blob: e451acc139118aada40fcf22fa30427a2bb06cda [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 Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070085 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
86 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070087 "src/f32-clamp/gen/scalar-x1.c",
88 "src/f32-clamp/gen/scalar-x2.c",
89 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080090 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080091 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070092 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070093 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070094 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080095 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080096 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070097 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x4-scalar.c",
99 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
100 "src/f32-dwconv/gen/up1x9-scalar.c",
101 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
102 "src/f32-dwconv/gen/up1x25-scalar.c",
103 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x4-scalar.c",
105 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
106 "src/f32-dwconv/gen/up2x9-scalar.c",
107 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
108 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700109 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
110 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
111 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700113 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
114 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700115 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
116 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
117 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
120 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700121 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700122 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
123 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700124 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
125 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700127 "src/f32-gemm/gen/1x4-scalar.c",
128 "src/f32-gemm/gen/2x4-scalar.c",
129 "src/f32-gemm/gen/4x2-scalar.c",
130 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-gemm/gen/1x4-minmax-scalar.c",
132 "src/f32-gemm/gen/2x4-minmax-scalar.c",
133 "src/f32-gemm/gen/4x2-minmax-scalar.c",
134 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800135 "src/f32-hswish/gen/scalar-x1.c",
136 "src/f32-hswish/gen/scalar-x2.c",
137 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700138 "src/f32-ibilinear/gen/scalar-c1.c",
139 "src/f32-ibilinear/gen/scalar-c2.c",
140 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-igemm/gen/1x4-scalar.c",
142 "src/f32-igemm/gen/2x4-scalar.c",
143 "src/f32-igemm/gen/4x2-scalar.c",
144 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-igemm/gen/1x4-minmax-scalar.c",
146 "src/f32-igemm/gen/2x4-minmax-scalar.c",
147 "src/f32-igemm/gen/4x2-minmax-scalar.c",
148 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700149 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
150 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700152 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
153 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
154 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800156 "src/f32-prelu/gen/scalar-2x1.c",
157 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800158 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700170 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800171 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
174 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
177 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700180 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
181 "src/f32-spmm/gen/1x1-minmax-scalar.c",
182 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar.c",
184 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar.c",
186 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar.c",
188 "src/f32-spmm/gen/8x2-minmax-scalar.c",
189 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700190 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
191 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
193 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
196 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
199 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800202 "src/f32-vbinary/gen/vmax-scalar-x1.c",
203 "src/f32-vbinary/gen/vmax-scalar-x2.c",
204 "src/f32-vbinary/gen/vmax-scalar-x4.c",
205 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
208 "src/f32-vbinary/gen/vmin-scalar-x1.c",
209 "src/f32-vbinary/gen/vmin-scalar-x2.c",
210 "src/f32-vbinary/gen/vmin-scalar-x4.c",
211 "src/f32-vbinary/gen/vminc-scalar-x1.c",
212 "src/f32-vbinary/gen/vminc-scalar-x2.c",
213 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700214 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
215 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
217 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
220 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700232 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
233 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800235 "src/math/expminus-scalar-lut2048-p1.c",
236 "src/math/expminus-scalar-lut64-p2.c",
237 "src/math/expminus-scalar-p5.c",
238 "src/math/sigmoid-scalar-lut2048-p1-div.c",
239 "src/math/sigmoid-scalar-lut64-p2-div.c",
240 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700241 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
242 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700243 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700244 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
245 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700246 "src/q8-gemm/2x2-minmax-scalar.c",
247 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700248 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700249 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700250 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700251 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700252 "src/u8-rmax/scalar.c",
253 "src/x32-packx/x2-scalar.c",
254 "src/x32-packx/x3-scalar.c",
255 "src/x32-packx/x4-scalar.c",
256 "src/x32-pad/x2-scalar.c",
257 "src/x32-unpool/scalar.c",
258 "src/x32-zip/x2-scalar.c",
259 "src/x32-zip/x3-scalar.c",
260 "src/x32-zip/x4-scalar.c",
261 "src/x32-zip/xm-scalar.c",
262 "src/x8-lut/scalar.c",
263 "src/x8-zip/x2-scalar.c",
264 "src/x8-zip/x3-scalar.c",
265 "src/x8-zip/x4-scalar.c",
266 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700267 "src/requantization/precise-scalar.c",
268 "src/requantization/fp32-scalar.c",
269 "src/requantization/q31-scalar.c",
270 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700271]
272
Marat Dukhan436ebe62019-12-04 15:10:12 -0800273WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700274 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
275 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700276 "src/f32-clamp/gen/wasm-x1.c",
277 "src/f32-clamp/gen/wasm-x2.c",
278 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700279 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
280 "src/f32-dwconv/gen/up1x4-wasm.c",
281 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
282 "src/f32-dwconv/gen/up1x9-wasm.c",
283 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
284 "src/f32-dwconv/gen/up1x25-wasm.c",
285 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
286 "src/f32-dwconv/gen/up2x4-wasm.c",
287 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
288 "src/f32-dwconv/gen/up2x9-wasm.c",
289 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
290 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700291 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
292 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
293 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
294 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700295 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
296 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700297 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
298 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
299 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
300 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700301 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
302 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700303 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
304 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700305 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
306 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
307 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700308 "src/f32-gemm/gen/1x4-wasm.c",
309 "src/f32-gemm/gen/2x4-wasm.c",
310 "src/f32-gemm/gen/4x2-wasm.c",
311 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700312 "src/f32-gemm/gen/1x4-minmax-wasm.c",
313 "src/f32-gemm/gen/2x4-minmax-wasm.c",
314 "src/f32-gemm/gen/4x2-minmax-wasm.c",
315 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800316 "src/f32-hswish/gen/wasm-x1.c",
317 "src/f32-hswish/gen/wasm-x2.c",
318 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700319 "src/f32-igemm/gen/1x4-wasm.c",
320 "src/f32-igemm/gen/2x4-wasm.c",
321 "src/f32-igemm/gen/4x2-wasm.c",
322 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700323 "src/f32-igemm/gen/1x4-minmax-wasm.c",
324 "src/f32-igemm/gen/2x4-minmax-wasm.c",
325 "src/f32-igemm/gen/4x2-minmax-wasm.c",
326 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700327 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
328 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
329 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700330 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
331 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
332 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
333 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
334 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
335 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
336 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
337 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
338 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
339 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
340 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
341 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800342 "src/f32-vbinary/gen/vmax-wasm-x1.c",
343 "src/f32-vbinary/gen/vmax-wasm-x2.c",
344 "src/f32-vbinary/gen/vmax-wasm-x4.c",
345 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
346 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
347 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
348 "src/f32-vbinary/gen/vmin-wasm-x1.c",
349 "src/f32-vbinary/gen/vmin-wasm-x2.c",
350 "src/f32-vbinary/gen/vmin-wasm-x4.c",
351 "src/f32-vbinary/gen/vminc-wasm-x1.c",
352 "src/f32-vbinary/gen/vminc-wasm-x2.c",
353 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700354 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
355 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
356 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
357 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
358 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
359 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
360 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
361 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
362 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
363 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
364 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
365 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
366 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
367 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
368 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
369 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
370 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
371 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700372 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
373 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
374 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800375]
376
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800377PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800378 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800379 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800380 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700381 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
382 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700383 "src/f32-clamp/gen/psimd-x4.c",
384 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700385 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
386 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
387 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
388 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
389 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
390 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
391 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
392 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
393 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
394 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
395 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
396 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
398 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700399 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
400 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
401 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
402 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
403 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
404 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
405 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
406 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
407 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
408 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
409 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
410 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
411 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
412 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
413 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
414 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
415 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
416 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
417 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800418 "src/f32-hswish/gen/psimd-x4.c",
419 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700420 "src/f32-ibilinear/gen/psimd-c4.c",
421 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
423 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
424 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
425 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
426 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
427 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
428 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
429 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
430 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
431 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700432 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
433 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
434 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700435 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800436 "src/f32-prelu/gen/psimd-2x4.c",
437 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800438 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700439 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
440 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
441 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
442 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
443 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
444 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
445 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
446 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800447 "src/f32-vbinary/gen/vmax-psimd-x4.c",
448 "src/f32-vbinary/gen/vmax-psimd-x8.c",
449 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
450 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
451 "src/f32-vbinary/gen/vmin-psimd-x4.c",
452 "src/f32-vbinary/gen/vmin-psimd-x8.c",
453 "src/f32-vbinary/gen/vminc-psimd-x4.c",
454 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700455 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
456 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
457 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
458 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
459 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
460 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
461 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
463 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
464 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
465 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
466 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700467 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
468 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700469 "src/x32-packx/x4-psimd.c",
470 "src/x32-pad/x2-psimd.c",
471 "src/x32-unpool/psimd.c",
472 "src/x32-zip/x2-psimd.c",
473 "src/x32-zip/x3-psimd.c",
474 "src/x32-zip/x4-psimd.c",
475 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700476 "src/requantization/precise-psimd.c",
477 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700478]
479
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800480PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800481 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
482 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
483 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
484 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
485 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
486 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
487 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
488 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
489 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
490 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
491 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
492 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800493 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
494 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
495 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
496 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
497 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
498 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
499 "src/math/sigmoid-psimd-p5-div.c",
500]
501
Marat Dukhan08c4a432019-10-03 09:29:21 -0700502# ISA-specific micro-kernels
503NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700504 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
505 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700506 "src/f32-clamp/gen/neon-x4.c",
507 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700508 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
509 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
510 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
511 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700512 "src/f32-gavgpool-spchw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700513 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
514 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700515 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
516 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
517 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
518 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
519 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
520 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
521 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
522 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
523 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
524 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
525 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
526 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
527 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
528 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
529 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
530 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
531 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
532 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
533 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
534 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
535 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
536 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
537 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
538 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
539 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
540 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
541 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
542 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
543 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
544 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
545 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800546 "src/f32-hswish/gen/neon-x4.c",
547 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700548 "src/f32-ibilinear/gen/neon-c4.c",
549 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700550 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
551 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
552 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
553 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
554 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
555 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
556 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
557 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
558 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
559 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
560 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
561 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
562 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
563 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
564 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
565 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700566 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
567 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
568 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700569 "src/f32-ppmm/gen/4x8-minmax-neon.c",
570 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800571 "src/f32-prelu/gen/neon-2x4.c",
572 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
580 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
581 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
582 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
583 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
584 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
585 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
586 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
587 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
588 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
589 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
590 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
591 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
592 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
593 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
594 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
595 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
596 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700597 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800598 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
601 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
602 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
603 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
604 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
605 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
606 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
607 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
608 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
609 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
610 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
611 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
612 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
613 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
614 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
615 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
616 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700617 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
618 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
619 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
620 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800621 "src/f32-vbinary/gen/vmax-neon-x4.c",
622 "src/f32-vbinary/gen/vmax-neon-x8.c",
623 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
624 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
625 "src/f32-vbinary/gen/vmin-neon-x4.c",
626 "src/f32-vbinary/gen/vmin-neon-x8.c",
627 "src/f32-vbinary/gen/vminc-neon-x4.c",
628 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700629 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
630 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
631 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
632 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
633 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
634 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
635 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
636 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
637 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
638 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700639 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
640 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
641 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
642 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700643 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700644 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
645 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700646 "src/q8-gemm/4x8-minmax-neon.c",
647 "src/q8-gemm/8x8-minmax-neon.c",
648 "src/q8-igemm/4x8-minmax-neon.c",
649 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700650 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700651 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700652 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700653 "src/u8-rmax/neon.c",
654 "src/x32-packx/x4-neon-st4.c",
655 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700656 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700657 "src/x32-zip/x2-neon.c",
658 "src/x32-zip/x3-neon.c",
659 "src/x32-zip/x4-neon.c",
660 "src/x32-zip/xm-neon.c",
661 "src/x8-zip/x2-neon.c",
662 "src/x8-zip/x3-neon.c",
663 "src/x8-zip/x4-neon.c",
664 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800665 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800666 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
667 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
668 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
669 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
670 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
671 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700672 "src/requantization/precise-neon.c",
673 "src/requantization/fp32-neon.c",
674 "src/requantization/q31-neon.c",
675 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700676]
677
678NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700679 "src/f32-ibilinear/gen/neonfma-c4.c",
680 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700681 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
682 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
683 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
684 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
685 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
686 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
687 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
688 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
689 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
690 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
691 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
692 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
693 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
694 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
695 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
696 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
697 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
698 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
699 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
700 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
701 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
702 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
703 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
704 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
705 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
706 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
707 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
708 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
709 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
710 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
711 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800712 "src/f32-hswish/gen/neonfma-x4.c",
713 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700714 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
715 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800716 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
717 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
718 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
719 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
720 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
722 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
723 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
725 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
726 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
727 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
728 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
729 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
730 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
731 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
732 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
733 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
734 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
735 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
736 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
737 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
738 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
739 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800740 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
741 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
742 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
743 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
744 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
745 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
746 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
747 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
748 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
749 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
750 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
753 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
754 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
755 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
756 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
768 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
769 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
770 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
771 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
772 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
773 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
774 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
775 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
776 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
777 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
778 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
779 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700794 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
795 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800796 "src/math/exp-neonfma-lut64-p2.c",
797 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800798 "src/math/expminus-neonfma-lut2048-p1.c",
799 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800800 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800801 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
802 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
803 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
804 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
805 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
806 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
807 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
808 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
809 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
810 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
811 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
812 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
813 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
814 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
815 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
816 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
817 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
818 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700819]
820
821AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700822 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
823 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
824 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
825 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
826 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
827 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700828 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
829 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
830 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
831 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
832 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
833 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
834 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
835 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
836 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
837 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
838 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
839 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
840 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
841 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
842 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
843 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
844 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
845 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
846 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
847 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
849 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
850 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
851 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
852 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
853 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
854 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800855 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
856 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
857 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
858 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
859 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
860 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
861 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
862 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
863 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
864 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
865 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
866 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
867 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
868 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
869 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
870 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
871 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
872 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700873 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
874 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
875 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
876 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
877 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
878 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
879 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
880 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
881 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
882 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
883 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
884 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
885 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
886 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
887 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
888 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
889 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
890 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800891 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
892 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
893 "src/math/sigmoid-neonfma-rr1-p5-div.c",
894 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
895 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
896 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700897]
898
899AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchard83ea9392020-04-20 20:15:14 -0700900 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
901 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
902 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
903 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
904 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
905 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
906 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
907 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700908 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
909 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
910 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
911 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
912 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
913 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
914 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
915 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700916]
917
918SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700919 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
920 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700921 "src/f32-clamp/gen/sse-x4.c",
922 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800923 "src/f32-dwconv-spchw/3x3p1-sse.c",
924 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700925 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
926 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
927 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
928 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
929 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
930 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
931 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
932 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
933 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
934 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
935 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
936 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937 "src/f32-gavgpool-spchw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700938 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
939 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700940 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
941 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
942 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
943 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
944 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
945 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
946 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
947 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
948 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
949 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
950 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
951 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
952 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800953 "src/f32-hswish/gen/sse-x4.c",
954 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700955 "src/f32-ibilinear/gen/sse-c4.c",
956 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700957 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
958 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
959 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
960 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
961 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
962 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
963 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700964 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
965 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
966 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700967 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700968 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700969 "src/f32-spmm/gen/4x1-minmax-sse.c",
970 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700971 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
972 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
973 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
974 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
975 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
976 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
977 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
978 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800979 "src/f32-vbinary/gen/vmax-sse-x4.c",
980 "src/f32-vbinary/gen/vmax-sse-x8.c",
981 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
982 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
983 "src/f32-vbinary/gen/vmin-sse-x4.c",
984 "src/f32-vbinary/gen/vmin-sse-x8.c",
985 "src/f32-vbinary/gen/vminc-sse-x4.c",
986 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700987 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
988 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
989 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
990 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
991 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
992 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
993 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
994 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
995 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
996 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
997 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
998 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700999 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1000 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001001 "src/x32-packx/x4-sse.c",
1002]
1003
1004SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001005 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1006 "src/f32-argmaxpool/4x-sse2-c4.c",
1007 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001008 "src/f32-prelu/gen/sse2-2x4.c",
1009 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001010 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1011 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1012 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1013 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1014 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1015 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1016 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1017 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1018 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1019 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1020 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1021 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001022 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001023 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001024 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001025 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001026 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1027 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001028 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1029 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001030 "src/q8-igemm/4x4c2-minmax-sse2.c",
1031 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001032 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1033 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001034 "src/q8-gemm/2x4c8-minmax-sse2.c",
1035 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001036 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001037 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001038 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001039 "src/u8-rmax/sse2.c",
1040 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001041 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001042 "src/x32-zip/x2-sse2.c",
1043 "src/x32-zip/x3-sse2.c",
1044 "src/x32-zip/x4-sse2.c",
1045 "src/x32-zip/xm-sse2.c",
1046 "src/x8-zip/x2-sse2.c",
1047 "src/x8-zip/x3-sse2.c",
1048 "src/x8-zip/x4-sse2.c",
1049 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001050 "src/math/exp-sse2-p5.c",
1051 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001052 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001053 "src/requantization/precise-sse2.c",
1054 "src/requantization/fp32-sse2.c",
1055 "src/requantization/q31-sse2.c",
1056 "src/requantization/gemmlowp-sse2.c",
1057]
1058
1059SSSE3_UKERNELS = [
1060 "src/requantization/precise-ssse3.c",
1061 "src/requantization/q31-ssse3.c",
1062 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001063]
1064
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001065SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001066 "src/f32-prelu/gen/sse41-2x4.c",
1067 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001068 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001069 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001070 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001071 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001072 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1073 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001074 "src/requantization/precise-sse4.c",
1075 "src/requantization/q31-sse4.c",
1076 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001077]
1078
Marat Dukhan08c4a432019-10-03 09:29:21 -07001079AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001080 "src/f32-clamp/gen/avx-x8.c",
1081 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001082 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1083 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1084 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1085 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1086 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1087 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1088 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1089 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1090 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1091 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1092 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1093 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1094 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1095 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1096 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1097 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1098 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1099 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1100 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1101 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1102 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1103 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1104 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1105 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1106 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1107 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1108 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1109 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1110 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1111 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001112 "src/f32-hswish/gen/avx-x8.c",
1113 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001114 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1115 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1116 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1117 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1118 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1119 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1120 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1121 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1122 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001123 "src/f32-prelu/gen/avx-2x8.c",
1124 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001125 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001126 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1127 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1128 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1129 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1130 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1131 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1132 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1133 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001134 "src/f32-vbinary/gen/vmax-avx-x8.c",
1135 "src/f32-vbinary/gen/vmax-avx-x16.c",
1136 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1137 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1138 "src/f32-vbinary/gen/vmin-avx-x8.c",
1139 "src/f32-vbinary/gen/vmin-avx-x16.c",
1140 "src/f32-vbinary/gen/vminc-avx-x8.c",
1141 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001142 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1143 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1144 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1145 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1146 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1147 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1148 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1149 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1150 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1151 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1152 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1153 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001154 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001155]
1156
Marat Dukhanfda12b82019-11-21 12:27:59 -08001157FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001158 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1159 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1160 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1161 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1162 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1163 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1164 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1165 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1166 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1167 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1168 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1169 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1170 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1171 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1172 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1173 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1174 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1175 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1176 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1177 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1178 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1179 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1180 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1181 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1182 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1183 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1184 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1185 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1186 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1187 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1188 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1189 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1190 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1191 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1192 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1193 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1194 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1195 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1196 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1197 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001198 "src/f32-hswish/gen/fma3-x8.c",
1199 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001200 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1201 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1202 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1203 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1204 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1205 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1206 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1207 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1208 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1209 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1210 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1211 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1212 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1213 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001214]
1215
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001216AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001217 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1218 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1219 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1220 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1221 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1222 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1223 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1224 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1225 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1226 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1227 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1228 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1229 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1230 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1231 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1232 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1233 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1234 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1235 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1236 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1237 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1238 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1239 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1240 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1241 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1242 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1243 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1244 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1245 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1246 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1247 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1248 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1249 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1250 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1251 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1252 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001253 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1254 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1255 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1256 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1257 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1258 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1259 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1260 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1261 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1262 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1263 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1264 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1265 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1266 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1267 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1268 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1269 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1270 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1271 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1272 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1273 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1274 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1275 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1276 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1277 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1278 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1279 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1280 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1281 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1282 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001283 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1284 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1285 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1286 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1287 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1288 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1289 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1290 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1291 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1292 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1293 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1294 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1295 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1296 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1297 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1298 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1299 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1300 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1301 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1302 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1303 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1304 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1305 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1306 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001307 "src/math/exp-avx2-p5.c",
1308 "src/math/exp-avx2-perm-p3.c",
1309 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001310 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001311 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001312 "src/math/sigmoid-avx2-rr2-p5-div.c",
1313 "src/math/sigmoid-avx2-rr1-p5-div.c",
1314 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1315 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1316 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1317 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001318]
1319
Marat Dukhan08c4a432019-10-03 09:29:21 -07001320AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001321 "src/f32-clamp/gen/avx512f-x16.c",
1322 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001323 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1324 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1325 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1326 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1327 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1328 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1329 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1330 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1331 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1332 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1333 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1334 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1335 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1336 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1337 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1338 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1339 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1340 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1341 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1342 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1343 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1344 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1345 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1346 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001347 "src/f32-hswish/gen/avx512f-x16.c",
1348 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001349 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1350 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1351 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1352 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1353 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1354 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001355 "src/f32-prelu/gen/avx512f-2x16.c",
1356 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001357 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1358 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1359 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1360 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1361 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1362 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1363 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1364 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1365 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1366 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1367 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1368 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1369 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1370 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1371 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1372 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1373 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1374 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1375 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1376 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1377 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1378 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1379 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1380 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1381 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1382 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1383 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1384 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1385 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1386 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1387 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1388 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1389 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1390 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1391 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1392 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001393 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001394 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1395 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1396 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1397 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1398 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1399 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1400 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1401 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001402 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1403 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1404 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1405 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1406 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1407 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1408 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1409 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001410 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1411 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1412 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1413 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1414 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1415 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1416 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1417 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1418 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1419 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1420 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1421 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001422 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001423 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1424 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1425 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1426 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1427 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1428 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1429 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1430 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1431 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1432 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1433 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1434 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1435 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1436 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1437 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1438 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1439 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1440 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1441 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1442 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1443 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1444 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1445 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1446 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001447 "src/math/exp-avx512f-p5-scalef.c",
1448 "src/math/exp-avx512f-p5.c",
1449 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001450 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001451 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001452]
1453
1454AARCH32_ASM_UKERNELS = [
Marat Dukhande06f492020-04-09 00:19:31 -07001455 "src/q8-dwconv/up8x9-minmax-aarch32-neon.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001456 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1457 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1458 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1459 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1460 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1461 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1462 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1463 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1464 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1465 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001466]
1467
1468AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001469 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001470 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1471 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001472 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001473 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1474 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001475 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1476 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1477 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001478 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001479 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1480 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1481 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001482 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001483 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1484 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001485 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001486 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1487 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1488 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1489 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1490 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1491 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1492 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1493 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1494 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1495 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1496 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1497 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1498 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1499 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1500 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1501 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1502 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1503 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1504 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1505 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1506 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001507 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001508 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1509 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1510 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1511 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1512 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1513 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1514 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1515 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1516 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1517 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1518 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1519 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1520 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1521 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1522 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1523 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1524 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1525 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1526 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1527 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1528 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1529 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1530 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1531 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1532 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1533 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1534 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1535 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1536 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1537 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1538 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1539 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1540 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1541 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1542 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1543 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1544 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1545 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001546]
1547
Marat Dukhan1b354632020-03-23 12:50:22 -07001548INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001549 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001550 "src/xnnpack/argmaxpool.h",
1551 "src/xnnpack/avgpool.h",
1552 "src/xnnpack/clamp.h",
1553 "src/xnnpack/common.h",
1554 "src/xnnpack/conv.h",
1555 "src/xnnpack/dwconv.h",
1556 "src/xnnpack/gavgpool.h",
1557 "src/xnnpack/gemm.h",
1558 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001559 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001560 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001561 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001562 "src/xnnpack/lut.h",
1563 "src/xnnpack/math.h",
1564 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001565 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001566 "src/xnnpack/packx.h",
1567 "src/xnnpack/pad.h",
1568 "src/xnnpack/params.h",
1569 "src/xnnpack/pavgpool.h",
1570 "src/xnnpack/ppmm.h",
1571 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001572 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001573 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001574 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001575 "src/xnnpack/rmax.h",
1576 "src/xnnpack/scalar-utils.h",
1577 "src/xnnpack/spmm.h",
1578 "src/xnnpack/unpool.h",
1579 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001580 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001582 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001583 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001584 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001585 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001586 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001587]
1588
1589INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001590 "include/xnnpack.h",
1591 "src/xnnpack/allocator.h",
1592 "src/xnnpack/compute.h",
1593 "src/xnnpack/im2col.h",
1594 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001595 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001596 "src/xnnpack/operator.h",
1597 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001598 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001599 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001600 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001601 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001602]
1603
Marat Dukhan1b354632020-03-23 12:50:22 -07001604ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001605 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606]
1607
Marat Dukhan1b354632020-03-23 12:50:22 -07001608MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001609 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610 "include/xnnpack.h",
1611]
1612
Marat Dukhan1b354632020-03-23 12:50:22 -07001613MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001614 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001615 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001616 "src/xnnpack/requantization.h",
1617 "include/xnnpack.h",
1618]
1619
1620OPERATOR_TEST_PARAMS_HDRS = [
1621 "src/xnnpack/params.h",
1622 "src/xnnpack/common.h",
1623]
1624
1625WEIGHTS_PACK_HDRS = [
1626 "src/xnnpack/pack.h",
1627 "src/xnnpack/operator.h",
1628 "src/xnnpack/compute.h",
1629]
1630
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001631LOGGING_COPTS = select({
1632 # No logging in optimized mode
1633 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1634 # Full logging in debug mode
1635 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1636 # Error-only logging in default (fastbuild) mode
1637 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1638})
1639
1640LOGGING_HDRS = [
1641 "src/xnnpack/log.h",
1642]
1643
Marat Dukhan08c4a432019-10-03 09:29:21 -07001644xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001645 name = "tables",
1646 srcs = TABLE_SRCS,
1647 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001648 gcc_copts = xnnpack_gcc_std_copts(),
1649 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001650)
1651
1652xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653 name = "scalar_ukernels",
1654 srcs = SCALAR_UKERNELS,
1655 hdrs = INTERNAL_HDRS,
1656 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001657 gcc_copts = xnnpack_gcc_std_copts(),
1658 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001659 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001660 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001661 "@FP16",
1662 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001663 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664 ],
1665)
1666
1667xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001668 name = "wasm_ukernels",
1669 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001670 gcc_copts = xnnpack_gcc_std_copts(),
1671 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001672 wasm_srcs = WASM_UKERNELS,
1673 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001674 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001675 "@FP16",
1676 "@FXdiv",
1677 "@pthreadpool",
1678 ],
1679)
1680
1681xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001682 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 hdrs = INTERNAL_HDRS,
1684 aarch32_copts = [
1685 "-marm",
1686 "-mfpu=neon",
1687 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001688 gcc_copts = xnnpack_gcc_std_copts(),
1689 gcc_x86_copts = ["-msse2"],
1690 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001691 optimized_copts = [
1692 "-O3",
1693 "-ffast-math",
1694 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001695 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001696 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001697 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001698 "@FP16",
1699 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001700 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001701 ],
1702)
1703
1704xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001705 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001706 hdrs = INTERNAL_HDRS,
1707 aarch32_copts = [
1708 "-marm",
1709 "-mfpu=neon",
1710 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001711 gcc_copts = xnnpack_gcc_std_copts(),
1712 gcc_x86_copts = ["-msse2"],
1713 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001714 optimized_copts = [
1715 "-O3",
1716 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001717 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001718 deps = [
1719 ":tables",
1720 "@FP16",
1721 "@psimd",
1722 "@pthreadpool",
1723 ],
1724)
1725
1726xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001727 name = "neon_ukernels",
1728 hdrs = INTERNAL_HDRS,
1729 aarch32_copts = [
1730 "-marm",
1731 "-mfpu=neon",
1732 ],
1733 aarch32_srcs = NEON_UKERNELS,
1734 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001735 gcc_copts = xnnpack_gcc_std_copts(),
1736 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001737 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001738 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001739 "@FP16",
1740 "@pthreadpool",
1741 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001742)
1743
1744xnnpack_cc_library(
1745 name = "neonfma_ukernels",
1746 hdrs = INTERNAL_HDRS,
1747 aarch32_copts = [
1748 "-marm",
1749 "-mfpu=neon-vfpv4",
1750 ],
1751 aarch32_srcs = NEONFMA_UKERNELS,
1752 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001753 gcc_copts = xnnpack_gcc_std_copts(),
1754 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001755 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001756 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001757 "@FP16",
1758 "@pthreadpool",
1759 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001760)
1761
1762xnnpack_cc_library(
1763 name = "neonfp16arith_ukernels",
1764 hdrs = INTERNAL_HDRS,
1765 aarch64_copts = ["-march=armv8.2-a+fp16"],
1766 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001767 gcc_copts = xnnpack_gcc_std_copts(),
1768 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001769 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001770 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001771 "@FP16",
1772 "@pthreadpool",
1773 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001774)
1775
1776xnnpack_cc_library(
1777 name = "sse2_ukernels",
1778 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001779 gcc_copts = xnnpack_gcc_std_copts(),
1780 gcc_x86_copts = ["-msse2"],
1781 msvc_copts = xnnpack_msvc_std_copts(),
1782 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001783 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001784 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001785 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001786 "@FP16",
1787 "@pthreadpool",
1788 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001789)
1790
1791xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001792 name = "ssse3_ukernels",
1793 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001794 gcc_copts = xnnpack_gcc_std_copts(),
1795 gcc_x86_copts = ["-mssse3"],
1796 msvc_copts = xnnpack_msvc_std_copts(),
1797 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001798 x86_srcs = SSSE3_UKERNELS,
1799 deps = [
1800 ":tables",
1801 "@FP16",
1802 "@pthreadpool",
1803 ],
1804)
1805
1806xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001807 name = "sse41_ukernels",
1808 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001809 gcc_copts = xnnpack_gcc_std_copts(),
1810 gcc_x86_copts = ["-msse4.1"],
1811 msvc_copts = xnnpack_msvc_std_copts(),
1812 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001813 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001814 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001815 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001816 "@FP16",
1817 "@pthreadpool",
1818 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001819)
1820
1821xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822 name = "avx_ukernels",
1823 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001824 gcc_copts = xnnpack_gcc_std_copts(),
1825 gcc_x86_copts = ["-mavx"],
1826 msvc_copts = xnnpack_msvc_std_copts(),
1827 msvc_x86_32_copts = ["/arch:AVX"],
1828 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001830 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001831 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001832 "@FP16",
1833 "@pthreadpool",
1834 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835)
1836
1837xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001838 name = "fma3_ukernels",
1839 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001840 gcc_copts = xnnpack_gcc_std_copts(),
1841 gcc_x86_copts = ["-mfma"],
1842 msvc_copts = xnnpack_msvc_std_copts(),
1843 msvc_x86_32_copts = ["/arch:AVX"],
1844 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001845 x86_srcs = FMA3_UKERNELS,
1846 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001847 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001848 "@FP16",
1849 "@pthreadpool",
1850 ],
1851)
1852
1853xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001854 name = "avx2_ukernels",
1855 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001856 gcc_copts = xnnpack_gcc_std_copts(),
1857 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001858 "-mfma",
1859 "-mavx2",
1860 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001861 msvc_copts = xnnpack_msvc_std_copts(),
1862 msvc_x86_32_copts = ["/arch:AVX2"],
1863 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001864 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001865 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001866 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001867 "@FP16",
1868 "@pthreadpool",
1869 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001870)
1871
1872xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873 name = "avx512f_ukernels",
1874 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001875 gcc_copts = xnnpack_gcc_std_copts(),
1876 gcc_x86_copts = ["-mavx512f"],
1877 mingw_copts = ["-fno-asynchronous-unwind-tables"],
1878 msvc_copts = xnnpack_msvc_std_copts(),
1879 msvc_x86_32_copts = ["/arch:AVX512"],
1880 msvc_x86_64_copts = ["/arch:AVX512"],
1881 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001882 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001883 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001884 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001885 "@FP16",
1886 "@pthreadpool",
1887 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001888)
1889
1890xnnpack_cc_library(
1891 name = "asm_ukernels",
1892 hdrs = ["src/xnnpack/assembly.h"],
1893 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07001894 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001895 aarch64_srcs = AARCH64_ASM_UKERNELS,
1896)
1897
1898xnnpack_aggregate_library(
1899 name = "ukernels",
1900 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001901 ":neon_ukernels",
1902 ":neonfma_ukernels",
1903 ":asm_ukernels",
1904 ],
1905 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906 ":neon_ukernels",
1907 ":neonfma_ukernels",
1908 ":neonfp16arith_ukernels",
1909 ":asm_ukernels",
1910 ],
1911 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07001912 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001913 ":psimd_fastmath_ukernels",
1914 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001916 wasm_deps = [":wasm_ukernels"],
1917 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001919 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001920 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001921 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001922 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001923 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001924 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925 ":avx512f_ukernels",
1926 ],
1927)
1928
1929xnnpack_cc_library(
1930 name = "im2col",
1931 srcs = ["src/im2col.c"],
1932 hdrs = [
1933 "src/xnnpack/common.h",
1934 "src/xnnpack/im2col.h",
1935 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001936 gcc_copts = xnnpack_gcc_std_copts(),
1937 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001938)
1939
1940xnnpack_cc_library(
1941 name = "indirection",
1942 srcs = ["src/indirection.c"],
1943 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001944 gcc_copts = xnnpack_gcc_std_copts(),
1945 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001946 deps = [
1947 "@FP16",
1948 "@FXdiv",
1949 "@pthreadpool",
1950 ],
1951)
1952
1953xnnpack_cc_library(
1954 name = "operator_run",
1955 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001956 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001957 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07001958 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1959 "//conditions:default": [],
1960 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001961 gcc_copts = xnnpack_gcc_std_copts(),
1962 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001963 deps = [
1964 "@FP16",
1965 "@FXdiv",
1966 "@clog",
1967 "@pthreadpool",
1968 ],
1969)
1970
1971cc_library(
1972 name = "enable_assembly",
1973 defines = select({
1974 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1975 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001976 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001977 }),
1978)
1979
Marat Dukhancf056b22019-10-07 10:26:29 -07001980xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001981 name = "operators",
1982 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001983 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001984 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001985 ],
1986 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001987 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001988 "-Isrc",
1989 "-Iinclude",
1990 ] + select({
1991 ":debug_build": [],
1992 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07001993 }) + select({
1994 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
1995 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001996 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07001997 gcc_copts = xnnpack_gcc_std_copts(),
1998 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhancf056b22019-10-07 10:26:29 -07001999 wasm_srcs = ["src/wasm-stubs.c"],
2000 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002002 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003 "@FP16",
2004 "@FXdiv",
2005 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002006 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002007 ],
2008)
2009
Marat Dukhan10a38082020-04-17 03:58:35 -07002010xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002011 name = "XNNPACK",
2012 srcs = [
2013 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002014 "src/runtime.c",
2015 "src/subgraph.c",
2016 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002017 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002018 hdrs = ["include/xnnpack.h"],
2019 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002020 "-Isrc",
2021 "-Iinclude",
2022 ] + select({
2023 ":debug_build": [],
2024 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002025 }) + select({
2026 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2027 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002028 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002029 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002030 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002031 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002032 visibility = xnnpack_visibility(),
2033 deps = [
2034 ":enable_assembly",
2035 ":ukernels",
2036 ":operator_run",
2037 ":operators",
2038 "@clog",
2039 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002040 ] + select({
2041 ":emscripten": [],
2042 "//conditions:default": ["@cpuinfo"],
2043 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002044)
2045
Marat Dukhan10a38082020-04-17 03:58:35 -07002046xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002047 name = "xnnpack_operators_nhwc_f32",
2048 srcs = [
2049 "src/init.c",
2050 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002051 hdrs = ["include/xnnpack.h"],
2052 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002053 "-Isrc",
2054 "-Iinclude",
2055 ] + select({
2056 ":debug_build": [],
2057 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002058 }) + select({
2059 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2060 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002061 }),
2062 defines = [
2063 "XNN_NO_Q8_OPERATORS",
2064 "XNN_NO_U8_OPERATORS",
2065 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002066 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002067 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002068 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002070 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002071 visibility = xnnpack_visibility(),
2072 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002073 ":enable_assembly",
2074 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002075 ":operator_run",
2076 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002077 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002078 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002079 ] + select({
2080 ":emscripten": [],
2081 "//conditions:default": ["@cpuinfo"],
2082 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002083)
2084
Marat Dukhancf056b22019-10-07 10:26:29 -07002085xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002086 name = "bench_utils",
2087 srcs = ["bench/utils.cc"],
2088 hdrs = ["bench/utils.h"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002089 gcc_copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002090 deps = [
2091 "@com_google_benchmark//:benchmark",
2092 "@cpuinfo",
2093 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094)
2095
Frank Barchard7e955972019-10-11 10:34:25 -07002096######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097
2098xnnpack_benchmark(
2099 name = "q8_gemm_bench",
2100 srcs = [
2101 "bench/gemm.h",
2102 "bench/q8-gemm.cc",
2103 "src/xnnpack/AlignedAllocator.h",
2104 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002105 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
2106 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002107 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2108)
2109
2110xnnpack_benchmark(
2111 name = "f16_gemm_bench",
2112 srcs = [
2113 "bench/f16-gemm.cc",
2114 "bench/gemm.h",
2115 "src/xnnpack/AlignedAllocator.h",
2116 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002117 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118 deps = MICROKERNEL_BENCHMARK_DEPS,
2119)
2120
2121xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002122 name = "f16_spmm_bench",
2123 srcs = [
2124 "bench/f16-spmm.cc",
2125 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002126 "src/xnnpack/AlignedAllocator.h",
2127 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002128 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002129 deps = MICROKERNEL_BENCHMARK_DEPS,
2130)
2131
2132xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002133 name = "f32_igemm_bench",
2134 srcs = [
2135 "bench/f32-igemm.cc",
2136 "bench/conv.h",
2137 "src/xnnpack/AlignedAllocator.h",
2138 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002139 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002140)
2141
2142xnnpack_benchmark(
2143 name = "f32_conv_hwc_bench",
2144 srcs = [
2145 "bench/f32-conv-hwc.cc",
2146 "bench/dconv.h",
2147 "src/xnnpack/AlignedAllocator.h",
2148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002149 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002150 deps = MICROKERNEL_BENCHMARK_DEPS,
2151)
2152
2153xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07002154 name = "f32_conv_hwc2spchw_bench",
2155 srcs = [
2156 "bench/f32-conv-hwc2spchw.cc",
2157 "bench/dconv.h",
2158 "src/xnnpack/AlignedAllocator.h",
2159 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002160 gcc_copts = ["-Wno-unused-function"],
Erich Elsen563df5f2019-10-23 08:02:21 -07002161 deps = MICROKERNEL_BENCHMARK_DEPS,
2162)
2163
2164xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002165 name = "f32_dwconv_bench",
2166 srcs = [
2167 "bench/f32-dwconv.cc",
2168 "bench/dwconv.h",
2169 "src/xnnpack/AlignedAllocator.h",
2170 ] + MICROKERNEL_BENCHMARK_HDRS,
2171 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2172)
2173
2174xnnpack_benchmark(
2175 name = "f32_dwconv_spchw_bench",
2176 srcs = [
2177 "bench/f32-dwconv-spchw.cc",
2178 "bench/dwconv.h",
2179 "src/xnnpack/AlignedAllocator.h",
2180 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2181 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2182)
2183
2184xnnpack_benchmark(
2185 name = "f32_gemm_bench",
2186 srcs = [
2187 "bench/f32-gemm.cc",
2188 "bench/gemm.h",
2189 "src/xnnpack/AlignedAllocator.h",
2190 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002191 copts = xnnpack_optional_ruy_copts(),
2192 gcc_copts = ["-Wno-unused-function"],
Frank Barchard7e955972019-10-11 10:34:25 -07002193 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002194)
2195
2196xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002197 name = "f32_raddexpminusmax_bench",
2198 srcs = [
2199 "bench/f32-raddexpminusmax.cc",
2200 "src/xnnpack/AlignedAllocator.h",
2201 ] + MICROKERNEL_BENCHMARK_HDRS,
2202 deps = MICROKERNEL_BENCHMARK_DEPS,
2203)
2204
2205xnnpack_benchmark(
2206 name = "f32_raddextexp_bench",
2207 srcs = [
2208 "bench/f32-raddextexp.cc",
2209 "src/xnnpack/AlignedAllocator.h",
2210 ] + MICROKERNEL_BENCHMARK_HDRS,
2211 deps = MICROKERNEL_BENCHMARK_DEPS,
2212)
2213
2214xnnpack_benchmark(
2215 name = "f32_raddstoreexpminusmax_bench",
2216 srcs = [
2217 "bench/f32-raddstoreexpminusmax.cc",
2218 "src/xnnpack/AlignedAllocator.h",
2219 ] + MICROKERNEL_BENCHMARK_HDRS,
2220 deps = MICROKERNEL_BENCHMARK_DEPS,
2221)
2222
2223xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002224 name = "f32_rmax_bench",
2225 srcs = [
2226 "bench/f32-rmax.cc",
2227 "src/xnnpack/AlignedAllocator.h",
2228 ] + MICROKERNEL_BENCHMARK_HDRS,
2229 deps = MICROKERNEL_BENCHMARK_DEPS,
2230)
2231
2232xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002233 name = "f32_sigmoid_bench",
2234 srcs = [
2235 "bench/f32-sigmoid.cc",
2236 "src/xnnpack/AlignedAllocator.h",
2237 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002238 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan14bec502019-11-18 11:35:31 -08002239 deps = MICROKERNEL_BENCHMARK_DEPS,
2240)
2241
2242xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243 name = "f32_spmm_bench",
2244 srcs = [
2245 "bench/f32-spmm.cc",
2246 "bench/gemm.h",
2247 "src/xnnpack/AlignedAllocator.h",
2248 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002249 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002250 deps = MICROKERNEL_BENCHMARK_DEPS,
2251)
2252
2253xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002254 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002255 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002256 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002257 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002258 copts = xnnpack_optional_dnnl_copts(),
2259 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002260 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002261)
2262
2263xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002264 name = "f32_vscaleexpminusmax_bench",
2265 srcs = [
2266 "bench/f32-vscaleexpminusmax.cc",
2267 "src/xnnpack/AlignedAllocator.h",
2268 ] + MICROKERNEL_BENCHMARK_HDRS,
2269 deps = MICROKERNEL_BENCHMARK_DEPS,
2270)
2271
2272xnnpack_benchmark(
2273 name = "f32_vscaleextexp_bench",
2274 srcs = [
2275 "bench/f32-vscaleextexp.cc",
2276 "src/xnnpack/AlignedAllocator.h",
2277 ] + MICROKERNEL_BENCHMARK_HDRS,
2278 deps = MICROKERNEL_BENCHMARK_DEPS,
2279)
2280
2281xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002282 name = "f32_im2col_gemm_bench",
2283 srcs = [
2284 "bench/f32-im2col-gemm.cc",
2285 "bench/conv.h",
2286 "src/xnnpack/AlignedAllocator.h",
2287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2288 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2289)
2290
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002291xnnpack_benchmark(
2292 name = "requantization_bench",
2293 srcs = [
2294 "bench/requantization.cc",
2295 "src/xnnpack/requantization-stubs.h",
2296 "src/xnnpack/AlignedAllocator.h",
2297 ] + MICROKERNEL_BENCHMARK_HDRS,
2298 deps = MICROKERNEL_BENCHMARK_DEPS,
2299)
2300
Marat Dukhan08c4a432019-10-03 09:29:21 -07002301########################### Benchmarks for operators ###########################
2302
2303xnnpack_benchmark(
2304 name = "add_bench",
2305 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002306 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002307)
2308
2309xnnpack_benchmark(
2310 name = "average_pooling_bench",
2311 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002312 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002313 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002314 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002315)
2316
2317xnnpack_benchmark(
2318 name = "channel_shuffle_bench",
2319 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002320 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002321)
2322
2323xnnpack_benchmark(
2324 name = "convolution_bench",
2325 srcs = ["bench/convolution.cc"],
2326 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002327 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002328 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002329)
2330
2331xnnpack_benchmark(
2332 name = "deconvolution_bench",
2333 srcs = ["bench/deconvolution.cc"],
2334 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002335 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002336 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337)
2338
2339xnnpack_benchmark(
2340 name = "global_average_pooling_bench",
2341 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002342 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343)
2344
2345xnnpack_benchmark(
2346 name = "max_pooling_bench",
2347 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002348 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002349)
2350
2351xnnpack_benchmark(
2352 name = "sigmoid_bench",
2353 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002354 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002355 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002356 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357)
2358
2359xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002360 name = "prelu_bench",
2361 srcs = ["bench/prelu.cc"],
2362 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002363 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002364 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002365)
2366
2367xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002368 name = "softmax_bench",
2369 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002370 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002371 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002372 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373)
2374
Marat Dukhanc068bb62019-10-04 13:24:39 -07002375############################# End-to-end benchmarks ############################
2376
2377cc_library(
2378 name = "mobilenet_v1",
2379 srcs = ["models/mobilenet-v1.cc"],
2380 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002381 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002382 linkstatic = True,
2383 deps = [
2384 ":XNNPACK",
2385 "@pthreadpool",
2386 ],
2387)
2388
2389cc_library(
2390 name = "mobilenet_v2",
2391 srcs = ["models/mobilenet-v2.cc"],
2392 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002393 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002394 linkstatic = True,
2395 deps = [
2396 ":XNNPACK",
2397 "@pthreadpool",
2398 ],
2399)
2400
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002401cc_library(
2402 name = "mobilenet_v3_large",
2403 srcs = ["models/mobilenet-v3-large.cc"],
2404 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002405 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002406 linkstatic = True,
2407 deps = [
2408 ":XNNPACK",
2409 "@pthreadpool",
2410 ],
2411)
2412
2413cc_library(
2414 name = "mobilenet_v3_small",
2415 srcs = ["models/mobilenet-v3-small.cc"],
2416 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002417 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002418 linkstatic = True,
2419 deps = [
2420 ":XNNPACK",
2421 "@pthreadpool",
2422 ],
2423)
2424
Marat Dukhanc068bb62019-10-04 13:24:39 -07002425xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002426 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002427 srcs = [
2428 "bench/f32-dwconv-e2e.cc",
2429 "bench/end2end.h",
2430 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002431 gcc_copts = ["-Wno-unused-function"],
Marat Dukhanef4416e2019-10-31 13:44:40 -07002432 deps = MICROKERNEL_BENCHMARK_DEPS + [
2433 ":XNNPACK",
2434 ":mobilenet_v1",
2435 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002436 ":mobilenet_v3_large",
2437 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002438 ],
2439)
2440
2441xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002442 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002443 srcs = [
2444 "bench/f32-gemm-e2e.cc",
2445 "bench/end2end.h",
2446 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002447 gcc_copts = ["-Wno-unused-function"],
Marat Dukhan5f18d262019-10-31 10:24:14 -07002448 deps = MICROKERNEL_BENCHMARK_DEPS + [
2449 ":XNNPACK",
2450 ":mobilenet_v1",
2451 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002452 ":mobilenet_v3_large",
2453 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002454 ],
2455)
2456
2457xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002458 name = "end2end_bench",
2459 srcs = ["bench/end2end.cc"],
2460 deps = [
2461 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002462 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002463 ":mobilenet_v1",
2464 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002465 ":mobilenet_v3_large",
2466 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002467 "@pthreadpool",
2468 ],
2469)
2470
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002471#################### Accuracy evaluation for math functions ####################
2472
2473xnnpack_benchmark(
2474 name = "f32_exp_eval",
2475 srcs = [
2476 "eval/f32-exp.cc",
2477 "src/xnnpack/AlignedAllocator.h",
2478 ] + ACCURACY_EVAL_HDRS,
2479 deps = ACCURACY_EVAL_DEPS,
2480)
2481
Marat Dukhan515c9772019-10-17 18:07:57 -07002482xnnpack_benchmark(
2483 name = "f32_expminus_eval",
2484 srcs = [
2485 "eval/f32-expminus.cc",
2486 "src/xnnpack/AlignedAllocator.h",
2487 ] + ACCURACY_EVAL_HDRS,
2488 deps = ACCURACY_EVAL_DEPS,
2489)
2490
Marat Dukhan98ba4412019-10-23 02:14:28 -07002491xnnpack_benchmark(
2492 name = "f32_extexp_eval",
2493 srcs = [
2494 "eval/f32-extexp.cc",
2495 "src/xnnpack/AlignedAllocator.h",
2496 ] + ACCURACY_EVAL_HDRS,
2497 deps = ACCURACY_EVAL_DEPS,
2498)
2499
Marat Dukhan346a9e52019-11-15 09:06:30 -08002500xnnpack_benchmark(
2501 name = "f32_sigmoid_eval",
2502 srcs = [
2503 "eval/f32-sigmoid.cc",
2504 "src/xnnpack/AlignedAllocator.h",
2505 ] + ACCURACY_EVAL_HDRS,
2506 deps = ACCURACY_EVAL_DEPS,
2507)
2508
Marat Dukhan08c4a432019-10-03 09:29:21 -07002509######################### Unit tests for micro-kernels #########################
2510
2511xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002512 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002513 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002514 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002515 "test/gemm-microkernel-tester.h",
2516 "src/xnnpack/AlignedAllocator.h",
2517 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2518 deps = MICROKERNEL_TEST_DEPS,
2519)
2520
2521xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002522 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002523 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002524 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002525 "test/spmm-microkernel-tester.h",
2526 "src/xnnpack/AlignedAllocator.h",
2527 ] + MICROKERNEL_TEST_HDRS,
2528 deps = MICROKERNEL_TEST_DEPS,
2529)
2530
2531xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002532 name = "f32_argmaxpool_test",
2533 srcs = [
2534 "test/f32-argmaxpool.cc",
2535 "test/argmaxpool-microkernel-tester.h",
2536 "src/xnnpack/AlignedAllocator.h",
2537 ] + MICROKERNEL_TEST_HDRS,
2538 deps = MICROKERNEL_TEST_DEPS,
2539)
2540
2541xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002542 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002543 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002544 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002545 "test/avgpool-microkernel-tester.h",
2546 "src/xnnpack/AlignedAllocator.h",
2547 ] + MICROKERNEL_TEST_HDRS,
2548 deps = MICROKERNEL_TEST_DEPS,
2549)
2550
2551xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002552 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002553 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002554 "test/f32-ibilinear.cc",
2555 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002556 "src/xnnpack/AlignedAllocator.h",
2557 ] + MICROKERNEL_TEST_HDRS,
2558 deps = MICROKERNEL_TEST_DEPS,
2559)
2560
2561xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 name = "f32_clamp_test",
2563 srcs = [
2564 "test/f32-clamp.cc",
2565 "test/clamp-microkernel-tester.h",
2566 ] + MICROKERNEL_TEST_HDRS,
2567 deps = MICROKERNEL_TEST_DEPS,
2568)
2569
2570xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002571 name = "f32_igemm_test",
2572 srcs = [
2573 "test/f32-igemm.cc",
2574 "test/gemm-microkernel-tester.h",
2575 "src/xnnpack/AlignedAllocator.h",
2576 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2577 deps = MICROKERNEL_TEST_DEPS,
2578)
2579
2580xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002581 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002582 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002583 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002584 "test/gemm-microkernel-tester.h",
2585 "src/xnnpack/AlignedAllocator.h",
2586 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002587 # Work-around for "too many sections" error
2588 mingw_copts = ["-Wa,-mbig-obj"],
2589 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002590 deps = MICROKERNEL_TEST_DEPS,
2591)
2592
2593xnnpack_unit_test(
2594 name = "f32_conv_hwc_test",
2595 srcs = [
2596 "test/f32-conv-hwc.cc",
2597 "test/conv-hwc-microkernel-tester.h",
2598 "src/xnnpack/AlignedAllocator.h",
2599 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2600 deps = MICROKERNEL_TEST_DEPS,
2601)
2602
2603xnnpack_unit_test(
2604 name = "f32_conv_hwc2spchw_test",
2605 srcs = [
2606 "test/f32-conv-hwc2spchw.cc",
2607 "test/conv-hwc2spchw-microkernel-tester.h",
2608 "src/xnnpack/AlignedAllocator.h",
2609 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2610 deps = MICROKERNEL_TEST_DEPS,
2611)
2612
2613xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002614 name = "f32_dwconv_test",
2615 srcs = [
2616 "test/f32-dwconv.cc",
2617 "test/dwconv-microkernel-tester.h",
2618 "src/xnnpack/AlignedAllocator.h",
2619 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2620 deps = MICROKERNEL_TEST_DEPS,
2621)
2622
2623xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002624 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002625 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002626 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002627 "test/dwconv-microkernel-tester.h",
2628 "src/xnnpack/AlignedAllocator.h",
2629 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2630 deps = MICROKERNEL_TEST_DEPS,
2631)
2632
2633xnnpack_unit_test(
2634 name = "f32_dwconv_spchw_test",
2635 srcs = [
2636 "test/f32-dwconv-spchw.cc",
2637 "test/dwconv-spchw-microkernel-tester.h",
2638 "src/xnnpack/AlignedAllocator.h",
2639 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2640 deps = MICROKERNEL_TEST_DEPS,
2641)
2642
2643xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002644 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002645 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002646 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002647 "test/gavgpool-microkernel-tester.h",
2648 "src/xnnpack/AlignedAllocator.h",
2649 ] + MICROKERNEL_TEST_HDRS,
2650 deps = MICROKERNEL_TEST_DEPS,
2651)
2652
2653xnnpack_unit_test(
2654 name = "f32_gavgpool_spchw_test",
2655 srcs = [
2656 "test/f32-gavgpool-spchw.cc",
2657 "test/gavgpool-spchw-microkernel-tester.h",
2658 "src/xnnpack/AlignedAllocator.h",
2659 ] + MICROKERNEL_TEST_HDRS,
2660 deps = MICROKERNEL_TEST_DEPS,
2661)
2662
2663xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002664 name = "f32_gemm_test",
2665 srcs = [
2666 "test/f32-gemm.cc",
2667 "test/gemm-microkernel-tester.h",
2668 "src/xnnpack/AlignedAllocator.h",
2669 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2670 deps = MICROKERNEL_TEST_DEPS,
2671)
2672
2673xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002674 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002676 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002677 "test/gemm-microkernel-tester.h",
2678 "src/xnnpack/AlignedAllocator.h",
2679 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2680 deps = MICROKERNEL_TEST_DEPS,
2681)
2682
2683xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002684 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002686 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687 "test/gemm-microkernel-tester.h",
2688 "src/xnnpack/AlignedAllocator.h",
2689 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2690 deps = MICROKERNEL_TEST_DEPS,
2691)
2692
2693xnnpack_unit_test(
2694 name = "f32_hswish_test",
2695 srcs = [
2696 "test/f32-hswish.cc",
2697 "test/hswish-microkernel-tester.h",
2698 ] + MICROKERNEL_TEST_HDRS,
2699 deps = MICROKERNEL_TEST_DEPS,
2700)
2701
2702xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002703 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002704 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002705 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706 "test/maxpool-microkernel-tester.h",
2707 ] + MICROKERNEL_TEST_HDRS,
2708 deps = MICROKERNEL_TEST_DEPS,
2709)
2710
2711xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002712 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002713 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002714 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002715 "test/avgpool-microkernel-tester.h",
2716 "src/xnnpack/AlignedAllocator.h",
2717 ] + MICROKERNEL_TEST_HDRS,
2718 deps = MICROKERNEL_TEST_DEPS,
2719)
2720
2721xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002722 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002723 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002724 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002725 "test/gemm-microkernel-tester.h",
2726 "src/xnnpack/AlignedAllocator.h",
2727 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2728 deps = MICROKERNEL_TEST_DEPS,
2729)
2730
2731xnnpack_unit_test(
2732 name = "f32_prelu_test",
2733 srcs = [
2734 "test/f32-prelu.cc",
2735 "test/prelu-microkernel-tester.h",
2736 "src/xnnpack/AlignedAllocator.h",
2737 ] + MICROKERNEL_TEST_HDRS,
2738 deps = MICROKERNEL_TEST_DEPS,
2739)
2740
2741xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002742 name = "f32_raddexpminusmax_test",
2743 srcs = [
2744 "test/f32-raddexpminusmax.cc",
2745 "test/raddexpminusmax-microkernel-tester.h",
2746 ] + MICROKERNEL_TEST_HDRS,
2747 deps = MICROKERNEL_TEST_DEPS,
2748)
2749
2750xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002751 name = "f32_raddextexp_test",
2752 srcs = [
2753 "test/f32-raddextexp.cc",
2754 "test/raddextexp-microkernel-tester.h",
2755 ] + MICROKERNEL_TEST_HDRS,
2756 deps = MICROKERNEL_TEST_DEPS,
2757)
2758
2759xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002760 name = "f32_raddstoreexpminusmax_test",
2761 srcs = [
2762 "test/f32-raddstoreexpminusmax.cc",
2763 "test/raddstoreexpminusmax-microkernel-tester.h",
2764 ] + MICROKERNEL_TEST_HDRS,
2765 deps = MICROKERNEL_TEST_DEPS,
2766)
2767
2768xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 name = "f32_rmax_test",
2770 srcs = [
2771 "test/f32-rmax.cc",
2772 "test/rmax-microkernel-tester.h",
2773 ] + MICROKERNEL_TEST_HDRS,
2774 deps = MICROKERNEL_TEST_DEPS,
2775)
2776
2777xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002778 name = "f32_sigmoid_test",
2779 srcs = [
2780 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002781 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002782 ] + MICROKERNEL_TEST_HDRS,
2783 deps = MICROKERNEL_TEST_DEPS,
2784)
2785
2786xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002787 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002788 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002789 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002790 "test/spmm-microkernel-tester.h",
2791 "src/xnnpack/AlignedAllocator.h",
2792 ] + MICROKERNEL_TEST_HDRS,
2793 deps = MICROKERNEL_TEST_DEPS,
2794)
2795
2796xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002797 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002799 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002800 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002801 ] + MICROKERNEL_TEST_HDRS,
2802 deps = MICROKERNEL_TEST_DEPS,
2803)
2804
2805xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002806 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002807 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002808 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002809 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002810 ] + MICROKERNEL_TEST_HDRS,
2811 deps = MICROKERNEL_TEST_DEPS,
2812)
2813
2814xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002815 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002816 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002817 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002818 "test/vbinary-microkernel-tester.h",
2819 ] + MICROKERNEL_TEST_HDRS,
2820 deps = MICROKERNEL_TEST_DEPS,
2821)
2822
2823xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002824 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002825 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002826 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002827 "test/vbinaryc-microkernel-tester.h",
2828 ] + MICROKERNEL_TEST_HDRS,
2829 deps = MICROKERNEL_TEST_DEPS,
2830)
2831
2832xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002833 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002834 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002835 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08002836 "test/vbinaryc-microkernel-tester.h",
2837 ] + MICROKERNEL_TEST_HDRS,
2838 deps = MICROKERNEL_TEST_DEPS,
2839)
2840
2841xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002842 name = "f32_vmax_test",
2843 srcs = [
2844 "test/f32-vmax.cc",
2845 "test/vbinary-microkernel-tester.h",
2846 ] + MICROKERNEL_TEST_HDRS,
2847 deps = MICROKERNEL_TEST_DEPS,
2848)
2849
2850xnnpack_unit_test(
2851 name = "f32_vmaxc_test",
2852 srcs = [
2853 "test/f32-vmaxc.cc",
2854 "test/vbinaryc-microkernel-tester.h",
2855 ] + MICROKERNEL_TEST_HDRS,
2856 deps = MICROKERNEL_TEST_DEPS,
2857)
2858
2859xnnpack_unit_test(
2860 name = "f32_vmin_test",
2861 srcs = [
2862 "test/f32-vmin.cc",
2863 "test/vbinary-microkernel-tester.h",
2864 ] + MICROKERNEL_TEST_HDRS,
2865 deps = MICROKERNEL_TEST_DEPS,
2866)
2867
2868xnnpack_unit_test(
2869 name = "f32_vminc_test",
2870 srcs = [
2871 "test/f32-vminc.cc",
2872 "test/vbinaryc-microkernel-tester.h",
2873 ] + MICROKERNEL_TEST_HDRS,
2874 deps = MICROKERNEL_TEST_DEPS,
2875)
2876
2877xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002878 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002880 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002881 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002882 ] + MICROKERNEL_TEST_HDRS,
2883 deps = MICROKERNEL_TEST_DEPS,
2884)
2885
2886xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002887 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002888 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002889 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002890 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 ] + MICROKERNEL_TEST_HDRS,
2892 deps = MICROKERNEL_TEST_DEPS,
2893)
2894
2895xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002896 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002898 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002899 "test/vmulcaddc-microkernel-tester.h",
2900 "src/xnnpack/AlignedAllocator.h",
2901 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2902 deps = MICROKERNEL_TEST_DEPS,
2903)
2904
2905xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002906 name = "f32_vscale_test",
2907 srcs = [
2908 "test/f32-vscale.cc",
2909 "test/vscale-microkernel-tester.h",
2910 ] + MICROKERNEL_TEST_HDRS,
2911 deps = MICROKERNEL_TEST_DEPS,
2912)
2913
2914xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002915 name = "f32_vscaleexpminusmax_test",
2916 srcs = [
2917 "test/f32-vscaleexpminusmax.cc",
2918 "test/vscaleexpminusmax-microkernel-tester.h",
2919 ] + MICROKERNEL_TEST_HDRS,
2920 deps = MICROKERNEL_TEST_DEPS,
2921)
2922
2923xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002924 name = "f32_vscaleextexp_test",
2925 srcs = [
2926 "test/f32-vscaleextexp.cc",
2927 "test/vscaleextexp-microkernel-tester.h",
2928 ] + MICROKERNEL_TEST_HDRS,
2929 deps = MICROKERNEL_TEST_DEPS,
2930)
2931
2932xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002933 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07002934 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002935 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002936 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002937 ] + MICROKERNEL_TEST_HDRS,
2938 deps = MICROKERNEL_TEST_DEPS,
2939)
2940
2941xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002942 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002943 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002944 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002945 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002946 ] + MICROKERNEL_TEST_HDRS,
2947 deps = MICROKERNEL_TEST_DEPS,
2948)
2949
2950xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002951 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002952 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002953 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002954 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002955 ] + MICROKERNEL_TEST_HDRS,
2956 deps = MICROKERNEL_TEST_DEPS,
2957)
2958
2959xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002960 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002962 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002963 "test/avgpool-microkernel-tester.h",
2964 "src/xnnpack/AlignedAllocator.h",
2965 ] + MICROKERNEL_TEST_HDRS,
2966 deps = MICROKERNEL_TEST_DEPS,
2967)
2968
2969xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002970 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002971 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002972 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002973 "test/gemm-microkernel-tester.h",
2974 "src/xnnpack/AlignedAllocator.h",
2975 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2976 deps = MICROKERNEL_TEST_DEPS,
2977)
2978
2979xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002980 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002981 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002982 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002983 "test/dwconv-microkernel-tester.h",
2984 "src/xnnpack/AlignedAllocator.h",
2985 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2986 deps = MICROKERNEL_TEST_DEPS,
2987)
2988
2989xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002990 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002992 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 "test/gavgpool-microkernel-tester.h",
2994 "src/xnnpack/AlignedAllocator.h",
2995 ] + MICROKERNEL_TEST_HDRS,
2996 deps = MICROKERNEL_TEST_DEPS,
2997)
2998
2999xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003000 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003002 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003003 "test/gemm-microkernel-tester.h",
3004 "src/xnnpack/AlignedAllocator.h",
3005 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3006 deps = MICROKERNEL_TEST_DEPS,
3007)
3008
3009xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003010 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003012 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 "test/vadd-microkernel-tester.h",
3014 ] + MICROKERNEL_TEST_HDRS,
3015 deps = MICROKERNEL_TEST_DEPS,
3016)
3017
3018xnnpack_unit_test(
3019 name = "u8_clamp_test",
3020 srcs = [
3021 "test/u8-clamp.cc",
3022 "test/clamp-microkernel-tester.h",
3023 ] + MICROKERNEL_TEST_HDRS,
3024 deps = MICROKERNEL_TEST_DEPS,
3025)
3026
3027xnnpack_unit_test(
3028 name = "u8_lut32norm_test",
3029 srcs = [
3030 "test/u8-lut32norm.cc",
3031 "test/lut-norm-microkernel-tester.h",
3032 ] + MICROKERNEL_TEST_HDRS,
3033 deps = MICROKERNEL_TEST_DEPS,
3034)
3035
3036xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003037 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003038 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003039 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003040 "test/maxpool-microkernel-tester.h",
3041 ] + MICROKERNEL_TEST_HDRS,
3042 deps = MICROKERNEL_TEST_DEPS,
3043)
3044
3045xnnpack_unit_test(
3046 name = "u8_rmax_test",
3047 srcs = [
3048 "test/u8-rmax.cc",
3049 "test/rmax-microkernel-tester.h",
3050 ] + MICROKERNEL_TEST_HDRS,
3051 deps = MICROKERNEL_TEST_DEPS,
3052)
3053
3054xnnpack_unit_test(
3055 name = "x32_packx_test",
3056 srcs = [
3057 "test/x32-packx.cc",
3058 "test/pack-microkernel-tester.h",
3059 "src/xnnpack/AlignedAllocator.h",
3060 ] + MICROKERNEL_TEST_HDRS,
3061 deps = MICROKERNEL_TEST_DEPS,
3062)
3063
3064xnnpack_unit_test(
3065 name = "x32_pad_test",
3066 srcs = [
3067 "test/x32-pad.cc",
3068 "test/pad-microkernel-tester.h",
3069 ] + MICROKERNEL_TEST_HDRS,
3070 deps = MICROKERNEL_TEST_DEPS,
3071)
3072
3073xnnpack_unit_test(
3074 name = "x32_unpool_test",
3075 srcs = [
3076 "test/x32-unpool.cc",
3077 "test/unpool-microkernel-tester.h",
3078 ] + MICROKERNEL_TEST_HDRS,
3079 deps = MICROKERNEL_TEST_DEPS,
3080)
3081
3082xnnpack_unit_test(
3083 name = "x32_zip_test",
3084 srcs = [
3085 "test/x32-zip.cc",
3086 "test/zip-microkernel-tester.h",
3087 ] + MICROKERNEL_TEST_HDRS,
3088 deps = MICROKERNEL_TEST_DEPS,
3089)
3090
3091xnnpack_unit_test(
3092 name = "x8_lut_test",
3093 srcs = [
3094 "test/x8-lut.cc",
3095 "test/lut-microkernel-tester.h",
3096 ] + MICROKERNEL_TEST_HDRS,
3097 deps = MICROKERNEL_TEST_DEPS,
3098)
3099
3100xnnpack_unit_test(
3101 name = "x8_zip_test",
3102 srcs = [
3103 "test/x8-zip.cc",
3104 "test/zip-microkernel-tester.h",
3105 ] + MICROKERNEL_TEST_HDRS,
3106 deps = MICROKERNEL_TEST_DEPS,
3107)
3108
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003109xnnpack_unit_test(
3110 name = "requantization_test",
3111 srcs = [
3112 "src/xnnpack/requantization-stubs.h",
3113 "test/requantization.cc",
3114 "test/requantization-tester.h",
3115 ] + MICROKERNEL_TEST_HDRS,
3116 deps = MICROKERNEL_TEST_DEPS,
3117)
3118
Marat Dukhan20c3b922020-03-10 03:45:06 -07003119########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120
3121xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003122 name = "operator_size_test",
3123 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003124 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125)
3126
Marat Dukhan20c3b922020-03-10 03:45:06 -07003127xnnpack_binary(
3128 name = "subgraph_size_test",
3129 srcs = ["test/subgraph-size.c"],
3130 deps = [":XNNPACK"],
3131)
3132
3133########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134
3135xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003136 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003138 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003139 "test/add-operator-tester.h",
3140 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003141 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142)
3143
3144xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003145 name = "add_nd_test",
3146 srcs = [
3147 "test/add-nd.cc",
3148 "test/binary-elementwise-operator-tester.h",
3149 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003150 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003151)
3152
3153xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003154 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003155 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003156 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003157 "test/argmax-pooling-operator-tester.h",
3158 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003159 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003160)
3161
3162xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003163 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003164 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003165 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166 "test/average-pooling-operator-tester.h",
3167 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003168 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169)
3170
3171xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003172 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003174 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175 "test/channel-pad-operator-tester.h",
3176 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003177 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178)
3179
3180xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003181 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003182 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003183 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 "test/channel-shuffle-operator-tester.h",
3185 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003186 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187)
3188
3189xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003190 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003192 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 "test/clamp-operator-tester.h",
3194 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003195 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003196)
3197
3198xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003199 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003201 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202 "test/convolution-operator-tester.h",
3203 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003204 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205)
3206
3207xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003208 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003209 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003210 "test/convolution-nchw.cc",
3211 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003212 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003213 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214)
3215
3216xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003217 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003218 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003219 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003220 "test/deconvolution-operator-tester.h",
3221 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003222 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003223)
3224
3225xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003226 name = "divide_nd_test",
3227 srcs = [
3228 "test/binary-elementwise-operator-tester.h",
3229 "test/divide-nd.cc",
3230 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003231 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003232)
3233
3234xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003235 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003237 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003238 "test/fully-connected-operator-tester.h",
3239 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003240 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003241)
3242
3243xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003244 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003246 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247 "test/global-average-pooling-operator-tester.h",
3248 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003249 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250)
3251
3252xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003253 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003254 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003255 "test/global-average-pooling-ncw.cc",
3256 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003257 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003258 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259)
3260
3261xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003262 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003263 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003264 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265 "test/hardswish-operator-tester.h",
3266 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003267 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268)
3269
3270xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003271 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003273 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003274 "test/leaky-relu-operator-tester.h",
3275 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003276 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277)
3278
3279xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003280 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003282 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003283 "test/max-pooling-operator-tester.h",
3284 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003285 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003286)
3287
3288xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003289 name = "maximum_nd_test",
3290 srcs = [
3291 "test/binary-elementwise-operator-tester.h",
3292 "test/maximum-nd.cc",
3293 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003294 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003295)
3296
3297xnnpack_unit_test(
3298 name = "minimum_nd_test",
3299 srcs = [
3300 "test/binary-elementwise-operator-tester.h",
3301 "test/minimum-nd.cc",
3302 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003303 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003304)
3305
3306xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003307 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003308 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003309 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003310 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003311 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003312 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003313)
3314
3315xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003316 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003317 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003318 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003319 "test/prelu-operator-tester.h",
3320 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003321 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322)
3323
3324xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003325 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003326 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003327 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003328 "test/resize-bilinear-operator-tester.h",
3329 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003330 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003331)
3332
3333xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003334 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003336 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337 "test/sigmoid-operator-tester.h",
3338 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003339 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003340)
3341
3342xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003343 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003345 "test/softmax-nc.cc",
3346 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003348 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003349)
3350
3351xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003352 name = "subtract_nd_test",
3353 srcs = [
3354 "test/binary-elementwise-operator-tester.h",
3355 "test/subtract-nd.cc",
3356 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003357 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003358)
3359
3360xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003361 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003363 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "test/unpooling-operator-tester.h",
3365 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003366 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003367)
3368
3369############################# Build configurations #############################
3370
Marat Dukhanb8642352019-10-30 15:43:02 -07003371# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003373 name = "xnn_enable_assembly_explicit_true",
3374 define_values = {"xnn_enable_assembly": "true"},
3375)
3376
3377# Disables usage of assembly kernels.
3378config_setting(
3379 name = "xnn_enable_assembly_explicit_false",
3380 define_values = {"xnn_enable_assembly": "false"},
3381)
3382
Marat Dukhan05702cf2020-03-26 15:41:33 -07003383# Disables usage of HMP-aware optimizations.
3384config_setting(
3385 name = "xnn_enable_hmp_explicit_false",
3386 define_values = {"xnn_enable_hmp": "false"},
3387)
3388
Marat Dukhanb8642352019-10-30 15:43:02 -07003389# Builds with -c dbg
3390config_setting(
3391 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003393 "compilation_mode": "dbg",
3394 },
3395)
3396
3397# Builds with -c opt
3398config_setting(
3399 name = "optimized_build",
3400 values = {
3401 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 },
3403)
3404
3405config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003406 name = "linux_k8",
3407 values = {"cpu": "k8"},
3408)
3409
3410config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003411 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003412 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003413)
3414
3415config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003416 name = "linux_armhf",
3417 values = {"cpu": "armhf"},
3418)
3419
3420config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421 name = "android",
3422 values = {"crosstool_top": "//external:android/crosstool"},
3423)
3424
3425config_setting(
3426 name = "android_armv7",
3427 values = {
3428 "crosstool_top": "//external:android/crosstool",
3429 "cpu": "armeabi-v7a",
3430 },
3431)
3432
3433config_setting(
3434 name = "android_arm64",
3435 values = {
3436 "crosstool_top": "//external:android/crosstool",
3437 "cpu": "arm64-v8a",
3438 },
3439)
3440
3441config_setting(
3442 name = "android_x86",
3443 values = {
3444 "crosstool_top": "//external:android/crosstool",
3445 "cpu": "x86",
3446 },
3447)
3448
3449config_setting(
3450 name = "android_x86_64",
3451 values = {
3452 "crosstool_top": "//external:android/crosstool",
3453 "cpu": "x86_64",
3454 },
3455)
3456
3457config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003458 name = "windows_x86_64",
3459 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003460)
3461
3462config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003463 name = "windows_x86_64_clang",
3464 values = {
3465 "compiler": "clang-cl",
3466 "cpu": "x64_windows",
3467 },
3468)
3469
3470config_setting(
3471 name = "windows_x86_64_mingw",
3472 values = {
3473 "compiler": "mingw-gcc",
3474 "cpu": "x64_windows",
3475 },
3476)
3477
3478config_setting(
3479 name = "windows_x86_64_msys",
3480 values = {
3481 "compiler": "msys-gcc",
3482 "cpu": "x64_windows",
3483 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003484)
3485
3486config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003487 name = "macos_x86_64",
3488 values = {
3489 "apple_platform_type": "macos",
3490 "cpu": "darwin",
3491 },
3492)
3493
3494config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003495 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003496 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497)
3498
3499config_setting(
3500 name = "emscripten_wasm",
3501 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003502 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003503 "cpu": "wasm",
3504 },
3505)
3506
3507config_setting(
3508 name = "emscripten_wasmsimd",
3509 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003510 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003512 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513 },
3514)
3515
3516config_setting(
3517 name = "emscripten_asmjs",
3518 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003519 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003520 "cpu": "asmjs",
3521 },
3522)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003523
3524config_setting(
3525 name = "ios_armv7",
3526 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003527 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003528 "cpu": "ios_armv7",
3529 },
3530)
3531
3532config_setting(
3533 name = "ios_arm64",
3534 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003535 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003536 "cpu": "ios_arm64",
3537 },
3538)
3539
3540config_setting(
3541 name = "ios_arm64e",
3542 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003543 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003544 "cpu": "ios_arm64e",
3545 },
3546)
3547
3548config_setting(
3549 name = "ios_x86",
3550 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003551 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003552 "cpu": "ios_i386",
3553 },
3554)
3555
3556config_setting(
3557 name = "ios_x86_64",
3558 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003559 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003560 "cpu": "ios_x86_64",
3561 },
3562)
3563
3564config_setting(
3565 name = "watchos_armv7k",
3566 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003567 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003568 "cpu": "watchos_armv7k",
3569 },
3570)
3571
3572config_setting(
3573 name = "watchos_arm64_32",
3574 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003575 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003576 "cpu": "watchos_arm64_32",
3577 },
3578)
3579
3580config_setting(
3581 name = "watchos_x86",
3582 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003583 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003584 "cpu": "watchos_i386",
3585 },
3586)
3587
3588config_setting(
3589 name = "watchos_x86_64",
3590 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003591 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003592 "cpu": "watchos_x86_64",
3593 },
3594)
3595
3596config_setting(
3597 name = "tvos_arm64",
3598 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003599 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003600 "cpu": "tvos_arm64",
3601 },
3602)
3603
3604config_setting(
3605 name = "tvos_x86_64",
3606 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003607 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003608 "cpu": "tvos_x86_64",
3609 },
3610)