blob: df27a4752e485860dfc9054bd3ad851d0f98157b [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",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/add-nc.c",
54 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
57 "src/operators/channel-pad-nc.c",
58 "src/operators/channel-shuffle-nc.c",
59 "src/operators/clamp-nc.c",
60 "src/operators/convolution-nchw.c",
61 "src/operators/convolution-nhwc.c",
62 "src/operators/deconvolution-nhwc.c",
63 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
66 "src/operators/hardswish-nc.c",
67 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
Marat Dukhan4662b192020-05-21 15:52:03 -070069 "src/operators/pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070070 "src/operators/prelu-nc.c",
71 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
74 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070075]
76
Marat Dukhan3a77ea72019-12-23 12:10:24 -080077TABLE_SRCS = [
78 "src/tables/exp2-k-over-64.c",
79 "src/tables/exp2-k-over-2048.c",
80]
81
Marat Dukhan08c4a432019-10-03 09:29:21 -070082SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080083 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080084 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080085 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070086 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
87 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070088 "src/f32-clamp/gen/scalar-x1.c",
89 "src/f32-clamp/gen/scalar-x2.c",
90 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080091 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080092 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -070093 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
94 "src/f32-dwconv-chw/3x3p1-scalar.c",
95 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
96 "src/f32-dwconv-chw/5x5p2-scalar.c",
97 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070098 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
99 "src/f32-dwconv/gen/up1x4-scalar.c",
100 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
101 "src/f32-dwconv/gen/up1x9-scalar.c",
102 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
103 "src/f32-dwconv/gen/up1x25-scalar.c",
104 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
105 "src/f32-dwconv/gen/up2x4-scalar.c",
106 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
107 "src/f32-dwconv/gen/up2x9-scalar.c",
108 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
109 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700110 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
111 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
113 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700114 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
115 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700116 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
117 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
119 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700120 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
121 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700122 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700123 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
124 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700125 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
127 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700128 "src/f32-gemm/gen/1x4-scalar.c",
129 "src/f32-gemm/gen/2x4-scalar.c",
130 "src/f32-gemm/gen/4x2-scalar.c",
131 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700132 "src/f32-gemm/gen/1x4-minmax-scalar.c",
133 "src/f32-gemm/gen/2x4-minmax-scalar.c",
134 "src/f32-gemm/gen/4x2-minmax-scalar.c",
135 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800136 "src/f32-hswish/gen/scalar-x1.c",
137 "src/f32-hswish/gen/scalar-x2.c",
138 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700139 "src/f32-ibilinear/gen/scalar-c1.c",
140 "src/f32-ibilinear/gen/scalar-c2.c",
141 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700142 "src/f32-igemm/gen/1x4-scalar.c",
143 "src/f32-igemm/gen/2x4-scalar.c",
144 "src/f32-igemm/gen/4x2-scalar.c",
145 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-igemm/gen/1x4-minmax-scalar.c",
147 "src/f32-igemm/gen/2x4-minmax-scalar.c",
148 "src/f32-igemm/gen/4x2-minmax-scalar.c",
149 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700150 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
152 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700153 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
154 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
156 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800157 "src/f32-prelu/gen/scalar-2x1.c",
158 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
170 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700171 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
174 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
177 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
180 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700181 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
182 "src/f32-spmm/gen/1x1-minmax-scalar.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
184 "src/f32-spmm/gen/2x1-minmax-scalar.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
186 "src/f32-spmm/gen/4x1-minmax-scalar.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
188 "src/f32-spmm/gen/8x1-minmax-scalar.c",
189 "src/f32-spmm/gen/8x2-minmax-scalar.c",
190 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700191 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
193 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
196 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
199 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
202 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800203 "src/f32-vbinary/gen/vmax-scalar-x1.c",
204 "src/f32-vbinary/gen/vmax-scalar-x2.c",
205 "src/f32-vbinary/gen/vmax-scalar-x4.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
208 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
209 "src/f32-vbinary/gen/vmin-scalar-x1.c",
210 "src/f32-vbinary/gen/vmin-scalar-x2.c",
211 "src/f32-vbinary/gen/vmin-scalar-x4.c",
212 "src/f32-vbinary/gen/vminc-scalar-x1.c",
213 "src/f32-vbinary/gen/vminc-scalar-x2.c",
214 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700215 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
217 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
220 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
223 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
226 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
229 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
232 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700233 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
235 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800236 "src/math/expminus-scalar-lut2048-p1.c",
237 "src/math/expminus-scalar-lut64-p2.c",
238 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700239 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700240 "src/math/roundne-scalar-nearbyint.c",
241 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700242 "src/math/roundd-scalar-addsub.c",
243 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700244 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700245 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700246 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700247 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700248 "src/math/roundz-scalar-addsub.c",
249 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700250 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800251 "src/math/sigmoid-scalar-lut2048-p1-div.c",
252 "src/math/sigmoid-scalar-lut64-p2-div.c",
253 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700254 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
255 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700256 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700257 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
258 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700259 "src/q8-gemm/2x2-minmax-scalar.c",
260 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700261 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700262 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700263 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700264 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700265 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700266 "src/x32-fill/scalar-float.c",
267 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700268 "src/x32-packx/x2-scalar.c",
269 "src/x32-packx/x3-scalar.c",
270 "src/x32-packx/x4-scalar.c",
271 "src/x32-pad/x2-scalar.c",
272 "src/x32-unpool/scalar.c",
273 "src/x32-zip/x2-scalar.c",
274 "src/x32-zip/x3-scalar.c",
275 "src/x32-zip/x4-scalar.c",
276 "src/x32-zip/xm-scalar.c",
277 "src/x8-lut/scalar.c",
278 "src/x8-zip/x2-scalar.c",
279 "src/x8-zip/x3-scalar.c",
280 "src/x8-zip/x4-scalar.c",
281 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700282 "src/requantization/precise-scalar.c",
283 "src/requantization/fp32-scalar.c",
284 "src/requantization/q31-scalar.c",
285 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700286]
287
Marat Dukhan436ebe62019-12-04 15:10:12 -0800288WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700289 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
290 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700291 "src/f32-clamp/gen/wasm-x1.c",
292 "src/f32-clamp/gen/wasm-x2.c",
293 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700294 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
295 "src/f32-dwconv/gen/up1x4-wasm.c",
296 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
297 "src/f32-dwconv/gen/up1x9-wasm.c",
298 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
299 "src/f32-dwconv/gen/up1x25-wasm.c",
300 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
301 "src/f32-dwconv/gen/up2x4-wasm.c",
302 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
303 "src/f32-dwconv/gen/up2x9-wasm.c",
304 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
305 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700306 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
307 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
308 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
309 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700310 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
311 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700312 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
313 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
314 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
315 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700316 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
317 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700318 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
319 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700320 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
321 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
322 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700323 "src/f32-gemm/gen/1x4-wasm.c",
324 "src/f32-gemm/gen/2x4-wasm.c",
325 "src/f32-gemm/gen/4x2-wasm.c",
326 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700327 "src/f32-gemm/gen/1x4-minmax-wasm.c",
328 "src/f32-gemm/gen/2x4-minmax-wasm.c",
329 "src/f32-gemm/gen/4x2-minmax-wasm.c",
330 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800331 "src/f32-hswish/gen/wasm-x1.c",
332 "src/f32-hswish/gen/wasm-x2.c",
333 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700334 "src/f32-igemm/gen/1x4-wasm.c",
335 "src/f32-igemm/gen/2x4-wasm.c",
336 "src/f32-igemm/gen/4x2-wasm.c",
337 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700338 "src/f32-igemm/gen/1x4-minmax-wasm.c",
339 "src/f32-igemm/gen/2x4-minmax-wasm.c",
340 "src/f32-igemm/gen/4x2-minmax-wasm.c",
341 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700342 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
343 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
344 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700345 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
346 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
347 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
348 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
349 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
350 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
351 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
352 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
353 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
354 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
355 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
356 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800357 "src/f32-vbinary/gen/vmax-wasm-x1.c",
358 "src/f32-vbinary/gen/vmax-wasm-x2.c",
359 "src/f32-vbinary/gen/vmax-wasm-x4.c",
360 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
361 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
362 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
363 "src/f32-vbinary/gen/vmin-wasm-x1.c",
364 "src/f32-vbinary/gen/vmin-wasm-x2.c",
365 "src/f32-vbinary/gen/vmin-wasm-x4.c",
366 "src/f32-vbinary/gen/vminc-wasm-x1.c",
367 "src/f32-vbinary/gen/vminc-wasm-x2.c",
368 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700369 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
370 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
371 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
372 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
373 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
374 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
375 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
376 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
377 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
378 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
379 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
380 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
381 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
382 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
383 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
384 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
385 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
386 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700387 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
388 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
389 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800390]
391
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800392PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800393 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800394 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800395 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700396 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
397 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700398 "src/f32-clamp/gen/psimd-x4.c",
399 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700400 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
401 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
402 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
403 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
404 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
405 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
406 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
407 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
408 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
409 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
410 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
411 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700412 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
413 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700414 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
415 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
416 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
417 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
418 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
419 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
420 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
421 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
422 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
423 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
424 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
425 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
426 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
427 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
428 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
429 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
430 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
431 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
432 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800433 "src/f32-hswish/gen/psimd-x4.c",
434 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700435 "src/f32-ibilinear/gen/psimd-c4.c",
436 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700437 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
438 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
439 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
440 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
441 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
442 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
443 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
444 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
445 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
446 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700447 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
448 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
449 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700450 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800451 "src/f32-prelu/gen/psimd-2x4.c",
452 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800453 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700454 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
455 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
456 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
457 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
458 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
459 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
460 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
461 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800462 "src/f32-vbinary/gen/vmax-psimd-x4.c",
463 "src/f32-vbinary/gen/vmax-psimd-x8.c",
464 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
465 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
466 "src/f32-vbinary/gen/vmin-psimd-x4.c",
467 "src/f32-vbinary/gen/vmin-psimd-x8.c",
468 "src/f32-vbinary/gen/vminc-psimd-x4.c",
469 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700470 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
471 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
472 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
473 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
474 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
475 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
476 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
477 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
478 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
479 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
480 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
481 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700482 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
483 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700484 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700485 "src/x32-packx/x4-psimd.c",
486 "src/x32-pad/x2-psimd.c",
487 "src/x32-unpool/psimd.c",
488 "src/x32-zip/x2-psimd.c",
489 "src/x32-zip/x3-psimd.c",
490 "src/x32-zip/x4-psimd.c",
491 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700492 "src/requantization/precise-psimd.c",
493 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700494]
495
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800496PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800497 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
498 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
499 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
500 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
501 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
502 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
503 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
504 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
505 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
506 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
507 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
508 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800509 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
510 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
511 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
512 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
513 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
514 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700515 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700516 "src/math/roundd-psimd-addsub.c",
517 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700518 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800519 "src/math/sigmoid-psimd-p5-div.c",
520]
521
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522# ISA-specific micro-kernels
523NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700524 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
525 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700526 "src/f32-clamp/gen/neon-x4.c",
527 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700528 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
529 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700530 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
531 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700532 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
533 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700534 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
535 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700536 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
537 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
538 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
539 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700540 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
541 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
542 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
543 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700544 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
545 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
546 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
547 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700548 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700549 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
550 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700551 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
552 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
553 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
554 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
555 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
556 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
557 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
558 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
559 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
560 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
561 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
562 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
563 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
564 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
565 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
566 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
567 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
568 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
569 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
570 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
571 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
572 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
573 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
574 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
575 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
576 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
577 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
578 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
579 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
580 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
581 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800582 "src/f32-hswish/gen/neon-x4.c",
583 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700584 "src/f32-ibilinear/gen/neon-c4.c",
585 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700586 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
587 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
588 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
589 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
590 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
591 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
592 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
593 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
594 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
595 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
596 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
597 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
598 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
599 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
600 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
601 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700602 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
603 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
604 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700605 "src/f32-ppmm/gen/4x8-minmax-neon.c",
606 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800607 "src/f32-prelu/gen/neon-2x4.c",
608 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800609 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
610 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
611 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
612 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
613 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
614 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
615 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
616 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
617 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
618 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
619 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
620 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
621 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
622 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
623 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
624 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
625 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
626 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
627 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
628 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
629 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
630 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
631 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
632 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700633 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800634 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800635 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
636 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
637 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
638 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
639 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
640 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
641 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
642 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
643 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
644 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
645 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
646 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
647 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
648 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
649 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
650 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
651 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
652 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700653 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
654 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
655 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
656 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800657 "src/f32-vbinary/gen/vmax-neon-x4.c",
658 "src/f32-vbinary/gen/vmax-neon-x8.c",
659 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
660 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
661 "src/f32-vbinary/gen/vmin-neon-x4.c",
662 "src/f32-vbinary/gen/vmin-neon-x8.c",
663 "src/f32-vbinary/gen/vminc-neon-x4.c",
664 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700665 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
666 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
667 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
668 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
669 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
670 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
671 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
672 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
673 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
674 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700675 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
676 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
677 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
678 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700679 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700680 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
681 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700682 "src/q8-gemm/4x8-minmax-neon.c",
683 "src/q8-gemm/8x8-minmax-neon.c",
684 "src/q8-igemm/4x8-minmax-neon.c",
685 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700686 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700687 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700688 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700689 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700690 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700691 "src/x32-packx/x4-neon-st4.c",
692 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700693 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700694 "src/x32-zip/x2-neon.c",
695 "src/x32-zip/x3-neon.c",
696 "src/x32-zip/x4-neon.c",
697 "src/x32-zip/xm-neon.c",
698 "src/x8-zip/x2-neon.c",
699 "src/x8-zip/x3-neon.c",
700 "src/x8-zip/x4-neon.c",
701 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700702 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700703 "src/math/roundd-neon-addsub.c",
704 "src/math/roundd-neon-cvt.c",
705 "src/math/roundu-neon-addsub.c",
706 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700707 "src/math/roundz-neon-addsub.c",
708 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800709 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800710 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
711 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
712 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
713 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
714 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
715 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700716 "src/requantization/precise-neon.c",
717 "src/requantization/fp32-neon.c",
718 "src/requantization/q31-neon.c",
719 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700720]
721
722NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700723 "src/f32-ibilinear/gen/neonfma-c4.c",
724 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700725 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
726 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
727 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
728 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
729 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
730 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
731 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
732 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
733 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700734 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
735 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
736 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
737 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700738 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
739 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
740 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
741 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700742 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
743 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
744 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
745 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700746 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
747 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
748 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
749 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
750 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
751 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
752 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
753 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
754 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
755 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
756 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
757 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
758 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
760 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
761 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
762 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
763 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800764 "src/f32-hswish/gen/neonfma-x4.c",
765 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700766 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
767 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800768 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
769 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
770 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
771 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
772 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
773 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
774 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
775 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
776 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
777 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
778 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
779 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
780 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
781 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
782 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
783 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
784 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
785 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
786 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
787 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
788 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
789 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
790 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
791 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800792 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
832 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
833 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
834 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
835 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
836 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
837 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
838 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
839 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
840 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
841 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
842 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
843 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
844 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
845 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700846 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
847 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800848 "src/math/exp-neonfma-lut64-p2.c",
849 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800850 "src/math/expminus-neonfma-lut2048-p1.c",
851 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800852 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800853 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
854 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
855 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
856 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
857 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
858 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
859 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
860 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
861 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
862 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
863 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
864 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
865 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
866 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
867 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
868 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
869 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
870 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700871]
872
873AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700874 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
875 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
876 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
877 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
878 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
879 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700880 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
881 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
882 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
883 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
884 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
885 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
886 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
887 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
888 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
889 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
890 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
891 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
892 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
893 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
894 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
895 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
896 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
897 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
898 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
899 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700900 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
903 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700904 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
905 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700906 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
907 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700908 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
909 "src/f32-dwconv-chw/3x3p1-neonfma.c",
910 "src/f32-dwconv-chw/5x5p2-neonfma.c",
911 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
912 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800913 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
914 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
915 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
916 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
917 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
918 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
919 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
925 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
926 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
927 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
928 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
929 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
930 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700931 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
932 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
933 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
934 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
935 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
936 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
937 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
938 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
939 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
940 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
941 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
942 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
943 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
944 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
945 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
946 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
947 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
948 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800949 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
950 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
951 "src/math/sigmoid-neonfma-rr1-p5-div.c",
952 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
953 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
954 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700955]
956
Marat Dukhan8853b822020-05-07 12:19:01 -0700957NEONV8_UKERNELS = [
958 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700959 "src/math/roundd-neonv8.c",
960 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700961 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700962]
963
Marat Dukhan08c4a432019-10-03 09:29:21 -0700964AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -0700965 "src/f16-clamp/gen/neonfp16arith-x16.c",
966 "src/f16-clamp/gen/neonfp16arith-x8.c",
967 "src/f16-hswish/gen/neonfp16arith-x16.c",
968 "src/f16-hswish/gen/neonfp16arith-x8.c",
969 "src/f16-prelu/gen/neonfp16arith-2x16.c",
970 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -0700971 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
972 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
973 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
974 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
975 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
976 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
977 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
978 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
979 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
980 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
981 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
982 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
983 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
984 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
985 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
986 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
987 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
988 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
989 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
990 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
991 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
992 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
993 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
994 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
995 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
996 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
997 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
998 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -0700999 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1000 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1001 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1002 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1003 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1004 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1005 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1006 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001007 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1008 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1009 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1010 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001011 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1012 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1013 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1014 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1015 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1016 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1017 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1018 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1019 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1020 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1021 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1022 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001023 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1024 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1025 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1026 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1027 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1028 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1029 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1030 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001031]
1032
1033SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001034 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1035 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001036 "src/f32-clamp/gen/sse-x4.c",
1037 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001038 "src/f32-dwconv-chw/3x3p1-sse.c",
1039 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001040 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1041 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1042 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1043 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1044 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1045 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1046 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1047 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1048 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1049 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1050 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1051 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001052 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001053 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1054 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001055 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1056 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1057 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1058 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1059 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1060 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1061 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1062 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1063 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1064 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1065 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1066 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1067 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001068 "src/f32-hswish/gen/sse-x4.c",
1069 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001070 "src/f32-ibilinear/gen/sse-c4.c",
1071 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001072 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1073 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1074 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1075 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1076 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1077 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1078 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001079 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1080 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1081 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001082 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001083 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001084 "src/f32-spmm/gen/4x1-minmax-sse.c",
1085 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001086 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1087 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1088 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1089 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1090 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1091 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1092 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1093 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001094 "src/f32-vbinary/gen/vmax-sse-x4.c",
1095 "src/f32-vbinary/gen/vmax-sse-x8.c",
1096 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1097 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1098 "src/f32-vbinary/gen/vmin-sse-x4.c",
1099 "src/f32-vbinary/gen/vmin-sse-x8.c",
1100 "src/f32-vbinary/gen/vminc-sse-x4.c",
1101 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001102 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1103 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1104 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1105 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1106 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1107 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1108 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1109 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1110 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1111 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1112 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1113 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001114 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1115 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001116 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001117 "src/x32-packx/x4-sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001118 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001119 "src/math/roundd-sse-addsub.c",
1120 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001121 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001122]
1123
1124SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001125 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1126 "src/f32-argmaxpool/4x-sse2-c4.c",
1127 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001128 "src/f32-prelu/gen/sse2-2x4.c",
1129 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001130 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1131 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1133 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1134 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1135 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1136 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1137 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1138 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1139 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1140 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1141 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001142 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001143 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001144 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001145 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001146 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1147 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001148 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1149 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001150 "src/q8-igemm/4x4c2-minmax-sse2.c",
1151 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001152 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1153 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001154 "src/q8-gemm/2x4c8-minmax-sse2.c",
1155 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001156 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001157 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001158 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001159 "src/u8-rmax/sse2.c",
1160 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001161 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001162 "src/x32-zip/x2-sse2.c",
1163 "src/x32-zip/x3-sse2.c",
1164 "src/x32-zip/x4-sse2.c",
1165 "src/x32-zip/xm-sse2.c",
1166 "src/x8-zip/x2-sse2.c",
1167 "src/x8-zip/x3-sse2.c",
1168 "src/x8-zip/x4-sse2.c",
1169 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001170 "src/math/exp-sse2-p5.c",
1171 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001172 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001173 "src/math/roundd-sse2-cvt.c",
1174 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001175 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001176 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001177 "src/requantization/precise-sse2.c",
1178 "src/requantization/fp32-sse2.c",
1179 "src/requantization/q31-sse2.c",
1180 "src/requantization/gemmlowp-sse2.c",
1181]
1182
1183SSSE3_UKERNELS = [
1184 "src/requantization/precise-ssse3.c",
1185 "src/requantization/q31-ssse3.c",
1186 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001187]
1188
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001189SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001190 "src/f32-prelu/gen/sse41-2x4.c",
1191 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001192 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001193 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001194 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001195 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001196 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1197 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001198 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001199 "src/math/roundd-sse41.c",
1200 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001201 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001202 "src/requantization/precise-sse4.c",
1203 "src/requantization/q31-sse4.c",
1204 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001205]
1206
Marat Dukhan08c4a432019-10-03 09:29:21 -07001207AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001208 "src/f32-clamp/gen/avx-x8.c",
1209 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001210 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1211 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1212 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1213 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1214 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1215 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1216 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1217 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1218 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1219 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1220 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1221 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1222 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1223 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1224 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1225 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1226 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1227 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1228 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1229 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1230 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1231 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1232 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1233 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1234 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1235 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1236 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1237 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1238 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1239 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001240 "src/f32-hswish/gen/avx-x8.c",
1241 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001242 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1243 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1244 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1245 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1246 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1247 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1248 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1249 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1250 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001251 "src/f32-prelu/gen/avx-2x8.c",
1252 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001253 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001254 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1255 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1256 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1257 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1258 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1259 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1260 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1261 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001262 "src/f32-vbinary/gen/vmax-avx-x8.c",
1263 "src/f32-vbinary/gen/vmax-avx-x16.c",
1264 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1265 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1266 "src/f32-vbinary/gen/vmin-avx-x8.c",
1267 "src/f32-vbinary/gen/vmin-avx-x16.c",
1268 "src/f32-vbinary/gen/vminc-avx-x8.c",
1269 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001270 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1271 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1272 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1273 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1274 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1275 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1276 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1277 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1278 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1279 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1280 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1281 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001282 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001283]
1284
Marat Dukhanfda12b82019-11-21 12:27:59 -08001285FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001286 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1287 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1288 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1289 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1290 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1291 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1292 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1293 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1294 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1295 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1296 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1297 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1298 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1299 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1300 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1301 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1302 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1303 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1304 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1305 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1306 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1307 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1308 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1309 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1310 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1311 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1312 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1313 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1314 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1315 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1316 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1317 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1318 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1319 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1320 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1321 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1322 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1323 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1324 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1325 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001326 "src/f32-hswish/gen/fma3-x8.c",
1327 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001328 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1329 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1330 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1331 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1332 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1333 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1334 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1335 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1336 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1337 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1338 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1339 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1340 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1341 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001342]
1343
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001344AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001345 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1346 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1347 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1348 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1349 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1350 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1351 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1352 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1353 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1354 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1355 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1356 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1357 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1358 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1359 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1360 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1361 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1362 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1363 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1364 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1365 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1366 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1367 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1368 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1369 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1370 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1373 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1374 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1375 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1376 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1377 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1378 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1379 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1380 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001381 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1382 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1383 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1384 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1385 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1386 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1387 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1388 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1389 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1390 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1403 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1404 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1405 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1406 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1407 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1408 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1409 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1410 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001411 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1412 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1415 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1416 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1417 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1418 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1419 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1420 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1421 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1422 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1423 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1424 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1425 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1426 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1427 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1428 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1429 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1430 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1431 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1432 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1433 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1434 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001435 "src/math/exp-avx2-p5.c",
1436 "src/math/exp-avx2-perm-p3.c",
1437 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001438 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001439 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001440 "src/math/sigmoid-avx2-rr2-p5-div.c",
1441 "src/math/sigmoid-avx2-rr1-p5-div.c",
1442 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1443 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1444 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1445 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001446]
1447
Marat Dukhan08c4a432019-10-03 09:29:21 -07001448AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001449 "src/f32-clamp/gen/avx512f-x16.c",
1450 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001451 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1452 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1453 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1454 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1455 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1456 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1457 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1458 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1459 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1460 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1461 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1462 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1463 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1464 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1465 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1466 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1467 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1468 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1469 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1470 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1471 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1472 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1473 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1474 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001475 "src/f32-hswish/gen/avx512f-x16.c",
1476 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001477 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1478 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1479 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1480 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1481 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1482 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001483 "src/f32-prelu/gen/avx512f-2x16.c",
1484 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001485 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1486 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1487 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1488 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1489 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1490 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1491 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1492 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1493 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1494 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1495 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1496 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1497 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1498 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1499 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1500 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1501 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1502 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1503 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1504 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1505 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1506 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1507 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1508 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1509 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1510 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1511 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1512 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1513 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1514 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1515 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1516 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1517 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1518 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1519 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1520 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001521 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001522 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1523 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1524 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1525 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1526 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1527 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1528 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1529 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001530 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1531 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1532 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1533 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1534 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1535 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1536 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1537 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001538 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1539 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1540 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1541 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1542 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1543 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1544 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1545 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1546 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1547 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1548 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1549 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001550 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001551 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1552 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1553 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1554 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1555 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1556 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1557 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1558 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1559 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1560 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1561 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1562 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1563 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1564 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1565 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1566 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1567 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1568 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1569 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1570 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1571 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1572 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1573 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1574 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001575 "src/math/exp-avx512f-p5-scalef.c",
1576 "src/math/exp-avx512f-p5.c",
1577 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001578 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001579 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580]
1581
1582AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001583 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001584 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001585 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1586 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1587 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1588 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1589 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1590 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1591 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1592 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1593 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1594 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595]
1596
1597AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001598 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001599 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1600 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001601 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001602 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1603 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001604 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1605 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1606 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001607 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001608 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1609 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1610 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001611 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001612 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1613 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001614 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001615 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1616 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1617 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1618 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1619 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1620 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1621 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1622 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1623 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1624 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1625 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1626 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1627 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1628 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1629 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1630 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1631 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1632 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1633 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1634 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1635 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001636 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001637 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1638 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1639 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1640 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1641 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1642 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1643 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1644 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1645 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1646 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1647 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1648 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1649 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1650 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1651 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1652 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1653 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1654 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1655 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1656 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1657 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1658 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1659 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1660 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1661 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1662 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1663 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1664 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1665 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1666 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1667 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1668 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1669 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1670 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1671 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1672 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1673 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1674 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001675]
1676
Marat Dukhan1b354632020-03-23 12:50:22 -07001677INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001678 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001679 "src/xnnpack/argmaxpool.h",
1680 "src/xnnpack/avgpool.h",
1681 "src/xnnpack/clamp.h",
1682 "src/xnnpack/common.h",
1683 "src/xnnpack/conv.h",
1684 "src/xnnpack/dwconv.h",
1685 "src/xnnpack/gavgpool.h",
1686 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001687 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001688 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001689 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001690 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001691 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001692 "src/xnnpack/lut.h",
1693 "src/xnnpack/math.h",
1694 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001695 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001696 "src/xnnpack/packx.h",
1697 "src/xnnpack/pad.h",
1698 "src/xnnpack/params.h",
1699 "src/xnnpack/pavgpool.h",
1700 "src/xnnpack/ppmm.h",
1701 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001702 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001703 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001704 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001705 "src/xnnpack/rmax.h",
1706 "src/xnnpack/scalar-utils.h",
1707 "src/xnnpack/spmm.h",
1708 "src/xnnpack/unpool.h",
1709 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001710 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001711 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001712 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001713 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001714 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001715 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001716 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001717]
1718
1719INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720 "include/xnnpack.h",
1721 "src/xnnpack/allocator.h",
1722 "src/xnnpack/compute.h",
1723 "src/xnnpack/im2col.h",
1724 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001725 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001726 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001727 "src/xnnpack/operator.h",
1728 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001729 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001731 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001732 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001733]
1734
Marat Dukhan1b354632020-03-23 12:50:22 -07001735ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001736 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001737]
1738
Marat Dukhan1b354632020-03-23 12:50:22 -07001739MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001740 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001741 "include/xnnpack.h",
1742]
1743
Marat Dukhan1b354632020-03-23 12:50:22 -07001744MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001745 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001746 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747 "src/xnnpack/requantization.h",
1748 "include/xnnpack.h",
1749]
1750
1751OPERATOR_TEST_PARAMS_HDRS = [
1752 "src/xnnpack/params.h",
1753 "src/xnnpack/common.h",
1754]
1755
1756WEIGHTS_PACK_HDRS = [
1757 "src/xnnpack/pack.h",
1758 "src/xnnpack/operator.h",
1759 "src/xnnpack/compute.h",
1760]
1761
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001762LOGGING_COPTS = select({
1763 # No logging in optimized mode
1764 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1765 # Full logging in debug mode
1766 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1767 # Error-only logging in default (fastbuild) mode
1768 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1769})
1770
1771LOGGING_HDRS = [
1772 "src/xnnpack/log.h",
1773]
1774
Marat Dukhan08c4a432019-10-03 09:29:21 -07001775xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001776 name = "tables",
1777 srcs = TABLE_SRCS,
1778 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001779 gcc_copts = xnnpack_gcc_std_copts(),
1780 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001781)
1782
1783xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001784 name = "scalar_ukernels",
1785 srcs = SCALAR_UKERNELS,
1786 hdrs = INTERNAL_HDRS,
1787 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001788 gcc_copts = xnnpack_gcc_std_copts(),
1789 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001791 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001792 "@FP16",
1793 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001794 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001795 ],
1796)
1797
1798xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001799 name = "wasm_ukernels",
1800 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001801 gcc_copts = xnnpack_gcc_std_copts(),
1802 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001803 wasm_srcs = WASM_UKERNELS,
1804 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001805 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001806 "@FP16",
1807 "@FXdiv",
1808 "@pthreadpool",
1809 ],
1810)
1811
1812xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001813 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 hdrs = INTERNAL_HDRS,
1815 aarch32_copts = [
1816 "-marm",
1817 "-mfpu=neon",
1818 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001819 gcc_copts = xnnpack_gcc_std_copts(),
1820 gcc_x86_copts = ["-msse2"],
1821 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822 optimized_copts = [
1823 "-O3",
1824 "-ffast-math",
1825 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001826 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001827 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001828 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829 "@FP16",
1830 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001831 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001832 ],
1833)
1834
1835xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001836 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001837 hdrs = INTERNAL_HDRS,
1838 aarch32_copts = [
1839 "-marm",
1840 "-mfpu=neon",
1841 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001842 gcc_copts = xnnpack_gcc_std_copts(),
1843 gcc_x86_copts = ["-msse2"],
1844 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001845 optimized_copts = [
1846 "-O3",
1847 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001848 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001849 deps = [
1850 ":tables",
1851 "@FP16",
1852 "@psimd",
1853 "@pthreadpool",
1854 ],
1855)
1856
1857xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001858 name = "neon_ukernels",
1859 hdrs = INTERNAL_HDRS,
1860 aarch32_copts = [
1861 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001862 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001863 "-mfpu=neon",
1864 ],
1865 aarch32_srcs = NEON_UKERNELS,
1866 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001867 gcc_copts = xnnpack_gcc_std_copts(),
1868 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001869 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001870 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001871 "@FP16",
1872 "@pthreadpool",
1873 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874)
1875
1876xnnpack_cc_library(
1877 name = "neonfma_ukernels",
1878 hdrs = INTERNAL_HDRS,
1879 aarch32_copts = [
1880 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001881 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001882 "-mfpu=neon-vfpv4",
1883 ],
1884 aarch32_srcs = NEONFMA_UKERNELS,
1885 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001886 gcc_copts = xnnpack_gcc_std_copts(),
1887 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001888 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001889 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001890 "@FP16",
1891 "@pthreadpool",
1892 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001893)
1894
1895xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001896 name = "neonv8_ukernels",
1897 hdrs = INTERNAL_HDRS,
1898 aarch32_copts = [
1899 "-marm",
1900 "-march=armv8-a",
1901 "-mfpu=neon-fp-armv8",
1902 ],
1903 aarch32_srcs = NEONV8_UKERNELS,
1904 aarch64_srcs = NEONV8_UKERNELS,
1905 gcc_copts = xnnpack_gcc_std_copts(),
1906 msvc_copts = xnnpack_msvc_std_copts(),
1907 deps = [
1908 ":tables",
1909 "@FP16",
1910 "@pthreadpool",
1911 ],
1912)
1913
1914xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915 name = "neonfp16arith_ukernels",
1916 hdrs = INTERNAL_HDRS,
1917 aarch64_copts = ["-march=armv8.2-a+fp16"],
1918 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001919 gcc_copts = xnnpack_gcc_std_copts(),
1920 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001921 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001922 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001923 "@FP16",
1924 "@pthreadpool",
1925 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001926)
1927
1928xnnpack_cc_library(
1929 name = "sse2_ukernels",
1930 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001931 gcc_copts = xnnpack_gcc_std_copts(),
1932 gcc_x86_copts = ["-msse2"],
1933 msvc_copts = xnnpack_msvc_std_copts(),
1934 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001936 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001937 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001938 "@FP16",
1939 "@pthreadpool",
1940 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001941)
1942
1943xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001944 name = "ssse3_ukernels",
1945 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001946 gcc_copts = xnnpack_gcc_std_copts(),
1947 gcc_x86_copts = ["-mssse3"],
1948 msvc_copts = xnnpack_msvc_std_copts(),
1949 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001950 x86_srcs = SSSE3_UKERNELS,
1951 deps = [
1952 ":tables",
1953 "@FP16",
1954 "@pthreadpool",
1955 ],
1956)
1957
1958xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001959 name = "sse41_ukernels",
1960 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001961 gcc_copts = xnnpack_gcc_std_copts(),
1962 gcc_x86_copts = ["-msse4.1"],
1963 msvc_copts = xnnpack_msvc_std_copts(),
1964 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001965 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001966 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001967 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001968 "@FP16",
1969 "@pthreadpool",
1970 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001971)
1972
1973xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001974 name = "avx_ukernels",
1975 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001976 gcc_copts = xnnpack_gcc_std_copts(),
1977 gcc_x86_copts = ["-mavx"],
1978 msvc_copts = xnnpack_msvc_std_copts(),
1979 msvc_x86_32_copts = ["/arch:AVX"],
1980 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001981 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001982 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001983 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001984 "@FP16",
1985 "@pthreadpool",
1986 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001987)
1988
1989xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001990 name = "fma3_ukernels",
1991 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001992 gcc_copts = xnnpack_gcc_std_copts(),
1993 gcc_x86_copts = ["-mfma"],
1994 msvc_copts = xnnpack_msvc_std_copts(),
1995 msvc_x86_32_copts = ["/arch:AVX"],
1996 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001997 x86_srcs = FMA3_UKERNELS,
1998 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001999 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002000 "@FP16",
2001 "@pthreadpool",
2002 ],
2003)
2004
2005xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002006 name = "avx2_ukernels",
2007 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002008 gcc_copts = xnnpack_gcc_std_copts(),
2009 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002010 "-mfma",
2011 "-mavx2",
2012 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002013 msvc_copts = xnnpack_msvc_std_copts(),
2014 msvc_x86_32_copts = ["/arch:AVX2"],
2015 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002016 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002017 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002018 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002019 "@FP16",
2020 "@pthreadpool",
2021 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002022)
2023
2024xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025 name = "avx512f_ukernels",
2026 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002027 gcc_copts = xnnpack_gcc_std_copts(),
2028 gcc_x86_copts = ["-mavx512f"],
2029 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2030 msvc_copts = xnnpack_msvc_std_copts(),
2031 msvc_x86_32_copts = ["/arch:AVX512"],
2032 msvc_x86_64_copts = ["/arch:AVX512"],
2033 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002034 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002035 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002036 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002037 "@FP16",
2038 "@pthreadpool",
2039 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002040)
2041
2042xnnpack_cc_library(
2043 name = "asm_ukernels",
2044 hdrs = ["src/xnnpack/assembly.h"],
2045 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002046 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002047 aarch64_srcs = AARCH64_ASM_UKERNELS,
2048)
2049
2050xnnpack_aggregate_library(
2051 name = "ukernels",
2052 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002053 ":neon_ukernels",
2054 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002055 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002056 ":asm_ukernels",
2057 ],
2058 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002059 ":neon_ukernels",
2060 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002061 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002062 ":neonfp16arith_ukernels",
2063 ":asm_ukernels",
2064 ],
2065 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002066 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002067 ":psimd_fastmath_ukernels",
2068 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002070 wasm_deps = [":wasm_ukernels"],
2071 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002073 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002074 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002075 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002076 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002077 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002078 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002079 ":avx512f_ukernels",
2080 ],
2081)
2082
2083xnnpack_cc_library(
2084 name = "im2col",
2085 srcs = ["src/im2col.c"],
2086 hdrs = [
2087 "src/xnnpack/common.h",
2088 "src/xnnpack/im2col.h",
2089 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002090 gcc_copts = xnnpack_gcc_std_copts(),
2091 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092)
2093
2094xnnpack_cc_library(
2095 name = "indirection",
2096 srcs = ["src/indirection.c"],
2097 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002098 gcc_copts = xnnpack_gcc_std_copts(),
2099 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002100 deps = [
2101 "@FP16",
2102 "@FXdiv",
2103 "@pthreadpool",
2104 ],
2105)
2106
2107xnnpack_cc_library(
2108 name = "operator_run",
2109 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002110 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002111 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002112 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2113 "//conditions:default": [],
2114 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002115 gcc_copts = xnnpack_gcc_std_copts(),
2116 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002117 deps = [
2118 "@FP16",
2119 "@FXdiv",
2120 "@clog",
2121 "@pthreadpool",
2122 ],
2123)
2124
Chao Mei6ddfc602020-05-13 22:29:36 -07002125xnnpack_cc_library(
2126 name = "memory_planner",
2127 srcs = ["src/memory-planner.c"],
2128 hdrs = INTERNAL_HDRS,
2129 defines = select({
2130 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2131 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2132 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2133 }),
2134 gcc_copts = xnnpack_gcc_std_copts(),
2135 msvc_copts = xnnpack_msvc_std_copts(),
2136 deps = [
2137 "@pthreadpool",
2138 ],
2139)
2140
Marat Dukhan08c4a432019-10-03 09:29:21 -07002141cc_library(
2142 name = "enable_assembly",
2143 defines = select({
2144 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2145 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002146 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 }),
2148)
2149
Marat Dukhancf056b22019-10-07 10:26:29 -07002150xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002151 name = "operators",
2152 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002153 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002155 ],
2156 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002157 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002158 "-Isrc",
2159 "-Iinclude",
2160 ] + select({
2161 ":debug_build": [],
2162 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002163 }) + select({
2164 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2165 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002167 gcc_copts = xnnpack_gcc_std_copts(),
2168 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002169 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002170 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002171 "@FP16",
2172 "@FXdiv",
2173 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002174 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002175 ],
2176)
2177
Marat Dukhan10a38082020-04-17 03:58:35 -07002178xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002179 name = "XNNPACK",
2180 srcs = [
2181 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002182 "src/runtime.c",
2183 "src/subgraph.c",
2184 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002185 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002186 hdrs = ["include/xnnpack.h"],
2187 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002188 "-Isrc",
2189 "-Iinclude",
2190 ] + select({
2191 ":debug_build": [],
2192 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002193 }) + select({
2194 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2195 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002196 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002197 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002198 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002199 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002200 visibility = xnnpack_visibility(),
2201 deps = [
2202 ":enable_assembly",
Chao Mei6ddfc602020-05-13 22:29:36 -07002203 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002204 ":ukernels",
2205 ":operator_run",
2206 ":operators",
2207 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07002208 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002209 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002210 ] + select({
2211 ":emscripten": [],
2212 "//conditions:default": ["@cpuinfo"],
2213 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002214)
2215
Marat Dukhan10a38082020-04-17 03:58:35 -07002216xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002217 name = "xnnpack_operators_nhwc_f32",
2218 srcs = [
2219 "src/init.c",
2220 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002221 hdrs = ["include/xnnpack.h"],
2222 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002223 "-Isrc",
2224 "-Iinclude",
2225 ] + select({
2226 ":debug_build": [],
2227 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002228 }) + select({
2229 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2230 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002231 }),
2232 defines = [
2233 "XNN_NO_Q8_OPERATORS",
2234 "XNN_NO_U8_OPERATORS",
2235 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002236 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002237 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002238 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002239 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002240 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002241 visibility = xnnpack_visibility(),
2242 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002243 ":enable_assembly",
2244 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002245 ":operator_run",
2246 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002247 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002248 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002249 ] + select({
2250 ":emscripten": [],
2251 "//conditions:default": ["@cpuinfo"],
2252 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002253)
2254
Marat Dukhancf056b22019-10-07 10:26:29 -07002255xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002256 name = "bench_utils",
2257 srcs = ["bench/utils.cc"],
2258 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002259 deps = [
2260 "@com_google_benchmark//:benchmark",
2261 "@cpuinfo",
2262 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002263)
2264
Frank Barchard7e955972019-10-11 10:34:25 -07002265######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002266
2267xnnpack_benchmark(
2268 name = "q8_gemm_bench",
2269 srcs = [
2270 "bench/gemm.h",
2271 "bench/q8-gemm.cc",
2272 "src/xnnpack/AlignedAllocator.h",
2273 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002274 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002275 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2276)
2277
2278xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002279 name = "f16_igemm_bench",
2280 srcs = [
2281 "bench/f16-igemm.cc",
2282 "bench/conv.h",
2283 "bench/google/conv.h",
2284 "src/xnnpack/AlignedAllocator.h",
2285 ] + MICROKERNEL_BENCHMARK_HDRS,
2286 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2287)
2288
2289xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002290 name = "f16_gemm_bench",
2291 srcs = [
2292 "bench/f16-gemm.cc",
2293 "bench/gemm.h",
2294 "src/xnnpack/AlignedAllocator.h",
2295 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002296 deps = MICROKERNEL_BENCHMARK_DEPS,
2297)
2298
2299xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002300 name = "f16_spmm_bench",
2301 srcs = [
2302 "bench/f16-spmm.cc",
2303 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002304 "src/xnnpack/AlignedAllocator.h",
2305 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002306 deps = MICROKERNEL_BENCHMARK_DEPS,
2307)
2308
2309xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002310 name = "f32_igemm_bench",
2311 srcs = [
2312 "bench/f32-igemm.cc",
2313 "bench/conv.h",
2314 "src/xnnpack/AlignedAllocator.h",
2315 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002316 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317)
2318
2319xnnpack_benchmark(
2320 name = "f32_conv_hwc_bench",
2321 srcs = [
2322 "bench/f32-conv-hwc.cc",
2323 "bench/dconv.h",
2324 "src/xnnpack/AlignedAllocator.h",
2325 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002326 deps = MICROKERNEL_BENCHMARK_DEPS,
2327)
2328
2329xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002330 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07002331 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002332 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07002333 "bench/dconv.h",
2334 "src/xnnpack/AlignedAllocator.h",
2335 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002336 deps = MICROKERNEL_BENCHMARK_DEPS,
2337)
2338
2339xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 name = "f32_dwconv_bench",
2341 srcs = [
2342 "bench/f32-dwconv.cc",
2343 "bench/dwconv.h",
2344 "src/xnnpack/AlignedAllocator.h",
2345 ] + MICROKERNEL_BENCHMARK_HDRS,
2346 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2347)
2348
2349xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002350 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002351 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002352 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002353 "bench/dwconv.h",
2354 "src/xnnpack/AlignedAllocator.h",
2355 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2356 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2357)
2358
2359xnnpack_benchmark(
2360 name = "f32_gemm_bench",
2361 srcs = [
2362 "bench/f32-gemm.cc",
2363 "bench/gemm.h",
2364 "src/xnnpack/AlignedAllocator.h",
2365 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002366 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002367 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002368)
2369
2370xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002371 name = "f32_raddexpminusmax_bench",
2372 srcs = [
2373 "bench/f32-raddexpminusmax.cc",
2374 "src/xnnpack/AlignedAllocator.h",
2375 ] + MICROKERNEL_BENCHMARK_HDRS,
2376 deps = MICROKERNEL_BENCHMARK_DEPS,
2377)
2378
2379xnnpack_benchmark(
2380 name = "f32_raddextexp_bench",
2381 srcs = [
2382 "bench/f32-raddextexp.cc",
2383 "src/xnnpack/AlignedAllocator.h",
2384 ] + MICROKERNEL_BENCHMARK_HDRS,
2385 deps = MICROKERNEL_BENCHMARK_DEPS,
2386)
2387
2388xnnpack_benchmark(
2389 name = "f32_raddstoreexpminusmax_bench",
2390 srcs = [
2391 "bench/f32-raddstoreexpminusmax.cc",
2392 "src/xnnpack/AlignedAllocator.h",
2393 ] + MICROKERNEL_BENCHMARK_HDRS,
2394 deps = MICROKERNEL_BENCHMARK_DEPS,
2395)
2396
2397xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002398 name = "f32_rmax_bench",
2399 srcs = [
2400 "bench/f32-rmax.cc",
2401 "src/xnnpack/AlignedAllocator.h",
2402 ] + MICROKERNEL_BENCHMARK_HDRS,
2403 deps = MICROKERNEL_BENCHMARK_DEPS,
2404)
2405
2406xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002407 name = "f32_sigmoid_bench",
2408 srcs = [
2409 "bench/f32-sigmoid.cc",
2410 "src/xnnpack/AlignedAllocator.h",
2411 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002412 deps = MICROKERNEL_BENCHMARK_DEPS,
2413)
2414
2415xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416 name = "f32_spmm_bench",
2417 srcs = [
2418 "bench/f32-spmm.cc",
2419 "bench/gemm.h",
2420 "src/xnnpack/AlignedAllocator.h",
2421 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002422 deps = MICROKERNEL_BENCHMARK_DEPS,
2423)
2424
2425xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002426 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002427 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002428 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002429 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002430 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002431 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002432)
2433
2434xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002435 name = "f32_vscaleexpminusmax_bench",
2436 srcs = [
2437 "bench/f32-vscaleexpminusmax.cc",
2438 "src/xnnpack/AlignedAllocator.h",
2439 ] + MICROKERNEL_BENCHMARK_HDRS,
2440 deps = MICROKERNEL_BENCHMARK_DEPS,
2441)
2442
2443xnnpack_benchmark(
2444 name = "f32_vscaleextexp_bench",
2445 srcs = [
2446 "bench/f32-vscaleextexp.cc",
2447 "src/xnnpack/AlignedAllocator.h",
2448 ] + MICROKERNEL_BENCHMARK_HDRS,
2449 deps = MICROKERNEL_BENCHMARK_DEPS,
2450)
2451
2452xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002453 name = "f32_im2col_gemm_bench",
2454 srcs = [
2455 "bench/f32-im2col-gemm.cc",
2456 "bench/conv.h",
2457 "src/xnnpack/AlignedAllocator.h",
2458 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2459 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2460)
2461
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002462xnnpack_benchmark(
2463 name = "requantization_bench",
2464 srcs = [
2465 "bench/requantization.cc",
2466 "src/xnnpack/requantization-stubs.h",
2467 "src/xnnpack/AlignedAllocator.h",
2468 ] + MICROKERNEL_BENCHMARK_HDRS,
2469 deps = MICROKERNEL_BENCHMARK_DEPS,
2470)
2471
Marat Dukhanffbf96a2020-05-14 02:59:08 -07002472xnnpack_benchmark(
2473 name = "rounding_bench",
2474 srcs = [
2475 "bench/rounding.cc",
2476 "src/xnnpack/math-stubs.h",
2477 "src/xnnpack/AlignedAllocator.h",
2478 ] + MICROKERNEL_BENCHMARK_HDRS,
2479 deps = MICROKERNEL_BENCHMARK_DEPS,
2480)
2481
Marat Dukhan08c4a432019-10-03 09:29:21 -07002482########################### Benchmarks for operators ###########################
2483
2484xnnpack_benchmark(
2485 name = "add_bench",
2486 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002487 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002488)
2489
2490xnnpack_benchmark(
2491 name = "average_pooling_bench",
2492 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002493 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002494 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002495 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002496)
2497
2498xnnpack_benchmark(
2499 name = "channel_shuffle_bench",
2500 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002501 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002502)
2503
2504xnnpack_benchmark(
2505 name = "convolution_bench",
2506 srcs = ["bench/convolution.cc"],
2507 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002508 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002509 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002510)
2511
2512xnnpack_benchmark(
2513 name = "deconvolution_bench",
2514 srcs = ["bench/deconvolution.cc"],
2515 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002516 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002517 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002518)
2519
2520xnnpack_benchmark(
2521 name = "global_average_pooling_bench",
2522 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002523 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002524)
2525
2526xnnpack_benchmark(
2527 name = "max_pooling_bench",
2528 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002529 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002530)
2531
2532xnnpack_benchmark(
2533 name = "sigmoid_bench",
2534 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002535 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002536 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002537 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538)
2539
2540xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002541 name = "prelu_bench",
2542 srcs = ["bench/prelu.cc"],
2543 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002544 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002545 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002546)
2547
2548xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002549 name = "softmax_bench",
2550 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002551 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002552 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002553 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554)
2555
Marat Dukhanc068bb62019-10-04 13:24:39 -07002556############################# End-to-end benchmarks ############################
2557
2558cc_library(
2559 name = "mobilenet_v1",
2560 srcs = ["models/mobilenet-v1.cc"],
2561 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002562 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002563 linkstatic = True,
2564 deps = [
2565 ":XNNPACK",
2566 "@pthreadpool",
2567 ],
2568)
2569
2570cc_library(
2571 name = "mobilenet_v2",
2572 srcs = ["models/mobilenet-v2.cc"],
2573 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002574 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002575 linkstatic = True,
2576 deps = [
2577 ":XNNPACK",
2578 "@pthreadpool",
2579 ],
2580)
2581
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002582cc_library(
2583 name = "mobilenet_v3_large",
2584 srcs = ["models/mobilenet-v3-large.cc"],
2585 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002586 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002587 linkstatic = True,
2588 deps = [
2589 ":XNNPACK",
2590 "@pthreadpool",
2591 ],
2592)
2593
2594cc_library(
2595 name = "mobilenet_v3_small",
2596 srcs = ["models/mobilenet-v3-small.cc"],
2597 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002598 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002599 linkstatic = True,
2600 deps = [
2601 ":XNNPACK",
2602 "@pthreadpool",
2603 ],
2604)
2605
Marat Dukhanc068bb62019-10-04 13:24:39 -07002606xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002607 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002608 srcs = [
2609 "bench/f32-dwconv-e2e.cc",
2610 "bench/end2end.h",
2611 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002612 deps = MICROKERNEL_BENCHMARK_DEPS + [
2613 ":XNNPACK",
2614 ":mobilenet_v1",
2615 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002616 ":mobilenet_v3_large",
2617 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002618 ],
2619)
2620
2621xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002622 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002623 srcs = [
2624 "bench/f32-gemm-e2e.cc",
2625 "bench/end2end.h",
2626 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002627 deps = MICROKERNEL_BENCHMARK_DEPS + [
2628 ":XNNPACK",
2629 ":mobilenet_v1",
2630 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002631 ":mobilenet_v3_large",
2632 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002633 ],
2634)
2635
2636xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002637 name = "end2end_bench",
2638 srcs = ["bench/end2end.cc"],
2639 deps = [
2640 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002641 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002642 ":mobilenet_v1",
2643 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002644 ":mobilenet_v3_large",
2645 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002646 "@pthreadpool",
2647 ],
2648)
2649
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002650#################### Accuracy evaluation for math functions ####################
2651
2652xnnpack_benchmark(
2653 name = "f32_exp_eval",
2654 srcs = [
2655 "eval/f32-exp.cc",
2656 "src/xnnpack/AlignedAllocator.h",
2657 ] + ACCURACY_EVAL_HDRS,
2658 deps = ACCURACY_EVAL_DEPS,
2659)
2660
Marat Dukhan515c9772019-10-17 18:07:57 -07002661xnnpack_benchmark(
2662 name = "f32_expminus_eval",
2663 srcs = [
2664 "eval/f32-expminus.cc",
2665 "src/xnnpack/AlignedAllocator.h",
2666 ] + ACCURACY_EVAL_HDRS,
2667 deps = ACCURACY_EVAL_DEPS,
2668)
2669
Marat Dukhan98ba4412019-10-23 02:14:28 -07002670xnnpack_benchmark(
2671 name = "f32_extexp_eval",
2672 srcs = [
2673 "eval/f32-extexp.cc",
2674 "src/xnnpack/AlignedAllocator.h",
2675 ] + ACCURACY_EVAL_HDRS,
2676 deps = ACCURACY_EVAL_DEPS,
2677)
2678
Marat Dukhan8853b822020-05-07 12:19:01 -07002679xnnpack_unit_test(
2680 name = "f32_roundne_eval",
2681 srcs = [
2682 "eval/f32-roundne.cc",
2683 "src/xnnpack/AlignedAllocator.h",
2684 "src/xnnpack/math-stubs.h",
2685 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002686 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002687 deps = MICROKERNEL_TEST_DEPS,
2688)
2689
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002690xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002691 name = "f32_roundd_eval",
2692 srcs = [
2693 "eval/f32-roundd.cc",
2694 "src/xnnpack/AlignedAllocator.h",
2695 "src/xnnpack/math-stubs.h",
2696 ] + MICROKERNEL_TEST_HDRS,
2697 automatic = False,
2698 deps = MICROKERNEL_TEST_DEPS,
2699)
2700
2701xnnpack_unit_test(
2702 name = "f32_roundu_eval",
2703 srcs = [
2704 "eval/f32-roundu.cc",
2705 "src/xnnpack/AlignedAllocator.h",
2706 "src/xnnpack/math-stubs.h",
2707 ] + MICROKERNEL_TEST_HDRS,
2708 automatic = False,
2709 deps = MICROKERNEL_TEST_DEPS,
2710)
2711
2712xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002713 name = "f32_roundz_eval",
2714 srcs = [
2715 "eval/f32-roundz.cc",
2716 "src/xnnpack/AlignedAllocator.h",
2717 "src/xnnpack/math-stubs.h",
2718 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002719 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002720 deps = MICROKERNEL_TEST_DEPS,
2721)
2722
Marat Dukhan346a9e52019-11-15 09:06:30 -08002723xnnpack_benchmark(
2724 name = "f32_sigmoid_eval",
2725 srcs = [
2726 "eval/f32-sigmoid.cc",
2727 "src/xnnpack/AlignedAllocator.h",
2728 ] + ACCURACY_EVAL_HDRS,
2729 deps = ACCURACY_EVAL_DEPS,
2730)
2731
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732######################### Unit tests for micro-kernels #########################
2733
2734xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002735 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002736 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002737 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002738 "test/gemm-microkernel-tester.h",
2739 "src/xnnpack/AlignedAllocator.h",
2740 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2741 deps = MICROKERNEL_TEST_DEPS,
2742)
2743
2744xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002745 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002746 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002747 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002748 "test/spmm-microkernel-tester.h",
2749 "src/xnnpack/AlignedAllocator.h",
2750 ] + MICROKERNEL_TEST_HDRS,
2751 deps = MICROKERNEL_TEST_DEPS,
2752)
2753
2754xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002755 name = "f32_argmaxpool_test",
2756 srcs = [
2757 "test/f32-argmaxpool.cc",
2758 "test/argmaxpool-microkernel-tester.h",
2759 "src/xnnpack/AlignedAllocator.h",
2760 ] + MICROKERNEL_TEST_HDRS,
2761 deps = MICROKERNEL_TEST_DEPS,
2762)
2763
2764xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002765 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002766 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002767 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002768 "test/avgpool-microkernel-tester.h",
2769 "src/xnnpack/AlignedAllocator.h",
2770 ] + MICROKERNEL_TEST_HDRS,
2771 deps = MICROKERNEL_TEST_DEPS,
2772)
2773
2774xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002775 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002776 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002777 "test/f32-ibilinear.cc",
2778 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002779 "src/xnnpack/AlignedAllocator.h",
2780 ] + MICROKERNEL_TEST_HDRS,
2781 deps = MICROKERNEL_TEST_DEPS,
2782)
2783
2784xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002785 name = "f16_clamp_test",
2786 srcs = [
2787 "test/f16-clamp.cc",
2788 "test/clamp-microkernel-tester.h",
2789 ] + MICROKERNEL_TEST_HDRS,
2790 deps = MICROKERNEL_TEST_DEPS,
2791)
2792
2793xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002794 name = "f32_clamp_test",
2795 srcs = [
2796 "test/f32-clamp.cc",
2797 "test/clamp-microkernel-tester.h",
2798 ] + MICROKERNEL_TEST_HDRS,
2799 deps = MICROKERNEL_TEST_DEPS,
2800)
2801
2802xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002803 name = "f32_igemm_test",
2804 srcs = [
2805 "test/f32-igemm.cc",
2806 "test/gemm-microkernel-tester.h",
2807 "src/xnnpack/AlignedAllocator.h",
2808 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2809 deps = MICROKERNEL_TEST_DEPS,
2810)
2811
2812xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002813 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002815 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002816 "test/gemm-microkernel-tester.h",
2817 "src/xnnpack/AlignedAllocator.h",
2818 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002819 # Work-around for "too many sections" error
2820 mingw_copts = ["-Wa,-mbig-obj"],
2821 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002822 deps = MICROKERNEL_TEST_DEPS,
2823)
2824
2825xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002826 name = "f16_igemm_minmax_test",
2827 srcs = [
2828 "test/f16-igemm-minmax.cc",
2829 "test/gemm-microkernel-tester.h",
2830 "src/xnnpack/AlignedAllocator.h",
2831 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2832 deps = MICROKERNEL_TEST_DEPS,
2833)
2834
2835xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002836 name = "f32_conv_hwc_test",
2837 srcs = [
2838 "test/f32-conv-hwc.cc",
2839 "test/conv-hwc-microkernel-tester.h",
2840 "src/xnnpack/AlignedAllocator.h",
2841 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2842 deps = MICROKERNEL_TEST_DEPS,
2843)
2844
2845xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002846 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002847 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002848 "test/f32-conv-hwc2chw.cc",
2849 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002850 "src/xnnpack/AlignedAllocator.h",
2851 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2852 deps = MICROKERNEL_TEST_DEPS,
2853)
2854
2855xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002856 name = "f32_dwconv_test",
2857 srcs = [
2858 "test/f32-dwconv.cc",
2859 "test/dwconv-microkernel-tester.h",
2860 "src/xnnpack/AlignedAllocator.h",
2861 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2862 deps = MICROKERNEL_TEST_DEPS,
2863)
2864
2865xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002866 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002868 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002869 "test/dwconv-microkernel-tester.h",
2870 "src/xnnpack/AlignedAllocator.h",
2871 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2872 deps = MICROKERNEL_TEST_DEPS,
2873)
2874
2875xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002876 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002877 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002878 "test/f32-dwconv-chw.cc",
2879 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 "src/xnnpack/AlignedAllocator.h",
2881 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2882 deps = MICROKERNEL_TEST_DEPS,
2883)
2884
2885xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002886 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002888 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002889 "test/gavgpool-microkernel-tester.h",
2890 "src/xnnpack/AlignedAllocator.h",
2891 ] + MICROKERNEL_TEST_HDRS,
2892 deps = MICROKERNEL_TEST_DEPS,
2893)
2894
2895xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002896 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002898 "test/f32-gavgpool-cw.cc",
2899 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002900 "src/xnnpack/AlignedAllocator.h",
2901 ] + MICROKERNEL_TEST_HDRS,
2902 deps = MICROKERNEL_TEST_DEPS,
2903)
2904
2905xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002906 name = "f32_gemm_test",
2907 srcs = [
2908 "test/f32-gemm.cc",
2909 "test/gemm-microkernel-tester.h",
2910 "src/xnnpack/AlignedAllocator.h",
2911 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2912 deps = MICROKERNEL_TEST_DEPS,
2913)
2914
2915xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002916 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002917 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002918 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 "test/gemm-microkernel-tester.h",
2920 "src/xnnpack/AlignedAllocator.h",
2921 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2922 deps = MICROKERNEL_TEST_DEPS,
2923)
2924
2925xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002926 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002927 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002928 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002929 "test/gemm-microkernel-tester.h",
2930 "src/xnnpack/AlignedAllocator.h",
2931 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2932 deps = MICROKERNEL_TEST_DEPS,
2933)
2934
2935xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002936 name = "f16_hswish_test",
2937 srcs = [
2938 "test/f16-hswish.cc",
2939 "test/hswish-microkernel-tester.h",
2940 ] + MICROKERNEL_TEST_HDRS,
2941 deps = MICROKERNEL_TEST_DEPS,
2942)
2943
2944xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 name = "f32_hswish_test",
2946 srcs = [
2947 "test/f32-hswish.cc",
2948 "test/hswish-microkernel-tester.h",
2949 ] + MICROKERNEL_TEST_HDRS,
2950 deps = MICROKERNEL_TEST_DEPS,
2951)
2952
2953xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002954 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002955 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002956 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002957 "test/maxpool-microkernel-tester.h",
2958 ] + MICROKERNEL_TEST_HDRS,
2959 deps = MICROKERNEL_TEST_DEPS,
2960)
2961
2962xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002963 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002965 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 "test/avgpool-microkernel-tester.h",
2967 "src/xnnpack/AlignedAllocator.h",
2968 ] + MICROKERNEL_TEST_HDRS,
2969 deps = MICROKERNEL_TEST_DEPS,
2970)
2971
2972xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002973 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002974 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002975 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976 "test/gemm-microkernel-tester.h",
2977 "src/xnnpack/AlignedAllocator.h",
2978 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2979 deps = MICROKERNEL_TEST_DEPS,
2980)
2981
2982xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002983 name = "f16_prelu_test",
2984 srcs = [
2985 "test/f16-prelu.cc",
2986 "test/prelu-microkernel-tester.h",
2987 "src/xnnpack/AlignedAllocator.h",
2988 ] + MICROKERNEL_TEST_HDRS,
2989 deps = MICROKERNEL_TEST_DEPS,
2990)
2991
2992xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 name = "f32_prelu_test",
2994 srcs = [
2995 "test/f32-prelu.cc",
2996 "test/prelu-microkernel-tester.h",
2997 "src/xnnpack/AlignedAllocator.h",
2998 ] + MICROKERNEL_TEST_HDRS,
2999 deps = MICROKERNEL_TEST_DEPS,
3000)
3001
3002xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003003 name = "f32_raddexpminusmax_test",
3004 srcs = [
3005 "test/f32-raddexpminusmax.cc",
3006 "test/raddexpminusmax-microkernel-tester.h",
3007 ] + MICROKERNEL_TEST_HDRS,
3008 deps = MICROKERNEL_TEST_DEPS,
3009)
3010
3011xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003012 name = "f32_raddextexp_test",
3013 srcs = [
3014 "test/f32-raddextexp.cc",
3015 "test/raddextexp-microkernel-tester.h",
3016 ] + MICROKERNEL_TEST_HDRS,
3017 deps = MICROKERNEL_TEST_DEPS,
3018)
3019
3020xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003021 name = "f32_raddstoreexpminusmax_test",
3022 srcs = [
3023 "test/f32-raddstoreexpminusmax.cc",
3024 "test/raddstoreexpminusmax-microkernel-tester.h",
3025 ] + MICROKERNEL_TEST_HDRS,
3026 deps = MICROKERNEL_TEST_DEPS,
3027)
3028
3029xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 name = "f32_rmax_test",
3031 srcs = [
3032 "test/f32-rmax.cc",
3033 "test/rmax-microkernel-tester.h",
3034 ] + MICROKERNEL_TEST_HDRS,
3035 deps = MICROKERNEL_TEST_DEPS,
3036)
3037
3038xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003039 name = "f32_sigmoid_test",
3040 srcs = [
3041 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003042 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003043 ] + MICROKERNEL_TEST_HDRS,
3044 deps = MICROKERNEL_TEST_DEPS,
3045)
3046
3047xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003048 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003050 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003051 "test/spmm-microkernel-tester.h",
3052 "src/xnnpack/AlignedAllocator.h",
3053 ] + MICROKERNEL_TEST_HDRS,
3054 deps = MICROKERNEL_TEST_DEPS,
3055)
3056
3057xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003058 name = "f16_vadd_minmax_test",
3059 srcs = [
3060 "test/f16-vadd-minmax.cc",
3061 "test/vbinary-microkernel-tester.h",
3062 ] + MICROKERNEL_TEST_HDRS,
3063 deps = MICROKERNEL_TEST_DEPS,
3064)
3065
3066xnnpack_unit_test(
3067 name = "f16_vaddc_minmax_test",
3068 srcs = [
3069 "test/f16-vaddc-minmax.cc",
3070 "test/vbinaryc-microkernel-tester.h",
3071 ] + MICROKERNEL_TEST_HDRS,
3072 deps = MICROKERNEL_TEST_DEPS,
3073)
3074
3075xnnpack_unit_test(
3076 name = "f16_vdiv_minmax_test",
3077 srcs = [
3078 "test/f16-vdiv-minmax.cc",
3079 "test/vbinary-microkernel-tester.h",
3080 ] + MICROKERNEL_TEST_HDRS,
3081 deps = MICROKERNEL_TEST_DEPS,
3082)
3083
3084xnnpack_unit_test(
3085 name = "f16_vdivc_minmax_test",
3086 srcs = [
3087 "test/f16-vdivc-minmax.cc",
3088 "test/vbinaryc-microkernel-tester.h",
3089 ] + MICROKERNEL_TEST_HDRS,
3090 deps = MICROKERNEL_TEST_DEPS,
3091)
3092
3093xnnpack_unit_test(
3094 name = "f16_vrdivc_minmax_test",
3095 srcs = [
3096 "test/f16-vrdivc-minmax.cc",
3097 "test/vbinaryc-microkernel-tester.h",
3098 ] + MICROKERNEL_TEST_HDRS,
3099 deps = MICROKERNEL_TEST_DEPS,
3100)
3101
3102xnnpack_unit_test(
3103 name = "f16_vmax_test",
3104 srcs = [
3105 "test/f16-vmax.cc",
3106 "test/vbinary-microkernel-tester.h",
3107 ] + MICROKERNEL_TEST_HDRS,
3108 deps = MICROKERNEL_TEST_DEPS,
3109)
3110
3111xnnpack_unit_test(
3112 name = "f16_vmaxc_test",
3113 srcs = [
3114 "test/f16-vmaxc.cc",
3115 "test/vbinaryc-microkernel-tester.h",
3116 ] + MICROKERNEL_TEST_HDRS,
3117 deps = MICROKERNEL_TEST_DEPS,
3118)
3119
3120xnnpack_unit_test(
3121 name = "f16_vmin_test",
3122 srcs = [
3123 "test/f16-vmin.cc",
3124 "test/vbinary-microkernel-tester.h",
3125 ] + MICROKERNEL_TEST_HDRS,
3126 deps = MICROKERNEL_TEST_DEPS,
3127)
3128
3129xnnpack_unit_test(
3130 name = "f16_vminc_test",
3131 srcs = [
3132 "test/f16-vminc.cc",
3133 "test/vbinaryc-microkernel-tester.h",
3134 ] + MICROKERNEL_TEST_HDRS,
3135 deps = MICROKERNEL_TEST_DEPS,
3136)
3137
3138xnnpack_unit_test(
3139 name = "f16_vmul_minmax_test",
3140 srcs = [
3141 "test/f16-vmul-minmax.cc",
3142 "test/vbinary-microkernel-tester.h",
3143 ] + MICROKERNEL_TEST_HDRS,
3144 deps = MICROKERNEL_TEST_DEPS,
3145)
3146
3147xnnpack_unit_test(
3148 name = "f16_vmulc_minmax_test",
3149 srcs = [
3150 "test/f16-vmulc-minmax.cc",
3151 "test/vbinaryc-microkernel-tester.h",
3152 ] + MICROKERNEL_TEST_HDRS,
3153 deps = MICROKERNEL_TEST_DEPS,
3154)
3155
3156xnnpack_unit_test(
3157 name = "f16_vsub_minmax_test",
3158 srcs = [
3159 "test/f16-vsub-minmax.cc",
3160 "test/vbinary-microkernel-tester.h",
3161 ] + MICROKERNEL_TEST_HDRS,
3162 deps = MICROKERNEL_TEST_DEPS,
3163)
3164
3165xnnpack_unit_test(
3166 name = "f16_vsubc_minmax_test",
3167 srcs = [
3168 "test/f16-vsubc-minmax.cc",
3169 "test/vbinaryc-microkernel-tester.h",
3170 ] + MICROKERNEL_TEST_HDRS,
3171 deps = MICROKERNEL_TEST_DEPS,
3172)
3173
3174xnnpack_unit_test(
3175 name = "f16_vrsubc_minmax_test",
3176 srcs = [
3177 "test/f16-vrsubc-minmax.cc",
3178 "test/vbinaryc-microkernel-tester.h",
3179 ] + MICROKERNEL_TEST_HDRS,
3180 deps = MICROKERNEL_TEST_DEPS,
3181)
3182
3183xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003184 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003185 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003186 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003187 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003188 ] + MICROKERNEL_TEST_HDRS,
3189 deps = MICROKERNEL_TEST_DEPS,
3190)
3191
3192xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003193 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003194 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003195 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003196 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197 ] + MICROKERNEL_TEST_HDRS,
3198 deps = MICROKERNEL_TEST_DEPS,
3199)
3200
3201xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003202 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003203 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003204 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003205 "test/vbinary-microkernel-tester.h",
3206 ] + MICROKERNEL_TEST_HDRS,
3207 deps = MICROKERNEL_TEST_DEPS,
3208)
3209
3210xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003211 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003212 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003213 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003214 "test/vbinaryc-microkernel-tester.h",
3215 ] + MICROKERNEL_TEST_HDRS,
3216 deps = MICROKERNEL_TEST_DEPS,
3217)
3218
3219xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003220 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003221 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003222 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003223 "test/vbinaryc-microkernel-tester.h",
3224 ] + MICROKERNEL_TEST_HDRS,
3225 deps = MICROKERNEL_TEST_DEPS,
3226)
3227
3228xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003229 name = "f32_vmax_test",
3230 srcs = [
3231 "test/f32-vmax.cc",
3232 "test/vbinary-microkernel-tester.h",
3233 ] + MICROKERNEL_TEST_HDRS,
3234 deps = MICROKERNEL_TEST_DEPS,
3235)
3236
3237xnnpack_unit_test(
3238 name = "f32_vmaxc_test",
3239 srcs = [
3240 "test/f32-vmaxc.cc",
3241 "test/vbinaryc-microkernel-tester.h",
3242 ] + MICROKERNEL_TEST_HDRS,
3243 deps = MICROKERNEL_TEST_DEPS,
3244)
3245
3246xnnpack_unit_test(
3247 name = "f32_vmin_test",
3248 srcs = [
3249 "test/f32-vmin.cc",
3250 "test/vbinary-microkernel-tester.h",
3251 ] + MICROKERNEL_TEST_HDRS,
3252 deps = MICROKERNEL_TEST_DEPS,
3253)
3254
3255xnnpack_unit_test(
3256 name = "f32_vminc_test",
3257 srcs = [
3258 "test/f32-vminc.cc",
3259 "test/vbinaryc-microkernel-tester.h",
3260 ] + MICROKERNEL_TEST_HDRS,
3261 deps = MICROKERNEL_TEST_DEPS,
3262)
3263
3264xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003265 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003266 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003267 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003268 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003269 ] + MICROKERNEL_TEST_HDRS,
3270 deps = MICROKERNEL_TEST_DEPS,
3271)
3272
3273xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003274 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003275 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003276 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003277 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278 ] + MICROKERNEL_TEST_HDRS,
3279 deps = MICROKERNEL_TEST_DEPS,
3280)
3281
3282xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003283 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003284 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003285 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286 "test/vmulcaddc-microkernel-tester.h",
3287 "src/xnnpack/AlignedAllocator.h",
3288 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3289 deps = MICROKERNEL_TEST_DEPS,
3290)
3291
3292xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003293 name = "f32_vscale_test",
3294 srcs = [
3295 "test/f32-vscale.cc",
3296 "test/vscale-microkernel-tester.h",
3297 ] + MICROKERNEL_TEST_HDRS,
3298 deps = MICROKERNEL_TEST_DEPS,
3299)
3300
3301xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003302 name = "f32_vscaleexpminusmax_test",
3303 srcs = [
3304 "test/f32-vscaleexpminusmax.cc",
3305 "test/vscaleexpminusmax-microkernel-tester.h",
3306 ] + MICROKERNEL_TEST_HDRS,
3307 deps = MICROKERNEL_TEST_DEPS,
3308)
3309
3310xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003311 name = "f32_vscaleextexp_test",
3312 srcs = [
3313 "test/f32-vscaleextexp.cc",
3314 "test/vscaleextexp-microkernel-tester.h",
3315 ] + MICROKERNEL_TEST_HDRS,
3316 deps = MICROKERNEL_TEST_DEPS,
3317)
3318
3319xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003320 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003321 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003322 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003323 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003324 ] + MICROKERNEL_TEST_HDRS,
3325 deps = MICROKERNEL_TEST_DEPS,
3326)
3327
3328xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003329 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003330 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003331 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003332 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003333 ] + MICROKERNEL_TEST_HDRS,
3334 deps = MICROKERNEL_TEST_DEPS,
3335)
3336
3337xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003338 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003339 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003340 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003341 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003342 ] + MICROKERNEL_TEST_HDRS,
3343 deps = MICROKERNEL_TEST_DEPS,
3344)
3345
3346xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003347 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003348 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003349 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 "test/avgpool-microkernel-tester.h",
3351 "src/xnnpack/AlignedAllocator.h",
3352 ] + MICROKERNEL_TEST_HDRS,
3353 deps = MICROKERNEL_TEST_DEPS,
3354)
3355
3356xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003357 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003358 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003359 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360 "test/gemm-microkernel-tester.h",
3361 "src/xnnpack/AlignedAllocator.h",
3362 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3363 deps = MICROKERNEL_TEST_DEPS,
3364)
3365
3366xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003367 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003368 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003369 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370 "test/dwconv-microkernel-tester.h",
3371 "src/xnnpack/AlignedAllocator.h",
3372 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3373 deps = MICROKERNEL_TEST_DEPS,
3374)
3375
3376xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003377 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003379 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 "test/gavgpool-microkernel-tester.h",
3381 "src/xnnpack/AlignedAllocator.h",
3382 ] + MICROKERNEL_TEST_HDRS,
3383 deps = MICROKERNEL_TEST_DEPS,
3384)
3385
3386xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003387 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003388 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003389 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003390 "test/gemm-microkernel-tester.h",
3391 "src/xnnpack/AlignedAllocator.h",
3392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3393 deps = MICROKERNEL_TEST_DEPS,
3394)
3395
3396xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003397 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003398 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003399 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003400 "test/vadd-microkernel-tester.h",
3401 ] + MICROKERNEL_TEST_HDRS,
3402 deps = MICROKERNEL_TEST_DEPS,
3403)
3404
3405xnnpack_unit_test(
3406 name = "u8_clamp_test",
3407 srcs = [
3408 "test/u8-clamp.cc",
3409 "test/clamp-microkernel-tester.h",
3410 ] + MICROKERNEL_TEST_HDRS,
3411 deps = MICROKERNEL_TEST_DEPS,
3412)
3413
3414xnnpack_unit_test(
3415 name = "u8_lut32norm_test",
3416 srcs = [
3417 "test/u8-lut32norm.cc",
3418 "test/lut-norm-microkernel-tester.h",
3419 ] + MICROKERNEL_TEST_HDRS,
3420 deps = MICROKERNEL_TEST_DEPS,
3421)
3422
3423xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003424 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003426 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 "test/maxpool-microkernel-tester.h",
3428 ] + MICROKERNEL_TEST_HDRS,
3429 deps = MICROKERNEL_TEST_DEPS,
3430)
3431
3432xnnpack_unit_test(
3433 name = "u8_rmax_test",
3434 srcs = [
3435 "test/u8-rmax.cc",
3436 "test/rmax-microkernel-tester.h",
3437 ] + MICROKERNEL_TEST_HDRS,
3438 deps = MICROKERNEL_TEST_DEPS,
3439)
3440
3441xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07003442 name = "x32_fill_test",
3443 srcs = [
3444 "test/x32-fill.cc",
3445 "test/fill-microkernel-tester.h",
3446 ] + MICROKERNEL_TEST_HDRS,
3447 deps = MICROKERNEL_TEST_DEPS,
3448)
3449
3450xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003451 name = "x32_packx_test",
3452 srcs = [
3453 "test/x32-packx.cc",
3454 "test/pack-microkernel-tester.h",
3455 "src/xnnpack/AlignedAllocator.h",
3456 ] + MICROKERNEL_TEST_HDRS,
3457 deps = MICROKERNEL_TEST_DEPS,
3458)
3459
3460xnnpack_unit_test(
3461 name = "x32_pad_test",
3462 srcs = [
3463 "test/x32-pad.cc",
3464 "test/pad-microkernel-tester.h",
3465 ] + MICROKERNEL_TEST_HDRS,
3466 deps = MICROKERNEL_TEST_DEPS,
3467)
3468
3469xnnpack_unit_test(
3470 name = "x32_unpool_test",
3471 srcs = [
3472 "test/x32-unpool.cc",
3473 "test/unpool-microkernel-tester.h",
3474 ] + MICROKERNEL_TEST_HDRS,
3475 deps = MICROKERNEL_TEST_DEPS,
3476)
3477
3478xnnpack_unit_test(
3479 name = "x32_zip_test",
3480 srcs = [
3481 "test/x32-zip.cc",
3482 "test/zip-microkernel-tester.h",
3483 ] + MICROKERNEL_TEST_HDRS,
3484 deps = MICROKERNEL_TEST_DEPS,
3485)
3486
3487xnnpack_unit_test(
3488 name = "x8_lut_test",
3489 srcs = [
3490 "test/x8-lut.cc",
3491 "test/lut-microkernel-tester.h",
3492 ] + MICROKERNEL_TEST_HDRS,
3493 deps = MICROKERNEL_TEST_DEPS,
3494)
3495
3496xnnpack_unit_test(
3497 name = "x8_zip_test",
3498 srcs = [
3499 "test/x8-zip.cc",
3500 "test/zip-microkernel-tester.h",
3501 ] + MICROKERNEL_TEST_HDRS,
3502 deps = MICROKERNEL_TEST_DEPS,
3503)
3504
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003505xnnpack_unit_test(
3506 name = "requantization_test",
3507 srcs = [
3508 "src/xnnpack/requantization-stubs.h",
3509 "test/requantization.cc",
3510 "test/requantization-tester.h",
3511 ] + MICROKERNEL_TEST_HDRS,
3512 deps = MICROKERNEL_TEST_DEPS,
3513)
3514
Marat Dukhan20c3b922020-03-10 03:45:06 -07003515########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516
3517xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003518 name = "operator_size_test",
3519 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003520 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521)
3522
Marat Dukhan20c3b922020-03-10 03:45:06 -07003523xnnpack_binary(
3524 name = "subgraph_size_test",
3525 srcs = ["test/subgraph-size.c"],
3526 deps = [":XNNPACK"],
3527)
3528
3529########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530
3531xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003532 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003533 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003534 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003535 "test/add-operator-tester.h",
3536 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003537 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538)
3539
3540xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003541 name = "add_nd_test",
3542 srcs = [
3543 "test/add-nd.cc",
3544 "test/binary-elementwise-operator-tester.h",
3545 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003546 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003547)
3548
3549xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003550 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003551 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003552 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 "test/argmax-pooling-operator-tester.h",
3554 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003555 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556)
3557
3558xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003559 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003560 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003561 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003562 "test/average-pooling-operator-tester.h",
3563 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003564 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003565)
3566
3567xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003568 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003570 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571 "test/channel-pad-operator-tester.h",
3572 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003573 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574)
3575
3576xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003577 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003578 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003579 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580 "test/channel-shuffle-operator-tester.h",
3581 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003582 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583)
3584
3585xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003586 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003588 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589 "test/clamp-operator-tester.h",
3590 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003591 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592)
3593
3594xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003595 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003596 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003597 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003598 "test/convolution-operator-tester.h",
3599 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003600 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601)
3602
3603xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003604 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003606 "test/convolution-nchw.cc",
3607 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003608 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003609 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003610)
3611
3612xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003613 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003614 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003615 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003616 "test/deconvolution-operator-tester.h",
3617 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003618 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619)
3620
3621xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003622 name = "divide_nd_test",
3623 srcs = [
3624 "test/binary-elementwise-operator-tester.h",
3625 "test/divide-nd.cc",
3626 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003627 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003628)
3629
3630xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003631 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003633 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 "test/fully-connected-operator-tester.h",
3635 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003636 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637)
3638
3639xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003640 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003642 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 "test/global-average-pooling-operator-tester.h",
3644 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003645 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646)
3647
3648xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003649 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003651 "test/global-average-pooling-ncw.cc",
3652 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003654 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003655)
3656
3657xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003658 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003660 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003661 "test/hardswish-operator-tester.h",
3662 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003663 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664)
3665
3666xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003667 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003668 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003669 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670 "test/leaky-relu-operator-tester.h",
3671 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003672 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003673)
3674
3675xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003676 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003677 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003678 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003679 "test/max-pooling-operator-tester.h",
3680 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003681 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003682)
3683
3684xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003685 name = "maximum_nd_test",
3686 srcs = [
3687 "test/binary-elementwise-operator-tester.h",
3688 "test/maximum-nd.cc",
3689 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003690 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003691)
3692
3693xnnpack_unit_test(
3694 name = "minimum_nd_test",
3695 srcs = [
3696 "test/binary-elementwise-operator-tester.h",
3697 "test/minimum-nd.cc",
3698 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003699 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003700)
3701
3702xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003703 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003704 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003705 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003706 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003707 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003708 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003709)
3710
3711xnnpack_unit_test(
Marat Dukhan4662b192020-05-21 15:52:03 -07003712 name = "pad_nd_test",
3713 srcs = [
3714 "test/pad-nd.cc",
3715 "test/pad-operator-tester.h",
3716 ],
3717 deps = OPERATOR_TEST_DEPS,
3718)
3719
3720xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003721 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003722 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003723 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003724 "test/prelu-operator-tester.h",
3725 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003726 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727)
3728
3729xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003730 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003731 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003732 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003733 "test/resize-bilinear-operator-tester.h",
3734 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003735 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003736)
3737
3738xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003739 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003740 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003741 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742 "test/sigmoid-operator-tester.h",
3743 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003744 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745)
3746
3747xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003748 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003750 "test/softmax-nc.cc",
3751 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003752 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003753 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003754)
3755
3756xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003757 name = "subtract_nd_test",
3758 srcs = [
3759 "test/binary-elementwise-operator-tester.h",
3760 "test/subtract-nd.cc",
3761 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003762 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003763)
3764
3765xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003766 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003767 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003768 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 "test/unpooling-operator-tester.h",
3770 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003771 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772)
3773
Chao Mei6ddfc602020-05-13 22:29:36 -07003774############################### Misc unit tests ###############################
3775
3776xnnpack_unit_test(
3777 name = "memory_planner_test",
3778 srcs = [
3779 "test/memory-planner-test.cc",
3780 ],
3781 deps = [
3782 ":XNNPACK",
3783 ":memory_planner",
3784 ],
3785)
3786
Marat Dukhan08c4a432019-10-03 09:29:21 -07003787############################# Build configurations #############################
3788
Marat Dukhanb8642352019-10-30 15:43:02 -07003789# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003790config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003791 name = "xnn_enable_assembly_explicit_true",
3792 define_values = {"xnn_enable_assembly": "true"},
3793)
3794
3795# Disables usage of assembly kernels.
3796config_setting(
3797 name = "xnn_enable_assembly_explicit_false",
3798 define_values = {"xnn_enable_assembly": "false"},
3799)
3800
Marat Dukhan05702cf2020-03-26 15:41:33 -07003801# Disables usage of HMP-aware optimizations.
3802config_setting(
3803 name = "xnn_enable_hmp_explicit_false",
3804 define_values = {"xnn_enable_hmp": "false"},
3805)
3806
Chao Mei6ddfc602020-05-13 22:29:36 -07003807# Enable usage of optimized memory allocation
3808config_setting(
3809 name = "xnn_enable_memopt_explicit_true",
3810 define_values = {"xnnpack_enable_memopt": "true"},
3811)
3812
3813# Disable usage of optimized memory allocation
3814config_setting(
3815 name = "xnn_enable_memopt_explicit_false",
3816 define_values = {"xnnpack_enable_memopt": "false"},
3817)
3818
Marat Dukhanb8642352019-10-30 15:43:02 -07003819# Builds with -c dbg
3820config_setting(
3821 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003823 "compilation_mode": "dbg",
3824 },
3825)
3826
3827# Builds with -c opt
3828config_setting(
3829 name = "optimized_build",
3830 values = {
3831 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832 },
3833)
3834
3835config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003836 name = "linux_k8",
3837 values = {"cpu": "k8"},
3838)
3839
3840config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003841 name = "linux_arm",
3842 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003843)
3844
3845config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003846 name = "linux_armhf",
3847 values = {"cpu": "armhf"},
3848)
3849
3850config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003851 name = "linux_aarch64",
3852 values = {"cpu": "aarch64"},
3853)
3854
3855config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003856 name = "android",
3857 values = {"crosstool_top": "//external:android/crosstool"},
3858)
3859
3860config_setting(
3861 name = "android_armv7",
3862 values = {
3863 "crosstool_top": "//external:android/crosstool",
3864 "cpu": "armeabi-v7a",
3865 },
3866)
3867
3868config_setting(
3869 name = "android_arm64",
3870 values = {
3871 "crosstool_top": "//external:android/crosstool",
3872 "cpu": "arm64-v8a",
3873 },
3874)
3875
3876config_setting(
3877 name = "android_x86",
3878 values = {
3879 "crosstool_top": "//external:android/crosstool",
3880 "cpu": "x86",
3881 },
3882)
3883
3884config_setting(
3885 name = "android_x86_64",
3886 values = {
3887 "crosstool_top": "//external:android/crosstool",
3888 "cpu": "x86_64",
3889 },
3890)
3891
3892config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003893 name = "windows_x86_64",
3894 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003895)
3896
3897config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003898 name = "windows_x86_64_clang",
3899 values = {
3900 "compiler": "clang-cl",
3901 "cpu": "x64_windows",
3902 },
3903)
3904
3905config_setting(
3906 name = "windows_x86_64_mingw",
3907 values = {
3908 "compiler": "mingw-gcc",
3909 "cpu": "x64_windows",
3910 },
3911)
3912
3913config_setting(
3914 name = "windows_x86_64_msys",
3915 values = {
3916 "compiler": "msys-gcc",
3917 "cpu": "x64_windows",
3918 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003919)
3920
3921config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003922 name = "macos_x86_64",
3923 values = {
3924 "apple_platform_type": "macos",
3925 "cpu": "darwin",
3926 },
3927)
3928
3929config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003930 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003931 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932)
3933
3934config_setting(
3935 name = "emscripten_wasm",
3936 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003937 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003938 "cpu": "wasm",
3939 },
3940)
3941
3942config_setting(
3943 name = "emscripten_wasmsimd",
3944 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003945 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003947 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003948 },
3949)
3950
3951config_setting(
3952 name = "emscripten_asmjs",
3953 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003954 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003955 "cpu": "asmjs",
3956 },
3957)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003958
3959config_setting(
3960 name = "ios_armv7",
3961 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003962 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003963 "cpu": "ios_armv7",
3964 },
3965)
3966
3967config_setting(
3968 name = "ios_arm64",
3969 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003970 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003971 "cpu": "ios_arm64",
3972 },
3973)
3974
3975config_setting(
3976 name = "ios_arm64e",
3977 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003978 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003979 "cpu": "ios_arm64e",
3980 },
3981)
3982
3983config_setting(
3984 name = "ios_x86",
3985 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003986 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003987 "cpu": "ios_i386",
3988 },
3989)
3990
3991config_setting(
3992 name = "ios_x86_64",
3993 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003994 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003995 "cpu": "ios_x86_64",
3996 },
3997)
3998
3999config_setting(
4000 name = "watchos_armv7k",
4001 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004002 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004003 "cpu": "watchos_armv7k",
4004 },
4005)
4006
4007config_setting(
4008 name = "watchos_arm64_32",
4009 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004010 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004011 "cpu": "watchos_arm64_32",
4012 },
4013)
4014
4015config_setting(
4016 name = "watchos_x86",
4017 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004018 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004019 "cpu": "watchos_i386",
4020 },
4021)
4022
4023config_setting(
4024 name = "watchos_x86_64",
4025 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004026 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004027 "cpu": "watchos_x86_64",
4028 },
4029)
4030
4031config_setting(
4032 name = "tvos_arm64",
4033 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004034 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004035 "cpu": "tvos_arm64",
4036 },
4037)
4038
4039config_setting(
4040 name = "tvos_x86_64",
4041 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004042 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004043 "cpu": "tvos_x86_64",
4044 },
4045)