blob: 319d9487148918ce8ff021cb56c44b421dd4140c [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/add-nc.c",
54 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
57 "src/operators/channel-pad-nc.c",
58 "src/operators/channel-shuffle-nc.c",
59 "src/operators/clamp-nc.c",
60 "src/operators/convolution-nchw.c",
61 "src/operators/convolution-nhwc.c",
62 "src/operators/deconvolution-nhwc.c",
63 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
66 "src/operators/hardswish-nc.c",
67 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
70 "src/operators/resize-bilinear-nhwc.c",
71 "src/operators/sigmoid-nc.c",
72 "src/operators/softmax-nc.c",
73 "src/operators/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",
Marat Dukhan1f29b802020-05-15 23:46:39 -070092 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
93 "src/f32-dwconv-chw/3x3p1-scalar.c",
94 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
95 "src/f32-dwconv-chw/5x5p2-scalar.c",
96 "src/f32-dwconv-chw/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",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700121 "src/f32-gavgpool-cw/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",
Marat Dukhan075088a2020-05-12 19:42:12 -0700238 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700239 "src/math/roundne-scalar-nearbyint.c",
240 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700241 "src/math/roundd-scalar-addsub.c",
242 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700243 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700244 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700245 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700246 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700247 "src/math/roundz-scalar-addsub.c",
248 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700249 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800250 "src/math/sigmoid-scalar-lut2048-p1-div.c",
251 "src/math/sigmoid-scalar-lut64-p2-div.c",
252 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700253 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
254 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700255 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700256 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
257 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700258 "src/q8-gemm/2x2-minmax-scalar.c",
259 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700260 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700261 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700262 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700263 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700264 "src/u8-rmax/scalar.c",
265 "src/x32-packx/x2-scalar.c",
266 "src/x32-packx/x3-scalar.c",
267 "src/x32-packx/x4-scalar.c",
268 "src/x32-pad/x2-scalar.c",
269 "src/x32-unpool/scalar.c",
270 "src/x32-zip/x2-scalar.c",
271 "src/x32-zip/x3-scalar.c",
272 "src/x32-zip/x4-scalar.c",
273 "src/x32-zip/xm-scalar.c",
274 "src/x8-lut/scalar.c",
275 "src/x8-zip/x2-scalar.c",
276 "src/x8-zip/x3-scalar.c",
277 "src/x8-zip/x4-scalar.c",
278 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700279 "src/requantization/precise-scalar.c",
280 "src/requantization/fp32-scalar.c",
281 "src/requantization/q31-scalar.c",
282 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700283]
284
Marat Dukhan436ebe62019-12-04 15:10:12 -0800285WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700286 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
287 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700288 "src/f32-clamp/gen/wasm-x1.c",
289 "src/f32-clamp/gen/wasm-x2.c",
290 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700291 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
292 "src/f32-dwconv/gen/up1x4-wasm.c",
293 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
294 "src/f32-dwconv/gen/up1x9-wasm.c",
295 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
296 "src/f32-dwconv/gen/up1x25-wasm.c",
297 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
298 "src/f32-dwconv/gen/up2x4-wasm.c",
299 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
300 "src/f32-dwconv/gen/up2x9-wasm.c",
301 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
302 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700303 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
304 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
305 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
306 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700307 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
308 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700309 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
310 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
311 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
312 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700313 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
314 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700315 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
316 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700317 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
318 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
319 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700320 "src/f32-gemm/gen/1x4-wasm.c",
321 "src/f32-gemm/gen/2x4-wasm.c",
322 "src/f32-gemm/gen/4x2-wasm.c",
323 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700324 "src/f32-gemm/gen/1x4-minmax-wasm.c",
325 "src/f32-gemm/gen/2x4-minmax-wasm.c",
326 "src/f32-gemm/gen/4x2-minmax-wasm.c",
327 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800328 "src/f32-hswish/gen/wasm-x1.c",
329 "src/f32-hswish/gen/wasm-x2.c",
330 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700331 "src/f32-igemm/gen/1x4-wasm.c",
332 "src/f32-igemm/gen/2x4-wasm.c",
333 "src/f32-igemm/gen/4x2-wasm.c",
334 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700335 "src/f32-igemm/gen/1x4-minmax-wasm.c",
336 "src/f32-igemm/gen/2x4-minmax-wasm.c",
337 "src/f32-igemm/gen/4x2-minmax-wasm.c",
338 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700339 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
340 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
341 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700342 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
343 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
344 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
345 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
346 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
347 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
348 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
349 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
350 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
351 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
352 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
353 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800354 "src/f32-vbinary/gen/vmax-wasm-x1.c",
355 "src/f32-vbinary/gen/vmax-wasm-x2.c",
356 "src/f32-vbinary/gen/vmax-wasm-x4.c",
357 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
358 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
359 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
360 "src/f32-vbinary/gen/vmin-wasm-x1.c",
361 "src/f32-vbinary/gen/vmin-wasm-x2.c",
362 "src/f32-vbinary/gen/vmin-wasm-x4.c",
363 "src/f32-vbinary/gen/vminc-wasm-x1.c",
364 "src/f32-vbinary/gen/vminc-wasm-x2.c",
365 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700366 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
367 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
368 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
369 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
370 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
371 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
374 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
375 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
376 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
377 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
378 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
379 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
380 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
381 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
382 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
383 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700384 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
385 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
386 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800387]
388
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800389PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800390 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800391 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800392 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700393 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
394 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700395 "src/f32-clamp/gen/psimd-x4.c",
396 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700397 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
398 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
399 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
400 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
401 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
402 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
403 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
404 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
405 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
406 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
407 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
408 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700409 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
410 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700411 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
412 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
413 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
414 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
415 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
416 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
417 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
418 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
419 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
420 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
421 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
422 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
423 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
424 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
425 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
427 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
428 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
429 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800430 "src/f32-hswish/gen/psimd-x4.c",
431 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700432 "src/f32-ibilinear/gen/psimd-c4.c",
433 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700434 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
435 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
436 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
437 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
438 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
439 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
440 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
441 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
442 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
443 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700444 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
445 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
446 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800448 "src/f32-prelu/gen/psimd-2x4.c",
449 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800450 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700451 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
452 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
453 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
454 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
455 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
456 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
457 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
458 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800459 "src/f32-vbinary/gen/vmax-psimd-x4.c",
460 "src/f32-vbinary/gen/vmax-psimd-x8.c",
461 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
462 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
463 "src/f32-vbinary/gen/vmin-psimd-x4.c",
464 "src/f32-vbinary/gen/vmin-psimd-x8.c",
465 "src/f32-vbinary/gen/vminc-psimd-x4.c",
466 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700467 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
468 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
469 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
470 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
471 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
472 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
473 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
474 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
475 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
476 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
477 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
478 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700479 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
480 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700481 "src/x32-packx/x4-psimd.c",
482 "src/x32-pad/x2-psimd.c",
483 "src/x32-unpool/psimd.c",
484 "src/x32-zip/x2-psimd.c",
485 "src/x32-zip/x3-psimd.c",
486 "src/x32-zip/x4-psimd.c",
487 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700488 "src/requantization/precise-psimd.c",
489 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700490]
491
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800492PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800493 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
494 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
495 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
496 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
497 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
498 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
499 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
500 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
501 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
502 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
503 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
504 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800505 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
506 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
507 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
508 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
509 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
510 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700511 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700512 "src/math/roundd-psimd-addsub.c",
513 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700514 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800515 "src/math/sigmoid-psimd-p5-div.c",
516]
517
Marat Dukhan08c4a432019-10-03 09:29:21 -0700518# ISA-specific micro-kernels
519NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700520 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
521 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700522 "src/f32-clamp/gen/neon-x4.c",
523 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700524 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
525 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700526 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
527 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700528 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
529 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700530 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
531 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700532 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
533 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
534 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
535 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700536 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
537 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
538 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
539 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700540 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
541 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
542 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
543 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700544 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700545 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
546 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700547 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
548 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
549 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
550 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
551 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
552 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
553 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
554 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
555 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
556 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
557 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
558 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
559 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
560 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
561 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
562 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
563 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
564 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
565 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
566 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
567 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
568 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
569 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
570 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
571 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
572 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
573 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
574 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
575 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
576 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
577 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800578 "src/f32-hswish/gen/neon-x4.c",
579 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700580 "src/f32-ibilinear/gen/neon-c4.c",
581 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700582 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
583 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
584 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
585 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
586 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
587 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
588 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
589 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
590 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
591 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
592 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
593 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
594 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
595 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
596 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
597 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700598 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
599 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
600 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700601 "src/f32-ppmm/gen/4x8-minmax-neon.c",
602 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800603 "src/f32-prelu/gen/neon-2x4.c",
604 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800605 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
606 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
607 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
608 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
609 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
610 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
611 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
612 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
613 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
614 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
615 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
616 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
617 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
618 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
619 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
620 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
621 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
622 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
623 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
624 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
625 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
626 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
627 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
628 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700629 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800630 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800631 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
632 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
633 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
634 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
635 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
636 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
637 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
638 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
639 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
640 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
641 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
642 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
643 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
644 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
645 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
646 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
647 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
648 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700649 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
650 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
651 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
652 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800653 "src/f32-vbinary/gen/vmax-neon-x4.c",
654 "src/f32-vbinary/gen/vmax-neon-x8.c",
655 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
656 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
657 "src/f32-vbinary/gen/vmin-neon-x4.c",
658 "src/f32-vbinary/gen/vmin-neon-x8.c",
659 "src/f32-vbinary/gen/vminc-neon-x4.c",
660 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700661 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
662 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
663 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
664 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
665 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
666 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
667 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
668 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
669 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
670 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700671 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
672 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
673 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
674 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700675 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700676 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
677 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700678 "src/q8-gemm/4x8-minmax-neon.c",
679 "src/q8-gemm/8x8-minmax-neon.c",
680 "src/q8-igemm/4x8-minmax-neon.c",
681 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700682 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700683 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700684 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700685 "src/u8-rmax/neon.c",
686 "src/x32-packx/x4-neon-st4.c",
687 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700688 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700689 "src/x32-zip/x2-neon.c",
690 "src/x32-zip/x3-neon.c",
691 "src/x32-zip/x4-neon.c",
692 "src/x32-zip/xm-neon.c",
693 "src/x8-zip/x2-neon.c",
694 "src/x8-zip/x3-neon.c",
695 "src/x8-zip/x4-neon.c",
696 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700697 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700698 "src/math/roundd-neon-addsub.c",
699 "src/math/roundd-neon-cvt.c",
700 "src/math/roundu-neon-addsub.c",
701 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700702 "src/math/roundz-neon-addsub.c",
703 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800704 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800705 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
706 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
707 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
708 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
709 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
710 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700711 "src/requantization/precise-neon.c",
712 "src/requantization/fp32-neon.c",
713 "src/requantization/q31-neon.c",
714 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700715]
716
717NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700718 "src/f32-ibilinear/gen/neonfma-c4.c",
719 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700720 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
721 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
722 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
723 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
724 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
725 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
726 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
727 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
728 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700729 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
730 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
731 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
732 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700733 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
734 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
735 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
736 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700737 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
738 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
739 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
740 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700741 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
742 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
743 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
744 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
745 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
746 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
747 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
748 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
749 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
750 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
751 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
752 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
753 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
754 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
755 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
756 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
757 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
758 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800759 "src/f32-hswish/gen/neonfma-x4.c",
760 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700761 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
762 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800763 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
764 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
765 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
766 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
767 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
768 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
769 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
770 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
771 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
772 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
773 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
774 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
775 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
776 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
777 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
778 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
779 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
780 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
781 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
782 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
783 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
784 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
785 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
786 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800787 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
832 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
833 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
834 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
835 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
836 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
837 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
838 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
839 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
840 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700841 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
842 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800843 "src/math/exp-neonfma-lut64-p2.c",
844 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800845 "src/math/expminus-neonfma-lut2048-p1.c",
846 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800847 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800848 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
849 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
850 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
851 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
852 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
853 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
854 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
855 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
856 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
857 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
858 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
859 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
860 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
861 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
862 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
863 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
864 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
865 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700866]
867
868AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700869 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
870 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
871 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
872 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
873 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
874 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700875 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
876 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
877 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
878 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
879 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
880 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
881 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
882 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
883 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
884 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
885 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
886 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
887 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
888 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
889 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
890 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
891 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
892 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
893 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
894 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700895 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
896 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700897 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
898 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700899 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
900 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700901 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
902 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700903 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
904 "src/f32-dwconv-chw/3x3p1-neonfma.c",
905 "src/f32-dwconv-chw/5x5p2-neonfma.c",
906 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
907 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800908 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
909 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
910 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
911 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
912 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
913 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
914 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
915 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
916 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
917 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
918 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
919 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
920 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
921 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
922 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
923 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
924 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
925 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700926 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
927 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
928 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
929 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
930 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
931 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
932 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
933 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
934 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
935 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
936 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
937 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
938 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
939 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
940 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
941 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
942 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
943 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800944 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
945 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
946 "src/math/sigmoid-neonfma-rr1-p5-div.c",
947 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
948 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
949 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700950]
951
Marat Dukhan8853b822020-05-07 12:19:01 -0700952NEONV8_UKERNELS = [
953 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700954 "src/math/roundd-neonv8.c",
955 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700956 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700957]
958
Marat Dukhan08c4a432019-10-03 09:29:21 -0700959AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -0700960 "src/f16-clamp/gen/neonfp16arith-x16.c",
961 "src/f16-clamp/gen/neonfp16arith-x8.c",
962 "src/f16-hswish/gen/neonfp16arith-x16.c",
963 "src/f16-hswish/gen/neonfp16arith-x8.c",
964 "src/f16-prelu/gen/neonfp16arith-2x16.c",
965 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -0700966 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
967 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
968 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
969 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
970 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
971 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
972 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
973 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
974 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
975 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
976 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
977 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
978 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
979 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
980 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
981 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
982 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
983 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
984 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
985 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
986 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
987 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
988 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
989 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
990 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
991 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
992 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
993 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -0700994 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
995 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
996 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
997 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
998 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
999 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1000 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1001 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001002 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1003 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1004 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1005 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001006 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1007 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1008 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1009 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1010 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1011 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1012 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1013 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1014 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1015 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1016 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1017 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001018 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1019 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1020 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1021 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1022 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1023 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1024 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1025 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001026]
1027
1028SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001029 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1030 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001031 "src/f32-clamp/gen/sse-x4.c",
1032 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001033 "src/f32-dwconv-chw/3x3p1-sse.c",
1034 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001035 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1036 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1037 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1038 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1039 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1040 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1041 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1042 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1043 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1044 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1045 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1046 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001047 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001048 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1049 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001050 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1051 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1052 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1053 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1054 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1055 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1056 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1057 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1058 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1059 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1060 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1061 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1062 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001063 "src/f32-hswish/gen/sse-x4.c",
1064 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001065 "src/f32-ibilinear/gen/sse-c4.c",
1066 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001067 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1068 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1069 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1070 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1071 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1072 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1073 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001074 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1075 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1076 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001077 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001078 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001079 "src/f32-spmm/gen/4x1-minmax-sse.c",
1080 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001081 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1082 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1083 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1084 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1085 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1086 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1087 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1088 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001089 "src/f32-vbinary/gen/vmax-sse-x4.c",
1090 "src/f32-vbinary/gen/vmax-sse-x8.c",
1091 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1092 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1093 "src/f32-vbinary/gen/vmin-sse-x4.c",
1094 "src/f32-vbinary/gen/vmin-sse-x8.c",
1095 "src/f32-vbinary/gen/vminc-sse-x4.c",
1096 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001097 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1098 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1099 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1100 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1101 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1102 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1103 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1104 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1105 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1106 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1107 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1108 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001109 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1110 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001111 "src/x32-packx/x4-sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001112 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001113 "src/math/roundd-sse-addsub.c",
1114 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001115 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001116]
1117
1118SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001119 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1120 "src/f32-argmaxpool/4x-sse2-c4.c",
1121 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001122 "src/f32-prelu/gen/sse2-2x4.c",
1123 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001124 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1125 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1126 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1127 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1128 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1129 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1130 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1131 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1133 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1134 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1135 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001136 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001137 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001138 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001139 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001140 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1141 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001142 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1143 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001144 "src/q8-igemm/4x4c2-minmax-sse2.c",
1145 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001146 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1147 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001148 "src/q8-gemm/2x4c8-minmax-sse2.c",
1149 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001150 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001151 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001152 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001153 "src/u8-rmax/sse2.c",
1154 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001155 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001156 "src/x32-zip/x2-sse2.c",
1157 "src/x32-zip/x3-sse2.c",
1158 "src/x32-zip/x4-sse2.c",
1159 "src/x32-zip/xm-sse2.c",
1160 "src/x8-zip/x2-sse2.c",
1161 "src/x8-zip/x3-sse2.c",
1162 "src/x8-zip/x4-sse2.c",
1163 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001164 "src/math/exp-sse2-p5.c",
1165 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001166 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001167 "src/math/roundd-sse2-cvt.c",
1168 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001169 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001170 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001171 "src/requantization/precise-sse2.c",
1172 "src/requantization/fp32-sse2.c",
1173 "src/requantization/q31-sse2.c",
1174 "src/requantization/gemmlowp-sse2.c",
1175]
1176
1177SSSE3_UKERNELS = [
1178 "src/requantization/precise-ssse3.c",
1179 "src/requantization/q31-ssse3.c",
1180 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001181]
1182
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001183SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001184 "src/f32-prelu/gen/sse41-2x4.c",
1185 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001186 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001187 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001188 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001189 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001190 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1191 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001192 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001193 "src/math/roundd-sse41.c",
1194 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001195 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001196 "src/requantization/precise-sse4.c",
1197 "src/requantization/q31-sse4.c",
1198 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001199]
1200
Marat Dukhan08c4a432019-10-03 09:29:21 -07001201AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001202 "src/f32-clamp/gen/avx-x8.c",
1203 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001204 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1205 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1206 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1207 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1208 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1209 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1210 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1211 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1212 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1213 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1214 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1215 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1216 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1217 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1218 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1219 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1220 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1221 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1222 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1223 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1224 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1225 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1226 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1227 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1228 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1229 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1230 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1231 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1232 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1233 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001234 "src/f32-hswish/gen/avx-x8.c",
1235 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001236 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1237 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1238 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1239 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1240 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1241 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1242 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1243 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1244 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001245 "src/f32-prelu/gen/avx-2x8.c",
1246 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001247 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001248 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1249 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1250 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1251 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1252 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1253 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1254 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1255 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001256 "src/f32-vbinary/gen/vmax-avx-x8.c",
1257 "src/f32-vbinary/gen/vmax-avx-x16.c",
1258 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1259 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1260 "src/f32-vbinary/gen/vmin-avx-x8.c",
1261 "src/f32-vbinary/gen/vmin-avx-x16.c",
1262 "src/f32-vbinary/gen/vminc-avx-x8.c",
1263 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001264 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1265 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1266 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1267 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1268 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1269 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1270 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1271 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1272 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1273 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1274 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1275 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001276 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001277]
1278
Marat Dukhanfda12b82019-11-21 12:27:59 -08001279FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001280 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1281 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1282 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1283 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1284 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1285 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1286 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1287 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1288 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1289 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1290 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1291 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1292 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1293 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1294 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1295 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1296 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1297 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1298 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1299 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1300 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1301 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1302 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1303 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1304 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1305 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1306 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1307 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1308 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1309 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1310 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1311 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1312 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1313 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1314 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1315 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1316 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1317 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1318 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1319 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001320 "src/f32-hswish/gen/fma3-x8.c",
1321 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001322 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1323 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1324 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1325 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1326 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1327 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1328 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1329 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1330 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1331 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1332 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1333 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1334 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1335 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001336]
1337
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001338AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001339 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1340 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1341 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1342 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1343 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1344 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1345 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1346 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1347 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1348 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1349 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1350 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1351 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1352 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1353 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1354 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1355 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1356 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1357 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1358 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1359 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1360 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1361 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1362 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1363 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1364 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1365 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1366 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1367 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1368 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1369 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1370 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1373 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1374 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001375 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1376 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1377 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1378 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1379 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1380 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1381 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1382 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1383 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1384 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1387 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1388 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1389 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1390 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1403 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1404 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001405 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1406 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1407 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1408 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1409 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1410 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1411 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1412 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1415 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1416 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1417 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1418 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1419 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1420 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1421 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1422 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1423 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1424 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1425 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1426 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1427 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1428 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001429 "src/math/exp-avx2-p5.c",
1430 "src/math/exp-avx2-perm-p3.c",
1431 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001432 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001433 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001434 "src/math/sigmoid-avx2-rr2-p5-div.c",
1435 "src/math/sigmoid-avx2-rr1-p5-div.c",
1436 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1437 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1438 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1439 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001440]
1441
Marat Dukhan08c4a432019-10-03 09:29:21 -07001442AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001443 "src/f32-clamp/gen/avx512f-x16.c",
1444 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001445 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1446 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1447 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1448 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1449 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1450 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1451 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1452 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1453 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1454 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1455 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1456 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1457 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1458 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1459 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1460 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1461 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1462 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1463 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1464 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1465 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1466 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1467 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1468 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001469 "src/f32-hswish/gen/avx512f-x16.c",
1470 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001471 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1472 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1473 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1474 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1475 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1476 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001477 "src/f32-prelu/gen/avx512f-2x16.c",
1478 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001479 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1480 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1481 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1482 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1483 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1484 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1485 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1486 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1487 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1488 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1489 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1490 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1491 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1492 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1493 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1494 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1495 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1496 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1497 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1498 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1499 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1500 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1501 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1502 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1503 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1504 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1505 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1506 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1507 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1508 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1509 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1510 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1511 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1512 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1513 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1514 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001515 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001516 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1517 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1518 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1519 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1520 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1521 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1522 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1523 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001524 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1525 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1526 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1527 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1528 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1529 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1530 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1531 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001532 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1533 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1534 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1535 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1536 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1537 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1538 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1539 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1540 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1541 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1542 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1543 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001544 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001545 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1546 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1547 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1548 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1549 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1550 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1551 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1552 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1553 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1554 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1555 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1556 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1557 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1558 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1559 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1560 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1561 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1562 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1563 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1564 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1565 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1566 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1567 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1568 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001569 "src/math/exp-avx512f-p5-scalef.c",
1570 "src/math/exp-avx512f-p5.c",
1571 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001572 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001573 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001574]
1575
1576AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001577 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001578 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001579 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1580 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1581 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1582 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1583 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1584 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1585 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1586 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1587 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1588 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001589]
1590
1591AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001592 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001593 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1594 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001595 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001596 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1597 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001598 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1599 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1600 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001601 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001602 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1603 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1604 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001605 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001606 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1607 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001608 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001609 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1610 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1611 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1612 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1613 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1614 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1615 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1616 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1617 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1618 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1619 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1620 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1621 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1622 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1623 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1624 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1625 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1626 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1627 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1628 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1629 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001630 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001631 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1632 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1633 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1634 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1635 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1636 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1637 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1638 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1639 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1640 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1641 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1642 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1643 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1644 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1645 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1646 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1647 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1648 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1649 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1650 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1651 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1652 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1653 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1654 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1655 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1656 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1657 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1658 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1659 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1660 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1661 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1662 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1663 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1664 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1665 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1666 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1667 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1668 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001669]
1670
Marat Dukhan1b354632020-03-23 12:50:22 -07001671INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001672 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001673 "src/xnnpack/argmaxpool.h",
1674 "src/xnnpack/avgpool.h",
1675 "src/xnnpack/clamp.h",
1676 "src/xnnpack/common.h",
1677 "src/xnnpack/conv.h",
1678 "src/xnnpack/dwconv.h",
1679 "src/xnnpack/gavgpool.h",
1680 "src/xnnpack/gemm.h",
1681 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001682 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001684 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 "src/xnnpack/lut.h",
1686 "src/xnnpack/math.h",
1687 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001688 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001689 "src/xnnpack/packx.h",
1690 "src/xnnpack/pad.h",
1691 "src/xnnpack/params.h",
1692 "src/xnnpack/pavgpool.h",
1693 "src/xnnpack/ppmm.h",
1694 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001695 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001696 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001697 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001698 "src/xnnpack/rmax.h",
1699 "src/xnnpack/scalar-utils.h",
1700 "src/xnnpack/spmm.h",
1701 "src/xnnpack/unpool.h",
1702 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001703 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001704 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001705 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001706 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001707 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001708 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001709 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001710]
1711
1712INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713 "include/xnnpack.h",
1714 "src/xnnpack/allocator.h",
1715 "src/xnnpack/compute.h",
1716 "src/xnnpack/im2col.h",
1717 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001718 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001719 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720 "src/xnnpack/operator.h",
1721 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001722 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001724 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001725 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001726]
1727
Marat Dukhan1b354632020-03-23 12:50:22 -07001728ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001729 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730]
1731
Marat Dukhan1b354632020-03-23 12:50:22 -07001732MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001733 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001734 "include/xnnpack.h",
1735]
1736
Marat Dukhan1b354632020-03-23 12:50:22 -07001737MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001738 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001739 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001740 "src/xnnpack/requantization.h",
1741 "include/xnnpack.h",
1742]
1743
1744OPERATOR_TEST_PARAMS_HDRS = [
1745 "src/xnnpack/params.h",
1746 "src/xnnpack/common.h",
1747]
1748
1749WEIGHTS_PACK_HDRS = [
1750 "src/xnnpack/pack.h",
1751 "src/xnnpack/operator.h",
1752 "src/xnnpack/compute.h",
1753]
1754
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001755LOGGING_COPTS = select({
1756 # No logging in optimized mode
1757 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1758 # Full logging in debug mode
1759 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1760 # Error-only logging in default (fastbuild) mode
1761 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1762})
1763
1764LOGGING_HDRS = [
1765 "src/xnnpack/log.h",
1766]
1767
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001769 name = "tables",
1770 srcs = TABLE_SRCS,
1771 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001772 gcc_copts = xnnpack_gcc_std_copts(),
1773 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001774)
1775
1776xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 name = "scalar_ukernels",
1778 srcs = SCALAR_UKERNELS,
1779 hdrs = INTERNAL_HDRS,
1780 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001781 gcc_copts = xnnpack_gcc_std_copts(),
1782 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001783 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001784 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001785 "@FP16",
1786 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001787 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001788 ],
1789)
1790
1791xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001792 name = "wasm_ukernels",
1793 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001794 gcc_copts = xnnpack_gcc_std_copts(),
1795 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001796 wasm_srcs = WASM_UKERNELS,
1797 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001798 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001799 "@FP16",
1800 "@FXdiv",
1801 "@pthreadpool",
1802 ],
1803)
1804
1805xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001806 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807 hdrs = INTERNAL_HDRS,
1808 aarch32_copts = [
1809 "-marm",
1810 "-mfpu=neon",
1811 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001812 gcc_copts = xnnpack_gcc_std_copts(),
1813 gcc_x86_copts = ["-msse2"],
1814 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001815 optimized_copts = [
1816 "-O3",
1817 "-ffast-math",
1818 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001819 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001821 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822 "@FP16",
1823 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001824 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825 ],
1826)
1827
1828xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001829 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001830 hdrs = INTERNAL_HDRS,
1831 aarch32_copts = [
1832 "-marm",
1833 "-mfpu=neon",
1834 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001835 gcc_copts = xnnpack_gcc_std_copts(),
1836 gcc_x86_copts = ["-msse2"],
1837 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001838 optimized_copts = [
1839 "-O3",
1840 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001841 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001842 deps = [
1843 ":tables",
1844 "@FP16",
1845 "@psimd",
1846 "@pthreadpool",
1847 ],
1848)
1849
1850xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001851 name = "neon_ukernels",
1852 hdrs = INTERNAL_HDRS,
1853 aarch32_copts = [
1854 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001855 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001856 "-mfpu=neon",
1857 ],
1858 aarch32_srcs = NEON_UKERNELS,
1859 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001860 gcc_copts = xnnpack_gcc_std_copts(),
1861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001862 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001863 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001864 "@FP16",
1865 "@pthreadpool",
1866 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867)
1868
1869xnnpack_cc_library(
1870 name = "neonfma_ukernels",
1871 hdrs = INTERNAL_HDRS,
1872 aarch32_copts = [
1873 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001874 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001875 "-mfpu=neon-vfpv4",
1876 ],
1877 aarch32_srcs = NEONFMA_UKERNELS,
1878 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001879 gcc_copts = xnnpack_gcc_std_copts(),
1880 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001881 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001882 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001883 "@FP16",
1884 "@pthreadpool",
1885 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001886)
1887
1888xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001889 name = "neonv8_ukernels",
1890 hdrs = INTERNAL_HDRS,
1891 aarch32_copts = [
1892 "-marm",
1893 "-march=armv8-a",
1894 "-mfpu=neon-fp-armv8",
1895 ],
1896 aarch32_srcs = NEONV8_UKERNELS,
1897 aarch64_srcs = NEONV8_UKERNELS,
1898 gcc_copts = xnnpack_gcc_std_copts(),
1899 msvc_copts = xnnpack_msvc_std_copts(),
1900 deps = [
1901 ":tables",
1902 "@FP16",
1903 "@pthreadpool",
1904 ],
1905)
1906
1907xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001908 name = "neonfp16arith_ukernels",
1909 hdrs = INTERNAL_HDRS,
1910 aarch64_copts = ["-march=armv8.2-a+fp16"],
1911 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001912 gcc_copts = xnnpack_gcc_std_copts(),
1913 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001914 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001915 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001916 "@FP16",
1917 "@pthreadpool",
1918 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001919)
1920
1921xnnpack_cc_library(
1922 name = "sse2_ukernels",
1923 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001924 gcc_copts = xnnpack_gcc_std_copts(),
1925 gcc_x86_copts = ["-msse2"],
1926 msvc_copts = xnnpack_msvc_std_copts(),
1927 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001928 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001929 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001930 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001931 "@FP16",
1932 "@pthreadpool",
1933 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001934)
1935
1936xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001937 name = "ssse3_ukernels",
1938 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001939 gcc_copts = xnnpack_gcc_std_copts(),
1940 gcc_x86_copts = ["-mssse3"],
1941 msvc_copts = xnnpack_msvc_std_copts(),
1942 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001943 x86_srcs = SSSE3_UKERNELS,
1944 deps = [
1945 ":tables",
1946 "@FP16",
1947 "@pthreadpool",
1948 ],
1949)
1950
1951xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001952 name = "sse41_ukernels",
1953 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001954 gcc_copts = xnnpack_gcc_std_copts(),
1955 gcc_x86_copts = ["-msse4.1"],
1956 msvc_copts = xnnpack_msvc_std_copts(),
1957 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001958 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001959 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001960 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001961 "@FP16",
1962 "@pthreadpool",
1963 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001964)
1965
1966xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 name = "avx_ukernels",
1968 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001969 gcc_copts = xnnpack_gcc_std_copts(),
1970 gcc_x86_copts = ["-mavx"],
1971 msvc_copts = xnnpack_msvc_std_copts(),
1972 msvc_x86_32_copts = ["/arch:AVX"],
1973 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001974 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001975 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001976 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001977 "@FP16",
1978 "@pthreadpool",
1979 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001980)
1981
1982xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001983 name = "fma3_ukernels",
1984 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001985 gcc_copts = xnnpack_gcc_std_copts(),
1986 gcc_x86_copts = ["-mfma"],
1987 msvc_copts = xnnpack_msvc_std_copts(),
1988 msvc_x86_32_copts = ["/arch:AVX"],
1989 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001990 x86_srcs = FMA3_UKERNELS,
1991 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001992 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001993 "@FP16",
1994 "@pthreadpool",
1995 ],
1996)
1997
1998xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001999 name = "avx2_ukernels",
2000 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002001 gcc_copts = xnnpack_gcc_std_copts(),
2002 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002003 "-mfma",
2004 "-mavx2",
2005 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002006 msvc_copts = xnnpack_msvc_std_copts(),
2007 msvc_x86_32_copts = ["/arch:AVX2"],
2008 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002009 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002010 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002011 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002012 "@FP16",
2013 "@pthreadpool",
2014 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002015)
2016
2017xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002018 name = "avx512f_ukernels",
2019 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002020 gcc_copts = xnnpack_gcc_std_copts(),
2021 gcc_x86_copts = ["-mavx512f"],
2022 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2023 msvc_copts = xnnpack_msvc_std_copts(),
2024 msvc_x86_32_copts = ["/arch:AVX512"],
2025 msvc_x86_64_copts = ["/arch:AVX512"],
2026 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002027 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002028 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002029 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002030 "@FP16",
2031 "@pthreadpool",
2032 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002033)
2034
2035xnnpack_cc_library(
2036 name = "asm_ukernels",
2037 hdrs = ["src/xnnpack/assembly.h"],
2038 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002039 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002040 aarch64_srcs = AARCH64_ASM_UKERNELS,
2041)
2042
2043xnnpack_aggregate_library(
2044 name = "ukernels",
2045 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002046 ":neon_ukernels",
2047 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002048 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002049 ":asm_ukernels",
2050 ],
2051 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002052 ":neon_ukernels",
2053 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002054 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002055 ":neonfp16arith_ukernels",
2056 ":asm_ukernels",
2057 ],
2058 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002059 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002060 ":psimd_fastmath_ukernels",
2061 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002062 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002063 wasm_deps = [":wasm_ukernels"],
2064 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002065 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002066 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002067 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002068 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002069 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002070 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002071 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 ":avx512f_ukernels",
2073 ],
2074)
2075
2076xnnpack_cc_library(
2077 name = "im2col",
2078 srcs = ["src/im2col.c"],
2079 hdrs = [
2080 "src/xnnpack/common.h",
2081 "src/xnnpack/im2col.h",
2082 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002083 gcc_copts = xnnpack_gcc_std_copts(),
2084 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085)
2086
2087xnnpack_cc_library(
2088 name = "indirection",
2089 srcs = ["src/indirection.c"],
2090 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002091 gcc_copts = xnnpack_gcc_std_copts(),
2092 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093 deps = [
2094 "@FP16",
2095 "@FXdiv",
2096 "@pthreadpool",
2097 ],
2098)
2099
2100xnnpack_cc_library(
2101 name = "operator_run",
2102 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002103 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002104 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002105 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2106 "//conditions:default": [],
2107 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002108 gcc_copts = xnnpack_gcc_std_copts(),
2109 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002110 deps = [
2111 "@FP16",
2112 "@FXdiv",
2113 "@clog",
2114 "@pthreadpool",
2115 ],
2116)
2117
Chao Mei6ddfc602020-05-13 22:29:36 -07002118xnnpack_cc_library(
2119 name = "memory_planner",
2120 srcs = ["src/memory-planner.c"],
2121 hdrs = INTERNAL_HDRS,
2122 defines = select({
2123 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2124 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2125 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2126 }),
2127 gcc_copts = xnnpack_gcc_std_copts(),
2128 msvc_copts = xnnpack_msvc_std_copts(),
2129 deps = [
2130 "@pthreadpool",
2131 ],
2132)
2133
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134cc_library(
2135 name = "enable_assembly",
2136 defines = select({
2137 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2138 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002139 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002140 }),
2141)
2142
Marat Dukhancf056b22019-10-07 10:26:29 -07002143xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002144 name = "operators",
2145 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002146 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002148 ],
2149 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002150 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002151 "-Isrc",
2152 "-Iinclude",
2153 ] + select({
2154 ":debug_build": [],
2155 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002156 }) + select({
2157 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2158 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002160 gcc_copts = xnnpack_gcc_std_copts(),
2161 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002162 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002163 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002164 "@FP16",
2165 "@FXdiv",
2166 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002167 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002168 ],
2169)
2170
Marat Dukhan10a38082020-04-17 03:58:35 -07002171xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002172 name = "XNNPACK",
2173 srcs = [
2174 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002175 "src/runtime.c",
2176 "src/subgraph.c",
2177 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002178 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002179 hdrs = ["include/xnnpack.h"],
2180 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002181 "-Isrc",
2182 "-Iinclude",
2183 ] + select({
2184 ":debug_build": [],
2185 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002186 }) + select({
2187 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2188 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002189 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002190 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002191 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002192 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002193 visibility = xnnpack_visibility(),
2194 deps = [
2195 ":enable_assembly",
Chao Mei6ddfc602020-05-13 22:29:36 -07002196 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002197 ":ukernels",
2198 ":operator_run",
2199 ":operators",
2200 "@clog",
2201 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002202 ] + select({
2203 ":emscripten": [],
2204 "//conditions:default": ["@cpuinfo"],
2205 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206)
2207
Marat Dukhan10a38082020-04-17 03:58:35 -07002208xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002209 name = "xnnpack_operators_nhwc_f32",
2210 srcs = [
2211 "src/init.c",
2212 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002213 hdrs = ["include/xnnpack.h"],
2214 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002215 "-Isrc",
2216 "-Iinclude",
2217 ] + select({
2218 ":debug_build": [],
2219 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002220 }) + select({
2221 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2222 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002223 }),
2224 defines = [
2225 "XNN_NO_Q8_OPERATORS",
2226 "XNN_NO_U8_OPERATORS",
2227 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002228 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002229 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002230 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002232 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002233 visibility = xnnpack_visibility(),
2234 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002235 ":enable_assembly",
2236 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237 ":operator_run",
2238 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002239 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002240 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002241 ] + select({
2242 ":emscripten": [],
2243 "//conditions:default": ["@cpuinfo"],
2244 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002245)
2246
Marat Dukhancf056b22019-10-07 10:26:29 -07002247xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002248 name = "bench_utils",
2249 srcs = ["bench/utils.cc"],
2250 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002251 deps = [
2252 "@com_google_benchmark//:benchmark",
2253 "@cpuinfo",
2254 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002255)
2256
Frank Barchard7e955972019-10-11 10:34:25 -07002257######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258
2259xnnpack_benchmark(
2260 name = "q8_gemm_bench",
2261 srcs = [
2262 "bench/gemm.h",
2263 "bench/q8-gemm.cc",
2264 "src/xnnpack/AlignedAllocator.h",
2265 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002266 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002267 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2268)
2269
2270xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002271 name = "f16_igemm_bench",
2272 srcs = [
2273 "bench/f16-igemm.cc",
2274 "bench/conv.h",
2275 "bench/google/conv.h",
2276 "src/xnnpack/AlignedAllocator.h",
2277 ] + MICROKERNEL_BENCHMARK_HDRS,
2278 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2279)
2280
2281xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002282 name = "f16_gemm_bench",
2283 srcs = [
2284 "bench/f16-gemm.cc",
2285 "bench/gemm.h",
2286 "src/xnnpack/AlignedAllocator.h",
2287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002288 deps = MICROKERNEL_BENCHMARK_DEPS,
2289)
2290
2291xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002292 name = "f16_spmm_bench",
2293 srcs = [
2294 "bench/f16-spmm.cc",
2295 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002296 "src/xnnpack/AlignedAllocator.h",
2297 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002298 deps = MICROKERNEL_BENCHMARK_DEPS,
2299)
2300
2301xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002302 name = "f32_igemm_bench",
2303 srcs = [
2304 "bench/f32-igemm.cc",
2305 "bench/conv.h",
2306 "src/xnnpack/AlignedAllocator.h",
2307 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002308 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309)
2310
2311xnnpack_benchmark(
2312 name = "f32_conv_hwc_bench",
2313 srcs = [
2314 "bench/f32-conv-hwc.cc",
2315 "bench/dconv.h",
2316 "src/xnnpack/AlignedAllocator.h",
2317 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002318 deps = MICROKERNEL_BENCHMARK_DEPS,
2319)
2320
2321xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002322 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07002323 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002324 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07002325 "bench/dconv.h",
2326 "src/xnnpack/AlignedAllocator.h",
2327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002328 deps = MICROKERNEL_BENCHMARK_DEPS,
2329)
2330
2331xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002332 name = "f32_dwconv_bench",
2333 srcs = [
2334 "bench/f32-dwconv.cc",
2335 "bench/dwconv.h",
2336 "src/xnnpack/AlignedAllocator.h",
2337 ] + MICROKERNEL_BENCHMARK_HDRS,
2338 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2339)
2340
2341xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002342 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002343 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002344 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345 "bench/dwconv.h",
2346 "src/xnnpack/AlignedAllocator.h",
2347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2348 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2349)
2350
2351xnnpack_benchmark(
2352 name = "f32_gemm_bench",
2353 srcs = [
2354 "bench/f32-gemm.cc",
2355 "bench/gemm.h",
2356 "src/xnnpack/AlignedAllocator.h",
2357 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002358 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002359 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002360)
2361
2362xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002363 name = "f32_raddexpminusmax_bench",
2364 srcs = [
2365 "bench/f32-raddexpminusmax.cc",
2366 "src/xnnpack/AlignedAllocator.h",
2367 ] + MICROKERNEL_BENCHMARK_HDRS,
2368 deps = MICROKERNEL_BENCHMARK_DEPS,
2369)
2370
2371xnnpack_benchmark(
2372 name = "f32_raddextexp_bench",
2373 srcs = [
2374 "bench/f32-raddextexp.cc",
2375 "src/xnnpack/AlignedAllocator.h",
2376 ] + MICROKERNEL_BENCHMARK_HDRS,
2377 deps = MICROKERNEL_BENCHMARK_DEPS,
2378)
2379
2380xnnpack_benchmark(
2381 name = "f32_raddstoreexpminusmax_bench",
2382 srcs = [
2383 "bench/f32-raddstoreexpminusmax.cc",
2384 "src/xnnpack/AlignedAllocator.h",
2385 ] + MICROKERNEL_BENCHMARK_HDRS,
2386 deps = MICROKERNEL_BENCHMARK_DEPS,
2387)
2388
2389xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002390 name = "f32_rmax_bench",
2391 srcs = [
2392 "bench/f32-rmax.cc",
2393 "src/xnnpack/AlignedAllocator.h",
2394 ] + MICROKERNEL_BENCHMARK_HDRS,
2395 deps = MICROKERNEL_BENCHMARK_DEPS,
2396)
2397
2398xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002399 name = "f32_sigmoid_bench",
2400 srcs = [
2401 "bench/f32-sigmoid.cc",
2402 "src/xnnpack/AlignedAllocator.h",
2403 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002404 deps = MICROKERNEL_BENCHMARK_DEPS,
2405)
2406
2407xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002408 name = "f32_spmm_bench",
2409 srcs = [
2410 "bench/f32-spmm.cc",
2411 "bench/gemm.h",
2412 "src/xnnpack/AlignedAllocator.h",
2413 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002414 deps = MICROKERNEL_BENCHMARK_DEPS,
2415)
2416
2417xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002418 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002419 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002420 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002421 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002422 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002423 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002424)
2425
2426xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002427 name = "f32_vscaleexpminusmax_bench",
2428 srcs = [
2429 "bench/f32-vscaleexpminusmax.cc",
2430 "src/xnnpack/AlignedAllocator.h",
2431 ] + MICROKERNEL_BENCHMARK_HDRS,
2432 deps = MICROKERNEL_BENCHMARK_DEPS,
2433)
2434
2435xnnpack_benchmark(
2436 name = "f32_vscaleextexp_bench",
2437 srcs = [
2438 "bench/f32-vscaleextexp.cc",
2439 "src/xnnpack/AlignedAllocator.h",
2440 ] + MICROKERNEL_BENCHMARK_HDRS,
2441 deps = MICROKERNEL_BENCHMARK_DEPS,
2442)
2443
2444xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002445 name = "f32_im2col_gemm_bench",
2446 srcs = [
2447 "bench/f32-im2col-gemm.cc",
2448 "bench/conv.h",
2449 "src/xnnpack/AlignedAllocator.h",
2450 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2451 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2452)
2453
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002454xnnpack_benchmark(
2455 name = "requantization_bench",
2456 srcs = [
2457 "bench/requantization.cc",
2458 "src/xnnpack/requantization-stubs.h",
2459 "src/xnnpack/AlignedAllocator.h",
2460 ] + MICROKERNEL_BENCHMARK_HDRS,
2461 deps = MICROKERNEL_BENCHMARK_DEPS,
2462)
2463
Marat Dukhanffbf96a2020-05-14 02:59:08 -07002464xnnpack_benchmark(
2465 name = "rounding_bench",
2466 srcs = [
2467 "bench/rounding.cc",
2468 "src/xnnpack/math-stubs.h",
2469 "src/xnnpack/AlignedAllocator.h",
2470 ] + MICROKERNEL_BENCHMARK_HDRS,
2471 deps = MICROKERNEL_BENCHMARK_DEPS,
2472)
2473
Marat Dukhan08c4a432019-10-03 09:29:21 -07002474########################### Benchmarks for operators ###########################
2475
2476xnnpack_benchmark(
2477 name = "add_bench",
2478 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002479 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002480)
2481
2482xnnpack_benchmark(
2483 name = "average_pooling_bench",
2484 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002485 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002486 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002487 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002488)
2489
2490xnnpack_benchmark(
2491 name = "channel_shuffle_bench",
2492 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002493 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002494)
2495
2496xnnpack_benchmark(
2497 name = "convolution_bench",
2498 srcs = ["bench/convolution.cc"],
2499 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002500 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002501 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002502)
2503
2504xnnpack_benchmark(
2505 name = "deconvolution_bench",
2506 srcs = ["bench/deconvolution.cc"],
2507 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002508 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002509 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002510)
2511
2512xnnpack_benchmark(
2513 name = "global_average_pooling_bench",
2514 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002515 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002516)
2517
2518xnnpack_benchmark(
2519 name = "max_pooling_bench",
2520 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002521 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002522)
2523
2524xnnpack_benchmark(
2525 name = "sigmoid_bench",
2526 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002527 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002528 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002529 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002530)
2531
2532xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002533 name = "prelu_bench",
2534 srcs = ["bench/prelu.cc"],
2535 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002536 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002537 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002538)
2539
2540xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002541 name = "softmax_bench",
2542 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002543 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002544 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002545 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002546)
2547
Marat Dukhanc068bb62019-10-04 13:24:39 -07002548############################# End-to-end benchmarks ############################
2549
2550cc_library(
2551 name = "mobilenet_v1",
2552 srcs = ["models/mobilenet-v1.cc"],
2553 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002554 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002555 linkstatic = True,
2556 deps = [
2557 ":XNNPACK",
2558 "@pthreadpool",
2559 ],
2560)
2561
2562cc_library(
2563 name = "mobilenet_v2",
2564 srcs = ["models/mobilenet-v2.cc"],
2565 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002566 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002567 linkstatic = True,
2568 deps = [
2569 ":XNNPACK",
2570 "@pthreadpool",
2571 ],
2572)
2573
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002574cc_library(
2575 name = "mobilenet_v3_large",
2576 srcs = ["models/mobilenet-v3-large.cc"],
2577 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002578 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002579 linkstatic = True,
2580 deps = [
2581 ":XNNPACK",
2582 "@pthreadpool",
2583 ],
2584)
2585
2586cc_library(
2587 name = "mobilenet_v3_small",
2588 srcs = ["models/mobilenet-v3-small.cc"],
2589 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002590 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002591 linkstatic = True,
2592 deps = [
2593 ":XNNPACK",
2594 "@pthreadpool",
2595 ],
2596)
2597
Marat Dukhanc068bb62019-10-04 13:24:39 -07002598xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002599 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002600 srcs = [
2601 "bench/f32-dwconv-e2e.cc",
2602 "bench/end2end.h",
2603 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002604 deps = MICROKERNEL_BENCHMARK_DEPS + [
2605 ":XNNPACK",
2606 ":mobilenet_v1",
2607 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002608 ":mobilenet_v3_large",
2609 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002610 ],
2611)
2612
2613xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002614 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002615 srcs = [
2616 "bench/f32-gemm-e2e.cc",
2617 "bench/end2end.h",
2618 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002619 deps = MICROKERNEL_BENCHMARK_DEPS + [
2620 ":XNNPACK",
2621 ":mobilenet_v1",
2622 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002623 ":mobilenet_v3_large",
2624 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002625 ],
2626)
2627
2628xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002629 name = "end2end_bench",
2630 srcs = ["bench/end2end.cc"],
2631 deps = [
2632 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002633 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002634 ":mobilenet_v1",
2635 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002636 ":mobilenet_v3_large",
2637 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002638 "@pthreadpool",
2639 ],
2640)
2641
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002642#################### Accuracy evaluation for math functions ####################
2643
2644xnnpack_benchmark(
2645 name = "f32_exp_eval",
2646 srcs = [
2647 "eval/f32-exp.cc",
2648 "src/xnnpack/AlignedAllocator.h",
2649 ] + ACCURACY_EVAL_HDRS,
2650 deps = ACCURACY_EVAL_DEPS,
2651)
2652
Marat Dukhan515c9772019-10-17 18:07:57 -07002653xnnpack_benchmark(
2654 name = "f32_expminus_eval",
2655 srcs = [
2656 "eval/f32-expminus.cc",
2657 "src/xnnpack/AlignedAllocator.h",
2658 ] + ACCURACY_EVAL_HDRS,
2659 deps = ACCURACY_EVAL_DEPS,
2660)
2661
Marat Dukhan98ba4412019-10-23 02:14:28 -07002662xnnpack_benchmark(
2663 name = "f32_extexp_eval",
2664 srcs = [
2665 "eval/f32-extexp.cc",
2666 "src/xnnpack/AlignedAllocator.h",
2667 ] + ACCURACY_EVAL_HDRS,
2668 deps = ACCURACY_EVAL_DEPS,
2669)
2670
Marat Dukhan8853b822020-05-07 12:19:01 -07002671xnnpack_unit_test(
2672 name = "f32_roundne_eval",
2673 srcs = [
2674 "eval/f32-roundne.cc",
2675 "src/xnnpack/AlignedAllocator.h",
2676 "src/xnnpack/math-stubs.h",
2677 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002678 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002679 deps = MICROKERNEL_TEST_DEPS,
2680)
2681
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002682xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002683 name = "f32_roundd_eval",
2684 srcs = [
2685 "eval/f32-roundd.cc",
2686 "src/xnnpack/AlignedAllocator.h",
2687 "src/xnnpack/math-stubs.h",
2688 ] + MICROKERNEL_TEST_HDRS,
2689 automatic = False,
2690 deps = MICROKERNEL_TEST_DEPS,
2691)
2692
2693xnnpack_unit_test(
2694 name = "f32_roundu_eval",
2695 srcs = [
2696 "eval/f32-roundu.cc",
2697 "src/xnnpack/AlignedAllocator.h",
2698 "src/xnnpack/math-stubs.h",
2699 ] + MICROKERNEL_TEST_HDRS,
2700 automatic = False,
2701 deps = MICROKERNEL_TEST_DEPS,
2702)
2703
2704xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002705 name = "f32_roundz_eval",
2706 srcs = [
2707 "eval/f32-roundz.cc",
2708 "src/xnnpack/AlignedAllocator.h",
2709 "src/xnnpack/math-stubs.h",
2710 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002711 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002712 deps = MICROKERNEL_TEST_DEPS,
2713)
2714
Marat Dukhan346a9e52019-11-15 09:06:30 -08002715xnnpack_benchmark(
2716 name = "f32_sigmoid_eval",
2717 srcs = [
2718 "eval/f32-sigmoid.cc",
2719 "src/xnnpack/AlignedAllocator.h",
2720 ] + ACCURACY_EVAL_HDRS,
2721 deps = ACCURACY_EVAL_DEPS,
2722)
2723
Marat Dukhan08c4a432019-10-03 09:29:21 -07002724######################### Unit tests for micro-kernels #########################
2725
2726xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002727 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002728 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002729 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002730 "test/gemm-microkernel-tester.h",
2731 "src/xnnpack/AlignedAllocator.h",
2732 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2733 deps = MICROKERNEL_TEST_DEPS,
2734)
2735
2736xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002737 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002738 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002739 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002740 "test/spmm-microkernel-tester.h",
2741 "src/xnnpack/AlignedAllocator.h",
2742 ] + MICROKERNEL_TEST_HDRS,
2743 deps = MICROKERNEL_TEST_DEPS,
2744)
2745
2746xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 name = "f32_argmaxpool_test",
2748 srcs = [
2749 "test/f32-argmaxpool.cc",
2750 "test/argmaxpool-microkernel-tester.h",
2751 "src/xnnpack/AlignedAllocator.h",
2752 ] + MICROKERNEL_TEST_HDRS,
2753 deps = MICROKERNEL_TEST_DEPS,
2754)
2755
2756xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002757 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002758 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002759 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002760 "test/avgpool-microkernel-tester.h",
2761 "src/xnnpack/AlignedAllocator.h",
2762 ] + MICROKERNEL_TEST_HDRS,
2763 deps = MICROKERNEL_TEST_DEPS,
2764)
2765
2766xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002767 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002768 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002769 "test/f32-ibilinear.cc",
2770 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002771 "src/xnnpack/AlignedAllocator.h",
2772 ] + MICROKERNEL_TEST_HDRS,
2773 deps = MICROKERNEL_TEST_DEPS,
2774)
2775
2776xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002777 name = "f16_clamp_test",
2778 srcs = [
2779 "test/f16-clamp.cc",
2780 "test/clamp-microkernel-tester.h",
2781 ] + MICROKERNEL_TEST_HDRS,
2782 deps = MICROKERNEL_TEST_DEPS,
2783)
2784
2785xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002786 name = "f32_clamp_test",
2787 srcs = [
2788 "test/f32-clamp.cc",
2789 "test/clamp-microkernel-tester.h",
2790 ] + MICROKERNEL_TEST_HDRS,
2791 deps = MICROKERNEL_TEST_DEPS,
2792)
2793
2794xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002795 name = "f32_igemm_test",
2796 srcs = [
2797 "test/f32-igemm.cc",
2798 "test/gemm-microkernel-tester.h",
2799 "src/xnnpack/AlignedAllocator.h",
2800 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2801 deps = MICROKERNEL_TEST_DEPS,
2802)
2803
2804xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002805 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002807 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002808 "test/gemm-microkernel-tester.h",
2809 "src/xnnpack/AlignedAllocator.h",
2810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002811 # Work-around for "too many sections" error
2812 mingw_copts = ["-Wa,-mbig-obj"],
2813 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 deps = MICROKERNEL_TEST_DEPS,
2815)
2816
2817xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002818 name = "f16_igemm_minmax_test",
2819 srcs = [
2820 "test/f16-igemm-minmax.cc",
2821 "test/gemm-microkernel-tester.h",
2822 "src/xnnpack/AlignedAllocator.h",
2823 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2824 deps = MICROKERNEL_TEST_DEPS,
2825)
2826
2827xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002828 name = "f32_conv_hwc_test",
2829 srcs = [
2830 "test/f32-conv-hwc.cc",
2831 "test/conv-hwc-microkernel-tester.h",
2832 "src/xnnpack/AlignedAllocator.h",
2833 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2834 deps = MICROKERNEL_TEST_DEPS,
2835)
2836
2837xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002838 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002839 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002840 "test/f32-conv-hwc2chw.cc",
2841 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002842 "src/xnnpack/AlignedAllocator.h",
2843 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2844 deps = MICROKERNEL_TEST_DEPS,
2845)
2846
2847xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002848 name = "f32_dwconv_test",
2849 srcs = [
2850 "test/f32-dwconv.cc",
2851 "test/dwconv-microkernel-tester.h",
2852 "src/xnnpack/AlignedAllocator.h",
2853 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2854 deps = MICROKERNEL_TEST_DEPS,
2855)
2856
2857xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002858 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002859 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002860 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002861 "test/dwconv-microkernel-tester.h",
2862 "src/xnnpack/AlignedAllocator.h",
2863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2864 deps = MICROKERNEL_TEST_DEPS,
2865)
2866
2867xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002868 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002869 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002870 "test/f32-dwconv-chw.cc",
2871 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 "src/xnnpack/AlignedAllocator.h",
2873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2874 deps = MICROKERNEL_TEST_DEPS,
2875)
2876
2877xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002878 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002880 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881 "test/gavgpool-microkernel-tester.h",
2882 "src/xnnpack/AlignedAllocator.h",
2883 ] + MICROKERNEL_TEST_HDRS,
2884 deps = MICROKERNEL_TEST_DEPS,
2885)
2886
2887xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002888 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002889 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002890 "test/f32-gavgpool-cw.cc",
2891 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002892 "src/xnnpack/AlignedAllocator.h",
2893 ] + MICROKERNEL_TEST_HDRS,
2894 deps = MICROKERNEL_TEST_DEPS,
2895)
2896
2897xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002898 name = "f32_gemm_test",
2899 srcs = [
2900 "test/f32-gemm.cc",
2901 "test/gemm-microkernel-tester.h",
2902 "src/xnnpack/AlignedAllocator.h",
2903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2904 deps = MICROKERNEL_TEST_DEPS,
2905)
2906
2907xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002908 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002909 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002910 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 "test/gemm-microkernel-tester.h",
2912 "src/xnnpack/AlignedAllocator.h",
2913 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2914 deps = MICROKERNEL_TEST_DEPS,
2915)
2916
2917xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002918 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002920 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 "test/gemm-microkernel-tester.h",
2922 "src/xnnpack/AlignedAllocator.h",
2923 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2924 deps = MICROKERNEL_TEST_DEPS,
2925)
2926
2927xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002928 name = "f16_hswish_test",
2929 srcs = [
2930 "test/f16-hswish.cc",
2931 "test/hswish-microkernel-tester.h",
2932 ] + MICROKERNEL_TEST_HDRS,
2933 deps = MICROKERNEL_TEST_DEPS,
2934)
2935
2936xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937 name = "f32_hswish_test",
2938 srcs = [
2939 "test/f32-hswish.cc",
2940 "test/hswish-microkernel-tester.h",
2941 ] + MICROKERNEL_TEST_HDRS,
2942 deps = MICROKERNEL_TEST_DEPS,
2943)
2944
2945xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002946 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002947 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002948 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 "test/maxpool-microkernel-tester.h",
2950 ] + MICROKERNEL_TEST_HDRS,
2951 deps = MICROKERNEL_TEST_DEPS,
2952)
2953
2954xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002955 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002956 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002957 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958 "test/avgpool-microkernel-tester.h",
2959 "src/xnnpack/AlignedAllocator.h",
2960 ] + MICROKERNEL_TEST_HDRS,
2961 deps = MICROKERNEL_TEST_DEPS,
2962)
2963
2964xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002965 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002967 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002968 "test/gemm-microkernel-tester.h",
2969 "src/xnnpack/AlignedAllocator.h",
2970 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2971 deps = MICROKERNEL_TEST_DEPS,
2972)
2973
2974xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002975 name = "f16_prelu_test",
2976 srcs = [
2977 "test/f16-prelu.cc",
2978 "test/prelu-microkernel-tester.h",
2979 "src/xnnpack/AlignedAllocator.h",
2980 ] + MICROKERNEL_TEST_HDRS,
2981 deps = MICROKERNEL_TEST_DEPS,
2982)
2983
2984xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 name = "f32_prelu_test",
2986 srcs = [
2987 "test/f32-prelu.cc",
2988 "test/prelu-microkernel-tester.h",
2989 "src/xnnpack/AlignedAllocator.h",
2990 ] + MICROKERNEL_TEST_HDRS,
2991 deps = MICROKERNEL_TEST_DEPS,
2992)
2993
2994xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002995 name = "f32_raddexpminusmax_test",
2996 srcs = [
2997 "test/f32-raddexpminusmax.cc",
2998 "test/raddexpminusmax-microkernel-tester.h",
2999 ] + MICROKERNEL_TEST_HDRS,
3000 deps = MICROKERNEL_TEST_DEPS,
3001)
3002
3003xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003004 name = "f32_raddextexp_test",
3005 srcs = [
3006 "test/f32-raddextexp.cc",
3007 "test/raddextexp-microkernel-tester.h",
3008 ] + MICROKERNEL_TEST_HDRS,
3009 deps = MICROKERNEL_TEST_DEPS,
3010)
3011
3012xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003013 name = "f32_raddstoreexpminusmax_test",
3014 srcs = [
3015 "test/f32-raddstoreexpminusmax.cc",
3016 "test/raddstoreexpminusmax-microkernel-tester.h",
3017 ] + MICROKERNEL_TEST_HDRS,
3018 deps = MICROKERNEL_TEST_DEPS,
3019)
3020
3021xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 name = "f32_rmax_test",
3023 srcs = [
3024 "test/f32-rmax.cc",
3025 "test/rmax-microkernel-tester.h",
3026 ] + MICROKERNEL_TEST_HDRS,
3027 deps = MICROKERNEL_TEST_DEPS,
3028)
3029
3030xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003031 name = "f32_sigmoid_test",
3032 srcs = [
3033 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003034 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003035 ] + MICROKERNEL_TEST_HDRS,
3036 deps = MICROKERNEL_TEST_DEPS,
3037)
3038
3039xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003040 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003041 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003042 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 "test/spmm-microkernel-tester.h",
3044 "src/xnnpack/AlignedAllocator.h",
3045 ] + MICROKERNEL_TEST_HDRS,
3046 deps = MICROKERNEL_TEST_DEPS,
3047)
3048
3049xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003050 name = "f16_vadd_minmax_test",
3051 srcs = [
3052 "test/f16-vadd-minmax.cc",
3053 "test/vbinary-microkernel-tester.h",
3054 ] + MICROKERNEL_TEST_HDRS,
3055 deps = MICROKERNEL_TEST_DEPS,
3056)
3057
3058xnnpack_unit_test(
3059 name = "f16_vaddc_minmax_test",
3060 srcs = [
3061 "test/f16-vaddc-minmax.cc",
3062 "test/vbinaryc-microkernel-tester.h",
3063 ] + MICROKERNEL_TEST_HDRS,
3064 deps = MICROKERNEL_TEST_DEPS,
3065)
3066
3067xnnpack_unit_test(
3068 name = "f16_vdiv_minmax_test",
3069 srcs = [
3070 "test/f16-vdiv-minmax.cc",
3071 "test/vbinary-microkernel-tester.h",
3072 ] + MICROKERNEL_TEST_HDRS,
3073 deps = MICROKERNEL_TEST_DEPS,
3074)
3075
3076xnnpack_unit_test(
3077 name = "f16_vdivc_minmax_test",
3078 srcs = [
3079 "test/f16-vdivc-minmax.cc",
3080 "test/vbinaryc-microkernel-tester.h",
3081 ] + MICROKERNEL_TEST_HDRS,
3082 deps = MICROKERNEL_TEST_DEPS,
3083)
3084
3085xnnpack_unit_test(
3086 name = "f16_vrdivc_minmax_test",
3087 srcs = [
3088 "test/f16-vrdivc-minmax.cc",
3089 "test/vbinaryc-microkernel-tester.h",
3090 ] + MICROKERNEL_TEST_HDRS,
3091 deps = MICROKERNEL_TEST_DEPS,
3092)
3093
3094xnnpack_unit_test(
3095 name = "f16_vmax_test",
3096 srcs = [
3097 "test/f16-vmax.cc",
3098 "test/vbinary-microkernel-tester.h",
3099 ] + MICROKERNEL_TEST_HDRS,
3100 deps = MICROKERNEL_TEST_DEPS,
3101)
3102
3103xnnpack_unit_test(
3104 name = "f16_vmaxc_test",
3105 srcs = [
3106 "test/f16-vmaxc.cc",
3107 "test/vbinaryc-microkernel-tester.h",
3108 ] + MICROKERNEL_TEST_HDRS,
3109 deps = MICROKERNEL_TEST_DEPS,
3110)
3111
3112xnnpack_unit_test(
3113 name = "f16_vmin_test",
3114 srcs = [
3115 "test/f16-vmin.cc",
3116 "test/vbinary-microkernel-tester.h",
3117 ] + MICROKERNEL_TEST_HDRS,
3118 deps = MICROKERNEL_TEST_DEPS,
3119)
3120
3121xnnpack_unit_test(
3122 name = "f16_vminc_test",
3123 srcs = [
3124 "test/f16-vminc.cc",
3125 "test/vbinaryc-microkernel-tester.h",
3126 ] + MICROKERNEL_TEST_HDRS,
3127 deps = MICROKERNEL_TEST_DEPS,
3128)
3129
3130xnnpack_unit_test(
3131 name = "f16_vmul_minmax_test",
3132 srcs = [
3133 "test/f16-vmul-minmax.cc",
3134 "test/vbinary-microkernel-tester.h",
3135 ] + MICROKERNEL_TEST_HDRS,
3136 deps = MICROKERNEL_TEST_DEPS,
3137)
3138
3139xnnpack_unit_test(
3140 name = "f16_vmulc_minmax_test",
3141 srcs = [
3142 "test/f16-vmulc-minmax.cc",
3143 "test/vbinaryc-microkernel-tester.h",
3144 ] + MICROKERNEL_TEST_HDRS,
3145 deps = MICROKERNEL_TEST_DEPS,
3146)
3147
3148xnnpack_unit_test(
3149 name = "f16_vsub_minmax_test",
3150 srcs = [
3151 "test/f16-vsub-minmax.cc",
3152 "test/vbinary-microkernel-tester.h",
3153 ] + MICROKERNEL_TEST_HDRS,
3154 deps = MICROKERNEL_TEST_DEPS,
3155)
3156
3157xnnpack_unit_test(
3158 name = "f16_vsubc_minmax_test",
3159 srcs = [
3160 "test/f16-vsubc-minmax.cc",
3161 "test/vbinaryc-microkernel-tester.h",
3162 ] + MICROKERNEL_TEST_HDRS,
3163 deps = MICROKERNEL_TEST_DEPS,
3164)
3165
3166xnnpack_unit_test(
3167 name = "f16_vrsubc_minmax_test",
3168 srcs = [
3169 "test/f16-vrsubc-minmax.cc",
3170 "test/vbinaryc-microkernel-tester.h",
3171 ] + MICROKERNEL_TEST_HDRS,
3172 deps = MICROKERNEL_TEST_DEPS,
3173)
3174
3175xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003176 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003177 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003178 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003179 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003180 ] + MICROKERNEL_TEST_HDRS,
3181 deps = MICROKERNEL_TEST_DEPS,
3182)
3183
3184xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003185 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003186 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003187 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003188 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 ] + MICROKERNEL_TEST_HDRS,
3190 deps = MICROKERNEL_TEST_DEPS,
3191)
3192
3193xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003194 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003195 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003196 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003197 "test/vbinary-microkernel-tester.h",
3198 ] + MICROKERNEL_TEST_HDRS,
3199 deps = MICROKERNEL_TEST_DEPS,
3200)
3201
3202xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003203 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003204 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003205 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003206 "test/vbinaryc-microkernel-tester.h",
3207 ] + MICROKERNEL_TEST_HDRS,
3208 deps = MICROKERNEL_TEST_DEPS,
3209)
3210
3211xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003212 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003213 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003214 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003215 "test/vbinaryc-microkernel-tester.h",
3216 ] + MICROKERNEL_TEST_HDRS,
3217 deps = MICROKERNEL_TEST_DEPS,
3218)
3219
3220xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003221 name = "f32_vmax_test",
3222 srcs = [
3223 "test/f32-vmax.cc",
3224 "test/vbinary-microkernel-tester.h",
3225 ] + MICROKERNEL_TEST_HDRS,
3226 deps = MICROKERNEL_TEST_DEPS,
3227)
3228
3229xnnpack_unit_test(
3230 name = "f32_vmaxc_test",
3231 srcs = [
3232 "test/f32-vmaxc.cc",
3233 "test/vbinaryc-microkernel-tester.h",
3234 ] + MICROKERNEL_TEST_HDRS,
3235 deps = MICROKERNEL_TEST_DEPS,
3236)
3237
3238xnnpack_unit_test(
3239 name = "f32_vmin_test",
3240 srcs = [
3241 "test/f32-vmin.cc",
3242 "test/vbinary-microkernel-tester.h",
3243 ] + MICROKERNEL_TEST_HDRS,
3244 deps = MICROKERNEL_TEST_DEPS,
3245)
3246
3247xnnpack_unit_test(
3248 name = "f32_vminc_test",
3249 srcs = [
3250 "test/f32-vminc.cc",
3251 "test/vbinaryc-microkernel-tester.h",
3252 ] + MICROKERNEL_TEST_HDRS,
3253 deps = MICROKERNEL_TEST_DEPS,
3254)
3255
3256xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003257 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003258 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003259 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003260 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003261 ] + MICROKERNEL_TEST_HDRS,
3262 deps = MICROKERNEL_TEST_DEPS,
3263)
3264
3265xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003266 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003267 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003268 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003269 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270 ] + MICROKERNEL_TEST_HDRS,
3271 deps = MICROKERNEL_TEST_DEPS,
3272)
3273
3274xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003275 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003277 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278 "test/vmulcaddc-microkernel-tester.h",
3279 "src/xnnpack/AlignedAllocator.h",
3280 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3281 deps = MICROKERNEL_TEST_DEPS,
3282)
3283
3284xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003285 name = "f32_vscale_test",
3286 srcs = [
3287 "test/f32-vscale.cc",
3288 "test/vscale-microkernel-tester.h",
3289 ] + MICROKERNEL_TEST_HDRS,
3290 deps = MICROKERNEL_TEST_DEPS,
3291)
3292
3293xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003294 name = "f32_vscaleexpminusmax_test",
3295 srcs = [
3296 "test/f32-vscaleexpminusmax.cc",
3297 "test/vscaleexpminusmax-microkernel-tester.h",
3298 ] + MICROKERNEL_TEST_HDRS,
3299 deps = MICROKERNEL_TEST_DEPS,
3300)
3301
3302xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003303 name = "f32_vscaleextexp_test",
3304 srcs = [
3305 "test/f32-vscaleextexp.cc",
3306 "test/vscaleextexp-microkernel-tester.h",
3307 ] + MICROKERNEL_TEST_HDRS,
3308 deps = MICROKERNEL_TEST_DEPS,
3309)
3310
3311xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003312 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003313 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003314 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003315 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003316 ] + MICROKERNEL_TEST_HDRS,
3317 deps = MICROKERNEL_TEST_DEPS,
3318)
3319
3320xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003321 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003322 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003323 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003324 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003325 ] + MICROKERNEL_TEST_HDRS,
3326 deps = MICROKERNEL_TEST_DEPS,
3327)
3328
3329xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003330 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003331 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003332 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003333 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003334 ] + MICROKERNEL_TEST_HDRS,
3335 deps = MICROKERNEL_TEST_DEPS,
3336)
3337
3338xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003339 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003340 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003341 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 "test/avgpool-microkernel-tester.h",
3343 "src/xnnpack/AlignedAllocator.h",
3344 ] + MICROKERNEL_TEST_HDRS,
3345 deps = MICROKERNEL_TEST_DEPS,
3346)
3347
3348xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003349 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003351 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003352 "test/gemm-microkernel-tester.h",
3353 "src/xnnpack/AlignedAllocator.h",
3354 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3355 deps = MICROKERNEL_TEST_DEPS,
3356)
3357
3358xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003359 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003361 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 "test/dwconv-microkernel-tester.h",
3363 "src/xnnpack/AlignedAllocator.h",
3364 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3365 deps = MICROKERNEL_TEST_DEPS,
3366)
3367
3368xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003369 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003371 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 "test/gavgpool-microkernel-tester.h",
3373 "src/xnnpack/AlignedAllocator.h",
3374 ] + MICROKERNEL_TEST_HDRS,
3375 deps = MICROKERNEL_TEST_DEPS,
3376)
3377
3378xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003379 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003381 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003382 "test/gemm-microkernel-tester.h",
3383 "src/xnnpack/AlignedAllocator.h",
3384 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3385 deps = MICROKERNEL_TEST_DEPS,
3386)
3387
3388xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003389 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003390 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003391 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003392 "test/vadd-microkernel-tester.h",
3393 ] + MICROKERNEL_TEST_HDRS,
3394 deps = MICROKERNEL_TEST_DEPS,
3395)
3396
3397xnnpack_unit_test(
3398 name = "u8_clamp_test",
3399 srcs = [
3400 "test/u8-clamp.cc",
3401 "test/clamp-microkernel-tester.h",
3402 ] + MICROKERNEL_TEST_HDRS,
3403 deps = MICROKERNEL_TEST_DEPS,
3404)
3405
3406xnnpack_unit_test(
3407 name = "u8_lut32norm_test",
3408 srcs = [
3409 "test/u8-lut32norm.cc",
3410 "test/lut-norm-microkernel-tester.h",
3411 ] + MICROKERNEL_TEST_HDRS,
3412 deps = MICROKERNEL_TEST_DEPS,
3413)
3414
3415xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003416 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003417 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003418 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003419 "test/maxpool-microkernel-tester.h",
3420 ] + MICROKERNEL_TEST_HDRS,
3421 deps = MICROKERNEL_TEST_DEPS,
3422)
3423
3424xnnpack_unit_test(
3425 name = "u8_rmax_test",
3426 srcs = [
3427 "test/u8-rmax.cc",
3428 "test/rmax-microkernel-tester.h",
3429 ] + MICROKERNEL_TEST_HDRS,
3430 deps = MICROKERNEL_TEST_DEPS,
3431)
3432
3433xnnpack_unit_test(
3434 name = "x32_packx_test",
3435 srcs = [
3436 "test/x32-packx.cc",
3437 "test/pack-microkernel-tester.h",
3438 "src/xnnpack/AlignedAllocator.h",
3439 ] + MICROKERNEL_TEST_HDRS,
3440 deps = MICROKERNEL_TEST_DEPS,
3441)
3442
3443xnnpack_unit_test(
3444 name = "x32_pad_test",
3445 srcs = [
3446 "test/x32-pad.cc",
3447 "test/pad-microkernel-tester.h",
3448 ] + MICROKERNEL_TEST_HDRS,
3449 deps = MICROKERNEL_TEST_DEPS,
3450)
3451
3452xnnpack_unit_test(
3453 name = "x32_unpool_test",
3454 srcs = [
3455 "test/x32-unpool.cc",
3456 "test/unpool-microkernel-tester.h",
3457 ] + MICROKERNEL_TEST_HDRS,
3458 deps = MICROKERNEL_TEST_DEPS,
3459)
3460
3461xnnpack_unit_test(
3462 name = "x32_zip_test",
3463 srcs = [
3464 "test/x32-zip.cc",
3465 "test/zip-microkernel-tester.h",
3466 ] + MICROKERNEL_TEST_HDRS,
3467 deps = MICROKERNEL_TEST_DEPS,
3468)
3469
3470xnnpack_unit_test(
3471 name = "x8_lut_test",
3472 srcs = [
3473 "test/x8-lut.cc",
3474 "test/lut-microkernel-tester.h",
3475 ] + MICROKERNEL_TEST_HDRS,
3476 deps = MICROKERNEL_TEST_DEPS,
3477)
3478
3479xnnpack_unit_test(
3480 name = "x8_zip_test",
3481 srcs = [
3482 "test/x8-zip.cc",
3483 "test/zip-microkernel-tester.h",
3484 ] + MICROKERNEL_TEST_HDRS,
3485 deps = MICROKERNEL_TEST_DEPS,
3486)
3487
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003488xnnpack_unit_test(
3489 name = "requantization_test",
3490 srcs = [
3491 "src/xnnpack/requantization-stubs.h",
3492 "test/requantization.cc",
3493 "test/requantization-tester.h",
3494 ] + MICROKERNEL_TEST_HDRS,
3495 deps = MICROKERNEL_TEST_DEPS,
3496)
3497
Marat Dukhan20c3b922020-03-10 03:45:06 -07003498########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499
3500xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003501 name = "operator_size_test",
3502 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003503 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504)
3505
Marat Dukhan20c3b922020-03-10 03:45:06 -07003506xnnpack_binary(
3507 name = "subgraph_size_test",
3508 srcs = ["test/subgraph-size.c"],
3509 deps = [":XNNPACK"],
3510)
3511
3512########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513
3514xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003515 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003517 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003518 "test/add-operator-tester.h",
3519 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003520 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521)
3522
3523xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003524 name = "add_nd_test",
3525 srcs = [
3526 "test/add-nd.cc",
3527 "test/binary-elementwise-operator-tester.h",
3528 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003529 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003530)
3531
3532xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003533 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003534 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003535 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536 "test/argmax-pooling-operator-tester.h",
3537 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003538 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539)
3540
3541xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003542 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003543 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003544 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003545 "test/average-pooling-operator-tester.h",
3546 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003547 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003548)
3549
3550xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003551 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003552 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003553 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 "test/channel-pad-operator-tester.h",
3555 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003556 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557)
3558
3559xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003560 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003562 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003563 "test/channel-shuffle-operator-tester.h",
3564 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003565 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566)
3567
3568xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003569 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003571 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003572 "test/clamp-operator-tester.h",
3573 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003574 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575)
3576
3577xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003578 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003580 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581 "test/convolution-operator-tester.h",
3582 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003583 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584)
3585
3586xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003587 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003589 "test/convolution-nchw.cc",
3590 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003591 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003592 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593)
3594
3595xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003596 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003597 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003598 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003599 "test/deconvolution-operator-tester.h",
3600 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003601 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602)
3603
3604xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003605 name = "divide_nd_test",
3606 srcs = [
3607 "test/binary-elementwise-operator-tester.h",
3608 "test/divide-nd.cc",
3609 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003610 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003611)
3612
3613xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003614 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003616 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617 "test/fully-connected-operator-tester.h",
3618 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003619 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620)
3621
3622xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003623 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003625 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003626 "test/global-average-pooling-operator-tester.h",
3627 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003628 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629)
3630
3631xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003632 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003634 "test/global-average-pooling-ncw.cc",
3635 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003637 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003638)
3639
3640xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003641 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003643 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 "test/hardswish-operator-tester.h",
3645 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003646 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647)
3648
3649xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003650 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003651 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003652 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 "test/leaky-relu-operator-tester.h",
3654 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003655 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656)
3657
3658xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003659 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003661 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 "test/max-pooling-operator-tester.h",
3663 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003664 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003665)
3666
3667xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003668 name = "maximum_nd_test",
3669 srcs = [
3670 "test/binary-elementwise-operator-tester.h",
3671 "test/maximum-nd.cc",
3672 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003673 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003674)
3675
3676xnnpack_unit_test(
3677 name = "minimum_nd_test",
3678 srcs = [
3679 "test/binary-elementwise-operator-tester.h",
3680 "test/minimum-nd.cc",
3681 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003682 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003683)
3684
3685xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003686 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003687 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003688 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003689 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003690 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003691 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003692)
3693
3694xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003695 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003697 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 "test/prelu-operator-tester.h",
3699 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003700 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701)
3702
3703xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003704 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003705 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003706 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003707 "test/resize-bilinear-operator-tester.h",
3708 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003709 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003710)
3711
3712xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003713 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003714 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003715 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716 "test/sigmoid-operator-tester.h",
3717 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003718 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719)
3720
3721xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003722 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003723 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003724 "test/softmax-nc.cc",
3725 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003726 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003727 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728)
3729
3730xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003731 name = "subtract_nd_test",
3732 srcs = [
3733 "test/binary-elementwise-operator-tester.h",
3734 "test/subtract-nd.cc",
3735 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003736 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003737)
3738
3739xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003740 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003741 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003742 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003743 "test/unpooling-operator-tester.h",
3744 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003745 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746)
3747
Chao Mei6ddfc602020-05-13 22:29:36 -07003748############################### Misc unit tests ###############################
3749
3750xnnpack_unit_test(
3751 name = "memory_planner_test",
3752 srcs = [
3753 "test/memory-planner-test.cc",
3754 ],
3755 deps = [
3756 ":XNNPACK",
3757 ":memory_planner",
3758 ],
3759)
3760
Marat Dukhan08c4a432019-10-03 09:29:21 -07003761############################# Build configurations #############################
3762
Marat Dukhanb8642352019-10-30 15:43:02 -07003763# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003765 name = "xnn_enable_assembly_explicit_true",
3766 define_values = {"xnn_enable_assembly": "true"},
3767)
3768
3769# Disables usage of assembly kernels.
3770config_setting(
3771 name = "xnn_enable_assembly_explicit_false",
3772 define_values = {"xnn_enable_assembly": "false"},
3773)
3774
Marat Dukhan05702cf2020-03-26 15:41:33 -07003775# Disables usage of HMP-aware optimizations.
3776config_setting(
3777 name = "xnn_enable_hmp_explicit_false",
3778 define_values = {"xnn_enable_hmp": "false"},
3779)
3780
Chao Mei6ddfc602020-05-13 22:29:36 -07003781# Enable usage of optimized memory allocation
3782config_setting(
3783 name = "xnn_enable_memopt_explicit_true",
3784 define_values = {"xnnpack_enable_memopt": "true"},
3785)
3786
3787# Disable usage of optimized memory allocation
3788config_setting(
3789 name = "xnn_enable_memopt_explicit_false",
3790 define_values = {"xnnpack_enable_memopt": "false"},
3791)
3792
Marat Dukhanb8642352019-10-30 15:43:02 -07003793# Builds with -c dbg
3794config_setting(
3795 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003796 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003797 "compilation_mode": "dbg",
3798 },
3799)
3800
3801# Builds with -c opt
3802config_setting(
3803 name = "optimized_build",
3804 values = {
3805 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003806 },
3807)
3808
3809config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003810 name = "linux_k8",
3811 values = {"cpu": "k8"},
3812)
3813
3814config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003815 name = "linux_arm",
3816 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003817)
3818
3819config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003820 name = "linux_armhf",
3821 values = {"cpu": "armhf"},
3822)
3823
3824config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003825 name = "linux_aarch64",
3826 values = {"cpu": "aarch64"},
3827)
3828
3829config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003830 name = "android",
3831 values = {"crosstool_top": "//external:android/crosstool"},
3832)
3833
3834config_setting(
3835 name = "android_armv7",
3836 values = {
3837 "crosstool_top": "//external:android/crosstool",
3838 "cpu": "armeabi-v7a",
3839 },
3840)
3841
3842config_setting(
3843 name = "android_arm64",
3844 values = {
3845 "crosstool_top": "//external:android/crosstool",
3846 "cpu": "arm64-v8a",
3847 },
3848)
3849
3850config_setting(
3851 name = "android_x86",
3852 values = {
3853 "crosstool_top": "//external:android/crosstool",
3854 "cpu": "x86",
3855 },
3856)
3857
3858config_setting(
3859 name = "android_x86_64",
3860 values = {
3861 "crosstool_top": "//external:android/crosstool",
3862 "cpu": "x86_64",
3863 },
3864)
3865
3866config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003867 name = "windows_x86_64",
3868 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003869)
3870
3871config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003872 name = "windows_x86_64_clang",
3873 values = {
3874 "compiler": "clang-cl",
3875 "cpu": "x64_windows",
3876 },
3877)
3878
3879config_setting(
3880 name = "windows_x86_64_mingw",
3881 values = {
3882 "compiler": "mingw-gcc",
3883 "cpu": "x64_windows",
3884 },
3885)
3886
3887config_setting(
3888 name = "windows_x86_64_msys",
3889 values = {
3890 "compiler": "msys-gcc",
3891 "cpu": "x64_windows",
3892 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003893)
3894
3895config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003896 name = "macos_x86_64",
3897 values = {
3898 "apple_platform_type": "macos",
3899 "cpu": "darwin",
3900 },
3901)
3902
3903config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003905 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906)
3907
3908config_setting(
3909 name = "emscripten_wasm",
3910 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003911 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003912 "cpu": "wasm",
3913 },
3914)
3915
3916config_setting(
3917 name = "emscripten_wasmsimd",
3918 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003919 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003920 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003921 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003922 },
3923)
3924
3925config_setting(
3926 name = "emscripten_asmjs",
3927 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003928 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929 "cpu": "asmjs",
3930 },
3931)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003932
3933config_setting(
3934 name = "ios_armv7",
3935 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003936 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003937 "cpu": "ios_armv7",
3938 },
3939)
3940
3941config_setting(
3942 name = "ios_arm64",
3943 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003944 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003945 "cpu": "ios_arm64",
3946 },
3947)
3948
3949config_setting(
3950 name = "ios_arm64e",
3951 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003952 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003953 "cpu": "ios_arm64e",
3954 },
3955)
3956
3957config_setting(
3958 name = "ios_x86",
3959 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003960 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003961 "cpu": "ios_i386",
3962 },
3963)
3964
3965config_setting(
3966 name = "ios_x86_64",
3967 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003968 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003969 "cpu": "ios_x86_64",
3970 },
3971)
3972
3973config_setting(
3974 name = "watchos_armv7k",
3975 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003976 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003977 "cpu": "watchos_armv7k",
3978 },
3979)
3980
3981config_setting(
3982 name = "watchos_arm64_32",
3983 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003984 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003985 "cpu": "watchos_arm64_32",
3986 },
3987)
3988
3989config_setting(
3990 name = "watchos_x86",
3991 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003992 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003993 "cpu": "watchos_i386",
3994 },
3995)
3996
3997config_setting(
3998 name = "watchos_x86_64",
3999 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004000 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004001 "cpu": "watchos_x86_64",
4002 },
4003)
4004
4005config_setting(
4006 name = "tvos_arm64",
4007 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004008 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004009 "cpu": "tvos_arm64",
4010 },
4011)
4012
4013config_setting(
4014 name = "tvos_x86_64",
4015 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004016 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004017 "cpu": "tvos_x86_64",
4018 },
4019)