subgraph-tester.h",
6577 ],
6578 deps = [
6579 ":XNNPACK",
6580 ],
6581)
6582
Marat Dukhan08c4a432019-10-03 09:29:21 -07006583############################# Build configurations #############################
6584
Marat Dukhanb8642352019-10-30 15:43:02 -07006585# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006586config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006587 name = "xnn_enable_assembly_explicit_true",
6588 define_values = {"xnn_enable_assembly": "true"},
6589)
6590
6591# Disables usage of assembly kernels.
6592config_setting(
6593 name = "xnn_enable_assembly_explicit_false",
6594 define_values = {"xnn_enable_assembly": "false"},
6595)
6596
Marat Dukhan9de90e02020-06-18 16:04:12 -07006597# Enables usage of sparse inference.
6598config_setting(
6599 name = "xnn_enable_sparse_explicit_true",
6600 define_values = {"xnn_enable_sparse": "true"},
6601)
6602
6603# Disables usage of sparse inference.
6604config_setting(
6605 name = "xnn_enable_sparse_explicit_false",
6606 define_values = {"xnn_enable_sparse": "false"},
6607)
6608
Marat Dukhan05702cf2020-03-26 15:41:33 -07006609# Disables usage of HMP-aware optimizations.
6610config_setting(
6611 name = "xnn_enable_hmp_explicit_false",
6612 define_values = {"xnn_enable_hmp": "false"},
6613)
6614
Chao Mei6ddfc602020-05-13 22:29:36 -07006615# Enable usage of optimized memory allocation
6616config_setting(
6617 name = "xnn_enable_memopt_explicit_true",
6618 define_values = {"xnnpack_enable_memopt": "true"},
6619)
6620
6621# Disable usage of optimized memory allocation
6622config_setting(
6623 name = "xnn_enable_memopt_explicit_false",
6624 define_values = {"xnnpack_enable_memopt": "false"},
6625)
6626
Marat Dukhanb8642352019-10-30 15:43:02 -07006627# Builds with -c dbg
6628config_setting(
6629 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006630 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006631 "compilation_mode": "dbg",
6632 },
6633)
6634
6635# Builds with -c opt
6636config_setting(
6637 name = "optimized_build",
6638 values = {
6639 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006640 },
6641)
6642
6643config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006644 name = "linux_k8",
6645 values = {"cpu": "k8"},
6646)
6647
6648config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006649 name = "linux_arm",
6650 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006651)
6652
6653config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006654 name = "linux_armeabi",
6655 values = {"cpu": "armeabi"},
6656)
6657
6658config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006659 name = "linux_armhf",
6660 values = {"cpu": "armhf"},
6661)
6662
6663config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006664 name = "linux_armv7a",
6665 values = {"cpu": "armv7a"},
6666)
6667
6668config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006669 name = "linux_aarch64",
6670 values = {"cpu": "aarch64"},
6671)
6672
6673config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006674 name = "android",
6675 values = {"crosstool_top": "//external:android/crosstool"},
6676)
6677
6678config_setting(
6679 name = "android_armv7",
6680 values = {
6681 "crosstool_top": "//external:android/crosstool",
6682 "cpu": "armeabi-v7a",
6683 },
6684)
6685
6686config_setting(
6687 name = "android_arm64",
6688 values = {
6689 "crosstool_top": "//external:android/crosstool",
6690 "cpu": "arm64-v8a",
6691 },
6692)
6693
6694config_setting(
6695 name = "android_x86",
6696 values = {
6697 "crosstool_top": "//external:android/crosstool",
6698 "cpu": "x86",
6699 },
6700)
6701
6702config_setting(
6703 name = "android_x86_64",
6704 values = {
6705 "crosstool_top": "//external:android/crosstool",
6706 "cpu": "x86_64",
6707 },
6708)
6709
6710config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006711 name = "windows_x86_64",
6712 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006713)
6714
6715config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006716 name = "windows_x86_64_clang",
6717 values = {
6718 "compiler": "clang-cl",
6719 "cpu": "x64_windows",
6720 },
6721)
6722
6723config_setting(
6724 name = "windows_x86_64_mingw",
6725 values = {
6726 "compiler": "mingw-gcc",
6727 "cpu": "x64_windows",
6728 },
6729)
6730
6731config_setting(
6732 name = "windows_x86_64_msys",
6733 values = {
6734 "compiler": "msys-gcc",
6735 "cpu": "x64_windows",
6736 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006737)
6738
6739config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006740 name = "macos_x86_64",
6741 values = {
6742 "apple_platform_type": "macos",
6743 "cpu": "darwin",
6744 },
6745)
6746
6747config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006748 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006749 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006750)
6751
6752config_setting(
6753 name = "emscripten_wasm",
6754 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006755 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006756 "cpu": "wasm",
6757 },
6758)
6759
6760config_setting(
6761 name = "emscripten_wasmsimd",
6762 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006763 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006764 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006765 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006766 },
6767)
6768
6769config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006770 name = "ios_armv7",
6771 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006772 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006773 "cpu": "ios_armv7",
6774 },
6775)
6776
6777config_setting(
6778 name = "ios_arm64",
6779 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006780 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006781 "cpu": "ios_arm64",
6782 },
6783)
6784
6785config_setting(
6786 name = "ios_arm64e",
6787 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006788 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006789 "cpu": "ios_arm64e",
6790 },
6791)
6792
6793config_setting(
6794 name = "ios_x86",
6795 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006796 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006797 "cpu": "ios_i386",
6798 },
6799)
6800
6801config_setting(
6802 name = "ios_x86_64",
6803 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006804 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006805 "cpu": "ios_x86_64",
6806 },
6807)
6808
6809config_setting(
6810 name = "watchos_armv7k",
6811 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006812 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006813 "cpu": "watchos_armv7k",
6814 },
6815)
6816
6817config_setting(
6818 name = "watchos_arm64_32",
6819 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006820 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006821 "cpu": "watchos_arm64_32",
6822 },
6823)
6824
6825config_setting(
6826 name = "watchos_x86",
6827 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006828 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006829 "cpu": "watchos_i386",
6830 },
6831)
6832
6833config_setting(
6834 name = "watchos_x86_64",
6835 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006836 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006837 "cpu": "watchos_x86_64",
6838 },
6839)
6840
6841config_setting(
6842 name = "tvos_arm64",
6843 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006844 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006845 "cpu": "tvos_arm64",
6846 },
6847)
6848
6849config_setting(
6850 name = "tvos_x86_64",
6851 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006852 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006853 "cpu": "tvos_x86_64",
6854 },
6855)