blob: 11984cc88532e9beb4c50ad547d451ff258d4c74 [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 Dukhanf5425ea2020-04-24 01:46:00 -0700524 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
525 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
526 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
527 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700528 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
529 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
530 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
531 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700532 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
533 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
534 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
535 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700536 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
538 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700539 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
540 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
541 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
542 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
543 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
544 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
545 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
546 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
547 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
548 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
549 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
550 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
551 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
552 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
553 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
554 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
555 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
556 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
557 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
558 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
559 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
560 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
561 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
562 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
563 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
564 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
565 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
566 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
567 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
568 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
569 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800570 "src/f32-hswish/gen/neon-x4.c",
571 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700572 "src/f32-ibilinear/gen/neon-c4.c",
573 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700574 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
575 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
576 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
577 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
578 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
579 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
580 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
581 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
582 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
583 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
584 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
585 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
586 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
587 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
588 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
589 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700590 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
591 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
592 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700593 "src/f32-ppmm/gen/4x8-minmax-neon.c",
594 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800595 "src/f32-prelu/gen/neon-2x4.c",
596 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800597 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
598 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
599 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
600 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
601 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
602 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
603 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
604 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
605 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
606 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
607 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
608 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
609 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
610 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
611 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
612 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
613 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
614 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
615 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
616 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
617 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
618 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
619 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
620 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700621 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800622 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800623 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
624 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
625 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
626 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
627 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
628 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
629 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
630 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
631 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
632 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
633 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
634 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
635 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
636 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
637 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
638 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
639 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
640 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700641 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
642 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
643 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
644 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800645 "src/f32-vbinary/gen/vmax-neon-x4.c",
646 "src/f32-vbinary/gen/vmax-neon-x8.c",
647 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
648 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
649 "src/f32-vbinary/gen/vmin-neon-x4.c",
650 "src/f32-vbinary/gen/vmin-neon-x8.c",
651 "src/f32-vbinary/gen/vminc-neon-x4.c",
652 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700653 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
654 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
655 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
656 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
657 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
658 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
659 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
660 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
661 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
662 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700663 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
664 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
665 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
666 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700667 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700668 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
669 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700670 "src/q8-gemm/4x8-minmax-neon.c",
671 "src/q8-gemm/8x8-minmax-neon.c",
672 "src/q8-igemm/4x8-minmax-neon.c",
673 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700674 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700675 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700676 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700677 "src/u8-rmax/neon.c",
678 "src/x32-packx/x4-neon-st4.c",
679 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700680 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700681 "src/x32-zip/x2-neon.c",
682 "src/x32-zip/x3-neon.c",
683 "src/x32-zip/x4-neon.c",
684 "src/x32-zip/xm-neon.c",
685 "src/x8-zip/x2-neon.c",
686 "src/x8-zip/x3-neon.c",
687 "src/x8-zip/x4-neon.c",
688 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700689 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700690 "src/math/roundd-neon-addsub.c",
691 "src/math/roundd-neon-cvt.c",
692 "src/math/roundu-neon-addsub.c",
693 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700694 "src/math/roundz-neon-addsub.c",
695 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800696 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800697 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
698 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
699 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
700 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
701 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
702 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700703 "src/requantization/precise-neon.c",
704 "src/requantization/fp32-neon.c",
705 "src/requantization/q31-neon.c",
706 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700707]
708
709NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700710 "src/f32-ibilinear/gen/neonfma-c4.c",
711 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700712 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
713 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
714 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
715 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
716 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
717 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
718 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
719 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
720 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700721 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
722 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
723 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
724 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700725 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
726 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
727 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
728 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700729 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
730 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
731 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
732 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700733 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
734 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
735 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
736 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
737 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
738 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
739 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
740 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
741 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
742 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
743 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
744 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
746 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
747 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
748 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
749 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
750 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800751 "src/f32-hswish/gen/neonfma-x4.c",
752 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700753 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
754 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800755 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
756 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
757 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
758 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
759 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
760 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
761 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
762 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
763 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
764 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
765 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
766 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
767 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
768 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
769 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
770 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
771 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
772 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
773 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
774 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
775 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
776 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
777 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
778 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800779 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
780 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
781 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
782 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
783 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
832 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700833 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
834 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800835 "src/math/exp-neonfma-lut64-p2.c",
836 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800837 "src/math/expminus-neonfma-lut2048-p1.c",
838 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800839 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800840 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
841 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
842 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
843 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
844 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
845 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
846 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
847 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
848 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
849 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
850 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
851 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
852 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
853 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
854 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
855 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
856 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
857 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700858]
859
860AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700861 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
862 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
863 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
864 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
865 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
866 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700867 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
868 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
869 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
870 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
871 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
872 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
873 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
874 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
875 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
876 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
877 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
878 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
879 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
880 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
881 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
882 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
883 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
884 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
885 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
886 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan7e4ca402020-05-15 18:50:12 -0700887 "src/f32-conv-hwc/3x3s2p0p1c3x4-neonfma-2x2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700888 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
889 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700890 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
891 "src/f32-dwconv-chw/3x3p1-neonfma.c",
892 "src/f32-dwconv-chw/5x5p2-neonfma.c",
893 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
894 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800895 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
896 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
897 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
898 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
899 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
900 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
901 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
902 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
903 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
904 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
905 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
906 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
907 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
908 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
909 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
910 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
911 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
912 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700913 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
914 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
915 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
916 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
917 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
918 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
919 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
920 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
921 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
922 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
923 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
924 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
925 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
926 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
927 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
928 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
929 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
930 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800931 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
932 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
933 "src/math/sigmoid-neonfma-rr1-p5-div.c",
934 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
935 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
936 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937]
938
Marat Dukhan8853b822020-05-07 12:19:01 -0700939NEONV8_UKERNELS = [
940 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700941 "src/math/roundd-neonv8.c",
942 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700943 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700944]
945
Marat Dukhan08c4a432019-10-03 09:29:21 -0700946AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -0700947 "src/f16-clamp/gen/neonfp16arith-x16.c",
948 "src/f16-clamp/gen/neonfp16arith-x8.c",
949 "src/f16-hswish/gen/neonfp16arith-x16.c",
950 "src/f16-hswish/gen/neonfp16arith-x8.c",
951 "src/f16-prelu/gen/neonfp16arith-2x16.c",
952 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -0700953 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
954 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
955 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
956 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
957 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
958 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
959 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
960 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
961 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
962 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
963 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
964 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
965 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
966 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
967 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
968 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
969 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
970 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
971 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
972 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
973 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
974 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
975 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
976 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
977 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
978 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
979 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
980 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -0700981 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
982 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
983 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
984 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
985 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
986 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
987 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
988 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700989 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
990 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
991 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
992 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -0700993 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
994 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
995 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
996 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
997 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
998 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
999 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1000 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1001 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1002 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1003 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1004 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001005 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1006 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1007 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1008 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1009 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1010 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1011 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1012 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001013]
1014
1015SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001016 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1017 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001018 "src/f32-clamp/gen/sse-x4.c",
1019 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001020 "src/f32-dwconv-chw/3x3p1-sse.c",
1021 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001022 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1023 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1024 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1025 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1026 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1027 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1028 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1029 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1030 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1031 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1032 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1033 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001034 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001035 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1036 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001037 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1038 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1039 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1040 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1041 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1042 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1043 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1044 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1045 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1046 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1047 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1048 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1049 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001050 "src/f32-hswish/gen/sse-x4.c",
1051 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001052 "src/f32-ibilinear/gen/sse-c4.c",
1053 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001054 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1055 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1056 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1057 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1058 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1059 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1060 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001061 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1062 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1063 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001064 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001065 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001066 "src/f32-spmm/gen/4x1-minmax-sse.c",
1067 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001068 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1069 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1070 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1071 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1072 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1073 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1074 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1075 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001076 "src/f32-vbinary/gen/vmax-sse-x4.c",
1077 "src/f32-vbinary/gen/vmax-sse-x8.c",
1078 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1079 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1080 "src/f32-vbinary/gen/vmin-sse-x4.c",
1081 "src/f32-vbinary/gen/vmin-sse-x8.c",
1082 "src/f32-vbinary/gen/vminc-sse-x4.c",
1083 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001084 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1085 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1086 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1087 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1088 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1089 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1090 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1091 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1092 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1093 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1094 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1095 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001096 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1097 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001098 "src/x32-packx/x4-sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001099 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001100 "src/math/roundd-sse-addsub.c",
1101 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001102 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001103]
1104
1105SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001106 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1107 "src/f32-argmaxpool/4x-sse2-c4.c",
1108 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001109 "src/f32-prelu/gen/sse2-2x4.c",
1110 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001111 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1112 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1113 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1114 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1115 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1116 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1117 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1118 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1119 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1120 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1121 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1122 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001123 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001124 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001125 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001126 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001127 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1128 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001129 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1130 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001131 "src/q8-igemm/4x4c2-minmax-sse2.c",
1132 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001133 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1134 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001135 "src/q8-gemm/2x4c8-minmax-sse2.c",
1136 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001137 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001138 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001139 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001140 "src/u8-rmax/sse2.c",
1141 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001142 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143 "src/x32-zip/x2-sse2.c",
1144 "src/x32-zip/x3-sse2.c",
1145 "src/x32-zip/x4-sse2.c",
1146 "src/x32-zip/xm-sse2.c",
1147 "src/x8-zip/x2-sse2.c",
1148 "src/x8-zip/x3-sse2.c",
1149 "src/x8-zip/x4-sse2.c",
1150 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001151 "src/math/exp-sse2-p5.c",
1152 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001153 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001154 "src/math/roundd-sse2-cvt.c",
1155 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001156 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001157 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001158 "src/requantization/precise-sse2.c",
1159 "src/requantization/fp32-sse2.c",
1160 "src/requantization/q31-sse2.c",
1161 "src/requantization/gemmlowp-sse2.c",
1162]
1163
1164SSSE3_UKERNELS = [
1165 "src/requantization/precise-ssse3.c",
1166 "src/requantization/q31-ssse3.c",
1167 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001168]
1169
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001170SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001171 "src/f32-prelu/gen/sse41-2x4.c",
1172 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001173 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001174 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001175 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001176 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001177 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1178 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001179 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001180 "src/math/roundd-sse41.c",
1181 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001182 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001183 "src/requantization/precise-sse4.c",
1184 "src/requantization/q31-sse4.c",
1185 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001186]
1187
Marat Dukhan08c4a432019-10-03 09:29:21 -07001188AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001189 "src/f32-clamp/gen/avx-x8.c",
1190 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001191 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1192 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1193 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1194 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1195 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1196 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1197 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1198 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1199 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1200 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1201 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1202 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1203 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1204 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1205 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1206 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1207 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1208 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1209 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1210 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1211 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1212 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1213 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1214 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1215 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1216 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1217 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1218 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1219 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1220 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001221 "src/f32-hswish/gen/avx-x8.c",
1222 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001223 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1224 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1225 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1226 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1227 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1228 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1229 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1230 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1231 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001232 "src/f32-prelu/gen/avx-2x8.c",
1233 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001234 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001235 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1236 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1237 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1238 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1239 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1240 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1241 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1242 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001243 "src/f32-vbinary/gen/vmax-avx-x8.c",
1244 "src/f32-vbinary/gen/vmax-avx-x16.c",
1245 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1246 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1247 "src/f32-vbinary/gen/vmin-avx-x8.c",
1248 "src/f32-vbinary/gen/vmin-avx-x16.c",
1249 "src/f32-vbinary/gen/vminc-avx-x8.c",
1250 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001251 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1252 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1253 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1254 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1255 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1256 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1257 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1258 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1259 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1260 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1261 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1262 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001263 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001264]
1265
Marat Dukhanfda12b82019-11-21 12:27:59 -08001266FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001267 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1268 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1269 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1270 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1271 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1272 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1273 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1274 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1275 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1276 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1277 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1278 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1279 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1280 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1281 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1282 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1283 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1284 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1285 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1286 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1287 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1288 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1289 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1290 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1291 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1292 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1293 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1294 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1295 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1296 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1297 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1298 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1299 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1300 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1301 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1302 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1303 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1304 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1305 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1306 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001307 "src/f32-hswish/gen/fma3-x8.c",
1308 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001309 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1310 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1311 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1312 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1313 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1314 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1315 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1316 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1317 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1318 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1319 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1320 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1321 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1322 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001323]
1324
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001325AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001326 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1327 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1328 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1329 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1330 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1331 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1332 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1333 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1334 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1335 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1336 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1337 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1338 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1339 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1340 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1341 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1342 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1343 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1344 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1345 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1346 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1347 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1348 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1349 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1350 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1351 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1352 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1353 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1354 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1355 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1356 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1357 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1358 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1359 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1360 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1361 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001362 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1363 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1364 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1365 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1366 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1367 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1368 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1369 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1370 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1371 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1372 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1373 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1374 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1375 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1376 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1377 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1378 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1379 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1380 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1381 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1382 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1383 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1384 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1387 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1388 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1389 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1390 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001392 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1393 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1394 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1395 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1396 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1397 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1398 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1399 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1400 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1401 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1402 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1403 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1404 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1405 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1406 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1407 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1408 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1409 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1410 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1411 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1412 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1413 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1414 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1415 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001416 "src/math/exp-avx2-p5.c",
1417 "src/math/exp-avx2-perm-p3.c",
1418 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001419 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001420 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001421 "src/math/sigmoid-avx2-rr2-p5-div.c",
1422 "src/math/sigmoid-avx2-rr1-p5-div.c",
1423 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1424 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1425 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1426 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001427]
1428
Marat Dukhan08c4a432019-10-03 09:29:21 -07001429AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001430 "src/f32-clamp/gen/avx512f-x16.c",
1431 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001432 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1433 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1434 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1435 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1436 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1437 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1438 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1439 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1440 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1441 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1442 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1443 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1444 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1445 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1446 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1447 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1448 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1449 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1450 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1451 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1452 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1453 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1454 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1455 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001456 "src/f32-hswish/gen/avx512f-x16.c",
1457 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001458 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1459 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1460 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1461 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1462 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1463 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001464 "src/f32-prelu/gen/avx512f-2x16.c",
1465 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001466 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1467 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1468 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1469 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1470 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1471 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1472 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1473 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1474 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1475 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1476 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1477 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1478 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1479 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1480 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1481 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1482 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1483 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1484 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1485 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1486 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1487 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1488 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1489 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1490 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1491 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1492 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1493 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1494 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1495 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1496 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1497 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1498 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1499 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1500 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1501 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001502 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001503 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1504 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1505 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1506 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1507 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1508 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1509 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1510 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001511 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1512 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1513 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1514 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1515 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1516 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1517 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1518 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001519 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1520 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1521 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1522 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1523 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1524 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1525 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1526 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1527 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1528 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1529 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1530 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001531 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001532 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1533 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1534 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1535 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1536 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1537 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1538 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1539 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1540 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1541 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1542 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1543 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1544 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1545 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1546 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1547 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1548 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1549 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1550 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1551 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1552 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1553 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1554 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1555 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001556 "src/math/exp-avx512f-p5-scalef.c",
1557 "src/math/exp-avx512f-p5.c",
1558 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001559 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001560 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001561]
1562
1563AARCH32_ASM_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001564 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1565 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1566 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1567 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1568 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1569 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1570 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1571 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1572 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1573 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001574]
1575
1576AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001577 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001578 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1579 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001580 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001581 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1582 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001583 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1584 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1585 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001586 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001587 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1588 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1589 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001590 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001591 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1592 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001593 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001594 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1595 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1596 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1597 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1598 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1599 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1600 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1601 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1602 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1603 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1604 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1605 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1606 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1607 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1608 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1609 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1610 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1611 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1612 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1613 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1614 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001615 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001616 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1617 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1618 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1619 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1620 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1621 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1622 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1623 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1624 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1625 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1626 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1627 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1628 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1629 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1630 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1631 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1632 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1633 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1634 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1635 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1636 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1637 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1638 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1639 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1640 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1641 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1642 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1643 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1644 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1645 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1646 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1647 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1648 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1649 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1650 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1651 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1652 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1653 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001654]
1655
Marat Dukhan1b354632020-03-23 12:50:22 -07001656INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001657 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658 "src/xnnpack/argmaxpool.h",
1659 "src/xnnpack/avgpool.h",
1660 "src/xnnpack/clamp.h",
1661 "src/xnnpack/common.h",
1662 "src/xnnpack/conv.h",
1663 "src/xnnpack/dwconv.h",
1664 "src/xnnpack/gavgpool.h",
1665 "src/xnnpack/gemm.h",
1666 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001667 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001668 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001669 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001670 "src/xnnpack/lut.h",
1671 "src/xnnpack/math.h",
1672 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001673 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001674 "src/xnnpack/packx.h",
1675 "src/xnnpack/pad.h",
1676 "src/xnnpack/params.h",
1677 "src/xnnpack/pavgpool.h",
1678 "src/xnnpack/ppmm.h",
1679 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001680 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001681 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001682 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 "src/xnnpack/rmax.h",
1684 "src/xnnpack/scalar-utils.h",
1685 "src/xnnpack/spmm.h",
1686 "src/xnnpack/unpool.h",
1687 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001688 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001689 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001690 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001691 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001692 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001693 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001694 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001695]
1696
1697INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001698 "include/xnnpack.h",
1699 "src/xnnpack/allocator.h",
1700 "src/xnnpack/compute.h",
1701 "src/xnnpack/im2col.h",
1702 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001703 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001704 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001705 "src/xnnpack/operator.h",
1706 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001707 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001708 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001709 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001710 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001711]
1712
Marat Dukhan1b354632020-03-23 12:50:22 -07001713ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001714 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001715]
1716
Marat Dukhan1b354632020-03-23 12:50:22 -07001717MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001718 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001719 "include/xnnpack.h",
1720]
1721
Marat Dukhan1b354632020-03-23 12:50:22 -07001722MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001724 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001725 "src/xnnpack/requantization.h",
1726 "include/xnnpack.h",
1727]
1728
1729OPERATOR_TEST_PARAMS_HDRS = [
1730 "src/xnnpack/params.h",
1731 "src/xnnpack/common.h",
1732]
1733
1734WEIGHTS_PACK_HDRS = [
1735 "src/xnnpack/pack.h",
1736 "src/xnnpack/operator.h",
1737 "src/xnnpack/compute.h",
1738]
1739
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001740LOGGING_COPTS = select({
1741 # No logging in optimized mode
1742 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1743 # Full logging in debug mode
1744 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1745 # Error-only logging in default (fastbuild) mode
1746 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1747})
1748
1749LOGGING_HDRS = [
1750 "src/xnnpack/log.h",
1751]
1752
Marat Dukhan08c4a432019-10-03 09:29:21 -07001753xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001754 name = "tables",
1755 srcs = TABLE_SRCS,
1756 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001757 gcc_copts = xnnpack_gcc_std_copts(),
1758 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001759)
1760
1761xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762 name = "scalar_ukernels",
1763 srcs = SCALAR_UKERNELS,
1764 hdrs = INTERNAL_HDRS,
1765 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001766 gcc_copts = xnnpack_gcc_std_copts(),
1767 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001769 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001770 "@FP16",
1771 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001772 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773 ],
1774)
1775
1776xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001777 name = "wasm_ukernels",
1778 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001779 gcc_copts = xnnpack_gcc_std_copts(),
1780 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001781 wasm_srcs = WASM_UKERNELS,
1782 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001783 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001784 "@FP16",
1785 "@FXdiv",
1786 "@pthreadpool",
1787 ],
1788)
1789
1790xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001791 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001792 hdrs = INTERNAL_HDRS,
1793 aarch32_copts = [
1794 "-marm",
1795 "-mfpu=neon",
1796 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001797 gcc_copts = xnnpack_gcc_std_copts(),
1798 gcc_x86_copts = ["-msse2"],
1799 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001800 optimized_copts = [
1801 "-O3",
1802 "-ffast-math",
1803 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001804 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001805 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001806 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807 "@FP16",
1808 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001809 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001810 ],
1811)
1812
1813xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001814 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001815 hdrs = INTERNAL_HDRS,
1816 aarch32_copts = [
1817 "-marm",
1818 "-mfpu=neon",
1819 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001820 gcc_copts = xnnpack_gcc_std_copts(),
1821 gcc_x86_copts = ["-msse2"],
1822 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001823 optimized_copts = [
1824 "-O3",
1825 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001826 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001827 deps = [
1828 ":tables",
1829 "@FP16",
1830 "@psimd",
1831 "@pthreadpool",
1832 ],
1833)
1834
1835xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001836 name = "neon_ukernels",
1837 hdrs = INTERNAL_HDRS,
1838 aarch32_copts = [
1839 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001840 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "-mfpu=neon",
1842 ],
1843 aarch32_srcs = NEON_UKERNELS,
1844 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001845 gcc_copts = xnnpack_gcc_std_copts(),
1846 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001847 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001848 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001849 "@FP16",
1850 "@pthreadpool",
1851 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001852)
1853
1854xnnpack_cc_library(
1855 name = "neonfma_ukernels",
1856 hdrs = INTERNAL_HDRS,
1857 aarch32_copts = [
1858 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001859 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001860 "-mfpu=neon-vfpv4",
1861 ],
1862 aarch32_srcs = NEONFMA_UKERNELS,
1863 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001864 gcc_copts = xnnpack_gcc_std_copts(),
1865 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001866 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001867 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001868 "@FP16",
1869 "@pthreadpool",
1870 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001871)
1872
1873xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001874 name = "neonv8_ukernels",
1875 hdrs = INTERNAL_HDRS,
1876 aarch32_copts = [
1877 "-marm",
1878 "-march=armv8-a",
1879 "-mfpu=neon-fp-armv8",
1880 ],
1881 aarch32_srcs = NEONV8_UKERNELS,
1882 aarch64_srcs = NEONV8_UKERNELS,
1883 gcc_copts = xnnpack_gcc_std_copts(),
1884 msvc_copts = xnnpack_msvc_std_copts(),
1885 deps = [
1886 ":tables",
1887 "@FP16",
1888 "@pthreadpool",
1889 ],
1890)
1891
1892xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001893 name = "neonfp16arith_ukernels",
1894 hdrs = INTERNAL_HDRS,
1895 aarch64_copts = ["-march=armv8.2-a+fp16"],
1896 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001897 gcc_copts = xnnpack_gcc_std_copts(),
1898 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001899 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001900 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001901 "@FP16",
1902 "@pthreadpool",
1903 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001904)
1905
1906xnnpack_cc_library(
1907 name = "sse2_ukernels",
1908 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001909 gcc_copts = xnnpack_gcc_std_copts(),
1910 gcc_x86_copts = ["-msse2"],
1911 msvc_copts = xnnpack_msvc_std_copts(),
1912 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001913 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
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(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001922 name = "ssse3_ukernels",
1923 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001924 gcc_copts = xnnpack_gcc_std_copts(),
1925 gcc_x86_copts = ["-mssse3"],
1926 msvc_copts = xnnpack_msvc_std_copts(),
1927 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001928 x86_srcs = SSSE3_UKERNELS,
1929 deps = [
1930 ":tables",
1931 "@FP16",
1932 "@pthreadpool",
1933 ],
1934)
1935
1936xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001937 name = "sse41_ukernels",
1938 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001939 gcc_copts = xnnpack_gcc_std_copts(),
1940 gcc_x86_copts = ["-msse4.1"],
1941 msvc_copts = xnnpack_msvc_std_copts(),
1942 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001943 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001944 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001945 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001946 "@FP16",
1947 "@pthreadpool",
1948 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001949)
1950
1951xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001952 name = "avx_ukernels",
1953 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001954 gcc_copts = xnnpack_gcc_std_copts(),
1955 gcc_x86_copts = ["-mavx"],
1956 msvc_copts = xnnpack_msvc_std_copts(),
1957 msvc_x86_32_copts = ["/arch:AVX"],
1958 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001959 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001960 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001961 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001962 "@FP16",
1963 "@pthreadpool",
1964 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001965)
1966
1967xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001968 name = "fma3_ukernels",
1969 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001970 gcc_copts = xnnpack_gcc_std_copts(),
1971 gcc_x86_copts = ["-mfma"],
1972 msvc_copts = xnnpack_msvc_std_copts(),
1973 msvc_x86_32_copts = ["/arch:AVX"],
1974 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001975 x86_srcs = FMA3_UKERNELS,
1976 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001977 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001978 "@FP16",
1979 "@pthreadpool",
1980 ],
1981)
1982
1983xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001984 name = "avx2_ukernels",
1985 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001986 gcc_copts = xnnpack_gcc_std_copts(),
1987 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001988 "-mfma",
1989 "-mavx2",
1990 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001991 msvc_copts = xnnpack_msvc_std_copts(),
1992 msvc_x86_32_copts = ["/arch:AVX2"],
1993 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001994 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001995 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001996 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001997 "@FP16",
1998 "@pthreadpool",
1999 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002000)
2001
2002xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003 name = "avx512f_ukernels",
2004 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002005 gcc_copts = xnnpack_gcc_std_copts(),
2006 gcc_x86_copts = ["-mavx512f"],
2007 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2008 msvc_copts = xnnpack_msvc_std_copts(),
2009 msvc_x86_32_copts = ["/arch:AVX512"],
2010 msvc_x86_64_copts = ["/arch:AVX512"],
2011 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002012 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002013 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002014 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002015 "@FP16",
2016 "@pthreadpool",
2017 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002018)
2019
2020xnnpack_cc_library(
2021 name = "asm_ukernels",
2022 hdrs = ["src/xnnpack/assembly.h"],
2023 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002024 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025 aarch64_srcs = AARCH64_ASM_UKERNELS,
2026)
2027
2028xnnpack_aggregate_library(
2029 name = "ukernels",
2030 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031 ":neon_ukernels",
2032 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002033 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002034 ":asm_ukernels",
2035 ],
2036 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002037 ":neon_ukernels",
2038 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002039 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002040 ":neonfp16arith_ukernels",
2041 ":asm_ukernels",
2042 ],
2043 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002044 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002045 ":psimd_fastmath_ukernels",
2046 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002047 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002048 wasm_deps = [":wasm_ukernels"],
2049 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002050 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002052 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002053 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002055 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002056 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057 ":avx512f_ukernels",
2058 ],
2059)
2060
2061xnnpack_cc_library(
2062 name = "im2col",
2063 srcs = ["src/im2col.c"],
2064 hdrs = [
2065 "src/xnnpack/common.h",
2066 "src/xnnpack/im2col.h",
2067 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002068 gcc_copts = xnnpack_gcc_std_copts(),
2069 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070)
2071
2072xnnpack_cc_library(
2073 name = "indirection",
2074 srcs = ["src/indirection.c"],
2075 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002076 gcc_copts = xnnpack_gcc_std_copts(),
2077 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002078 deps = [
2079 "@FP16",
2080 "@FXdiv",
2081 "@pthreadpool",
2082 ],
2083)
2084
2085xnnpack_cc_library(
2086 name = "operator_run",
2087 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002088 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002089 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002090 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2091 "//conditions:default": [],
2092 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002093 gcc_copts = xnnpack_gcc_std_copts(),
2094 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002095 deps = [
2096 "@FP16",
2097 "@FXdiv",
2098 "@clog",
2099 "@pthreadpool",
2100 ],
2101)
2102
Chao Mei6ddfc602020-05-13 22:29:36 -07002103xnnpack_cc_library(
2104 name = "memory_planner",
2105 srcs = ["src/memory-planner.c"],
2106 hdrs = INTERNAL_HDRS,
2107 defines = select({
2108 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2109 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2110 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2111 }),
2112 gcc_copts = xnnpack_gcc_std_copts(),
2113 msvc_copts = xnnpack_msvc_std_copts(),
2114 deps = [
2115 "@pthreadpool",
2116 ],
2117)
2118
Marat Dukhan08c4a432019-10-03 09:29:21 -07002119cc_library(
2120 name = "enable_assembly",
2121 defines = select({
2122 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2123 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002124 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002125 }),
2126)
2127
Marat Dukhancf056b22019-10-07 10:26:29 -07002128xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002129 name = "operators",
2130 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002131 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002132 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002133 ],
2134 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002135 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002136 "-Isrc",
2137 "-Iinclude",
2138 ] + select({
2139 ":debug_build": [],
2140 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002141 }) + select({
2142 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2143 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002144 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002145 gcc_copts = xnnpack_gcc_std_copts(),
2146 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002148 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002149 "@FP16",
2150 "@FXdiv",
2151 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002152 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002153 ],
2154)
2155
Marat Dukhan10a38082020-04-17 03:58:35 -07002156xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002157 name = "XNNPACK",
2158 srcs = [
2159 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002160 "src/runtime.c",
2161 "src/subgraph.c",
2162 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002163 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002164 hdrs = ["include/xnnpack.h"],
2165 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002166 "-Isrc",
2167 "-Iinclude",
2168 ] + select({
2169 ":debug_build": [],
2170 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002171 }) + select({
2172 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2173 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002174 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002175 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002176 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002177 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002178 visibility = xnnpack_visibility(),
2179 deps = [
2180 ":enable_assembly",
Chao Mei6ddfc602020-05-13 22:29:36 -07002181 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002182 ":ukernels",
2183 ":operator_run",
2184 ":operators",
2185 "@clog",
2186 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002187 ] + select({
2188 ":emscripten": [],
2189 "//conditions:default": ["@cpuinfo"],
2190 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002191)
2192
Marat Dukhan10a38082020-04-17 03:58:35 -07002193xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002194 name = "xnnpack_operators_nhwc_f32",
2195 srcs = [
2196 "src/init.c",
2197 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002198 hdrs = ["include/xnnpack.h"],
2199 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002200 "-Isrc",
2201 "-Iinclude",
2202 ] + select({
2203 ":debug_build": [],
2204 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002205 }) + select({
2206 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2207 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002208 }),
2209 defines = [
2210 "XNN_NO_Q8_OPERATORS",
2211 "XNN_NO_U8_OPERATORS",
2212 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002213 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002214 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002215 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002217 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002218 visibility = xnnpack_visibility(),
2219 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002220 ":enable_assembly",
2221 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002222 ":operator_run",
2223 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002224 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002225 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002226 ] + select({
2227 ":emscripten": [],
2228 "//conditions:default": ["@cpuinfo"],
2229 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002230)
2231
Marat Dukhancf056b22019-10-07 10:26:29 -07002232xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002233 name = "bench_utils",
2234 srcs = ["bench/utils.cc"],
2235 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002236 deps = [
2237 "@com_google_benchmark//:benchmark",
2238 "@cpuinfo",
2239 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002240)
2241
Frank Barchard7e955972019-10-11 10:34:25 -07002242######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243
2244xnnpack_benchmark(
2245 name = "q8_gemm_bench",
2246 srcs = [
2247 "bench/gemm.h",
2248 "bench/q8-gemm.cc",
2249 "src/xnnpack/AlignedAllocator.h",
2250 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002251 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2253)
2254
2255xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002256 name = "f16_igemm_bench",
2257 srcs = [
2258 "bench/f16-igemm.cc",
2259 "bench/conv.h",
2260 "bench/google/conv.h",
2261 "src/xnnpack/AlignedAllocator.h",
2262 ] + MICROKERNEL_BENCHMARK_HDRS,
2263 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2264)
2265
2266xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002267 name = "f16_gemm_bench",
2268 srcs = [
2269 "bench/f16-gemm.cc",
2270 "bench/gemm.h",
2271 "src/xnnpack/AlignedAllocator.h",
2272 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002273 deps = MICROKERNEL_BENCHMARK_DEPS,
2274)
2275
2276xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002277 name = "f16_spmm_bench",
2278 srcs = [
2279 "bench/f16-spmm.cc",
2280 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002281 "src/xnnpack/AlignedAllocator.h",
2282 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002283 deps = MICROKERNEL_BENCHMARK_DEPS,
2284)
2285
2286xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002287 name = "f32_igemm_bench",
2288 srcs = [
2289 "bench/f32-igemm.cc",
2290 "bench/conv.h",
2291 "src/xnnpack/AlignedAllocator.h",
2292 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002293 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002294)
2295
2296xnnpack_benchmark(
2297 name = "f32_conv_hwc_bench",
2298 srcs = [
2299 "bench/f32-conv-hwc.cc",
2300 "bench/dconv.h",
2301 "src/xnnpack/AlignedAllocator.h",
2302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002303 deps = MICROKERNEL_BENCHMARK_DEPS,
2304)
2305
2306xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002307 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07002308 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002309 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07002310 "bench/dconv.h",
2311 "src/xnnpack/AlignedAllocator.h",
2312 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002313 deps = MICROKERNEL_BENCHMARK_DEPS,
2314)
2315
2316xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317 name = "f32_dwconv_bench",
2318 srcs = [
2319 "bench/f32-dwconv.cc",
2320 "bench/dwconv.h",
2321 "src/xnnpack/AlignedAllocator.h",
2322 ] + MICROKERNEL_BENCHMARK_HDRS,
2323 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2324)
2325
2326xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002327 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002329 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002330 "bench/dwconv.h",
2331 "src/xnnpack/AlignedAllocator.h",
2332 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2333 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2334)
2335
2336xnnpack_benchmark(
2337 name = "f32_gemm_bench",
2338 srcs = [
2339 "bench/f32-gemm.cc",
2340 "bench/gemm.h",
2341 "src/xnnpack/AlignedAllocator.h",
2342 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002343 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002344 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345)
2346
2347xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002348 name = "f32_raddexpminusmax_bench",
2349 srcs = [
2350 "bench/f32-raddexpminusmax.cc",
2351 "src/xnnpack/AlignedAllocator.h",
2352 ] + MICROKERNEL_BENCHMARK_HDRS,
2353 deps = MICROKERNEL_BENCHMARK_DEPS,
2354)
2355
2356xnnpack_benchmark(
2357 name = "f32_raddextexp_bench",
2358 srcs = [
2359 "bench/f32-raddextexp.cc",
2360 "src/xnnpack/AlignedAllocator.h",
2361 ] + MICROKERNEL_BENCHMARK_HDRS,
2362 deps = MICROKERNEL_BENCHMARK_DEPS,
2363)
2364
2365xnnpack_benchmark(
2366 name = "f32_raddstoreexpminusmax_bench",
2367 srcs = [
2368 "bench/f32-raddstoreexpminusmax.cc",
2369 "src/xnnpack/AlignedAllocator.h",
2370 ] + MICROKERNEL_BENCHMARK_HDRS,
2371 deps = MICROKERNEL_BENCHMARK_DEPS,
2372)
2373
2374xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002375 name = "f32_rmax_bench",
2376 srcs = [
2377 "bench/f32-rmax.cc",
2378 "src/xnnpack/AlignedAllocator.h",
2379 ] + MICROKERNEL_BENCHMARK_HDRS,
2380 deps = MICROKERNEL_BENCHMARK_DEPS,
2381)
2382
2383xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002384 name = "f32_sigmoid_bench",
2385 srcs = [
2386 "bench/f32-sigmoid.cc",
2387 "src/xnnpack/AlignedAllocator.h",
2388 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002389 deps = MICROKERNEL_BENCHMARK_DEPS,
2390)
2391
2392xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002393 name = "f32_spmm_bench",
2394 srcs = [
2395 "bench/f32-spmm.cc",
2396 "bench/gemm.h",
2397 "src/xnnpack/AlignedAllocator.h",
2398 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002399 deps = MICROKERNEL_BENCHMARK_DEPS,
2400)
2401
2402xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002403 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002404 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002405 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002406 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002407 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002408 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002409)
2410
2411xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002412 name = "f32_vscaleexpminusmax_bench",
2413 srcs = [
2414 "bench/f32-vscaleexpminusmax.cc",
2415 "src/xnnpack/AlignedAllocator.h",
2416 ] + MICROKERNEL_BENCHMARK_HDRS,
2417 deps = MICROKERNEL_BENCHMARK_DEPS,
2418)
2419
2420xnnpack_benchmark(
2421 name = "f32_vscaleextexp_bench",
2422 srcs = [
2423 "bench/f32-vscaleextexp.cc",
2424 "src/xnnpack/AlignedAllocator.h",
2425 ] + MICROKERNEL_BENCHMARK_HDRS,
2426 deps = MICROKERNEL_BENCHMARK_DEPS,
2427)
2428
2429xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002430 name = "f32_im2col_gemm_bench",
2431 srcs = [
2432 "bench/f32-im2col-gemm.cc",
2433 "bench/conv.h",
2434 "src/xnnpack/AlignedAllocator.h",
2435 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2436 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2437)
2438
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002439xnnpack_benchmark(
2440 name = "requantization_bench",
2441 srcs = [
2442 "bench/requantization.cc",
2443 "src/xnnpack/requantization-stubs.h",
2444 "src/xnnpack/AlignedAllocator.h",
2445 ] + MICROKERNEL_BENCHMARK_HDRS,
2446 deps = MICROKERNEL_BENCHMARK_DEPS,
2447)
2448
Marat Dukhanffbf96a2020-05-14 02:59:08 -07002449xnnpack_benchmark(
2450 name = "rounding_bench",
2451 srcs = [
2452 "bench/rounding.cc",
2453 "src/xnnpack/math-stubs.h",
2454 "src/xnnpack/AlignedAllocator.h",
2455 ] + MICROKERNEL_BENCHMARK_HDRS,
2456 deps = MICROKERNEL_BENCHMARK_DEPS,
2457)
2458
Marat Dukhan08c4a432019-10-03 09:29:21 -07002459########################### Benchmarks for operators ###########################
2460
2461xnnpack_benchmark(
2462 name = "add_bench",
2463 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002464 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465)
2466
2467xnnpack_benchmark(
2468 name = "average_pooling_bench",
2469 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002470 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002471 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002472 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002473)
2474
2475xnnpack_benchmark(
2476 name = "channel_shuffle_bench",
2477 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002478 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002479)
2480
2481xnnpack_benchmark(
2482 name = "convolution_bench",
2483 srcs = ["bench/convolution.cc"],
2484 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002485 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002486 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002487)
2488
2489xnnpack_benchmark(
2490 name = "deconvolution_bench",
2491 srcs = ["bench/deconvolution.cc"],
2492 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002493 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002494 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002495)
2496
2497xnnpack_benchmark(
2498 name = "global_average_pooling_bench",
2499 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002500 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002501)
2502
2503xnnpack_benchmark(
2504 name = "max_pooling_bench",
2505 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002506 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507)
2508
2509xnnpack_benchmark(
2510 name = "sigmoid_bench",
2511 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002512 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002513 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002514 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002515)
2516
2517xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002518 name = "prelu_bench",
2519 srcs = ["bench/prelu.cc"],
2520 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002521 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002522 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002523)
2524
2525xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002526 name = "softmax_bench",
2527 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002528 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002529 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002530 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002531)
2532
Marat Dukhanc068bb62019-10-04 13:24:39 -07002533############################# End-to-end benchmarks ############################
2534
2535cc_library(
2536 name = "mobilenet_v1",
2537 srcs = ["models/mobilenet-v1.cc"],
2538 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002539 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002540 linkstatic = True,
2541 deps = [
2542 ":XNNPACK",
2543 "@pthreadpool",
2544 ],
2545)
2546
2547cc_library(
2548 name = "mobilenet_v2",
2549 srcs = ["models/mobilenet-v2.cc"],
2550 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002551 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002552 linkstatic = True,
2553 deps = [
2554 ":XNNPACK",
2555 "@pthreadpool",
2556 ],
2557)
2558
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002559cc_library(
2560 name = "mobilenet_v3_large",
2561 srcs = ["models/mobilenet-v3-large.cc"],
2562 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002563 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002564 linkstatic = True,
2565 deps = [
2566 ":XNNPACK",
2567 "@pthreadpool",
2568 ],
2569)
2570
2571cc_library(
2572 name = "mobilenet_v3_small",
2573 srcs = ["models/mobilenet-v3-small.cc"],
2574 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002575 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002576 linkstatic = True,
2577 deps = [
2578 ":XNNPACK",
2579 "@pthreadpool",
2580 ],
2581)
2582
Marat Dukhanc068bb62019-10-04 13:24:39 -07002583xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002584 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002585 srcs = [
2586 "bench/f32-dwconv-e2e.cc",
2587 "bench/end2end.h",
2588 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002589 deps = MICROKERNEL_BENCHMARK_DEPS + [
2590 ":XNNPACK",
2591 ":mobilenet_v1",
2592 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002593 ":mobilenet_v3_large",
2594 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002595 ],
2596)
2597
2598xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002599 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002600 srcs = [
2601 "bench/f32-gemm-e2e.cc",
2602 "bench/end2end.h",
2603 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -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 Dukhan5f18d262019-10-31 10:24:14 -07002610 ],
2611)
2612
2613xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002614 name = "end2end_bench",
2615 srcs = ["bench/end2end.cc"],
2616 deps = [
2617 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002618 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002619 ":mobilenet_v1",
2620 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002621 ":mobilenet_v3_large",
2622 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002623 "@pthreadpool",
2624 ],
2625)
2626
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002627#################### Accuracy evaluation for math functions ####################
2628
2629xnnpack_benchmark(
2630 name = "f32_exp_eval",
2631 srcs = [
2632 "eval/f32-exp.cc",
2633 "src/xnnpack/AlignedAllocator.h",
2634 ] + ACCURACY_EVAL_HDRS,
2635 deps = ACCURACY_EVAL_DEPS,
2636)
2637
Marat Dukhan515c9772019-10-17 18:07:57 -07002638xnnpack_benchmark(
2639 name = "f32_expminus_eval",
2640 srcs = [
2641 "eval/f32-expminus.cc",
2642 "src/xnnpack/AlignedAllocator.h",
2643 ] + ACCURACY_EVAL_HDRS,
2644 deps = ACCURACY_EVAL_DEPS,
2645)
2646
Marat Dukhan98ba4412019-10-23 02:14:28 -07002647xnnpack_benchmark(
2648 name = "f32_extexp_eval",
2649 srcs = [
2650 "eval/f32-extexp.cc",
2651 "src/xnnpack/AlignedAllocator.h",
2652 ] + ACCURACY_EVAL_HDRS,
2653 deps = ACCURACY_EVAL_DEPS,
2654)
2655
Marat Dukhan8853b822020-05-07 12:19:01 -07002656xnnpack_unit_test(
2657 name = "f32_roundne_eval",
2658 srcs = [
2659 "eval/f32-roundne.cc",
2660 "src/xnnpack/AlignedAllocator.h",
2661 "src/xnnpack/math-stubs.h",
2662 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002663 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002664 deps = MICROKERNEL_TEST_DEPS,
2665)
2666
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002667xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002668 name = "f32_roundd_eval",
2669 srcs = [
2670 "eval/f32-roundd.cc",
2671 "src/xnnpack/AlignedAllocator.h",
2672 "src/xnnpack/math-stubs.h",
2673 ] + MICROKERNEL_TEST_HDRS,
2674 automatic = False,
2675 deps = MICROKERNEL_TEST_DEPS,
2676)
2677
2678xnnpack_unit_test(
2679 name = "f32_roundu_eval",
2680 srcs = [
2681 "eval/f32-roundu.cc",
2682 "src/xnnpack/AlignedAllocator.h",
2683 "src/xnnpack/math-stubs.h",
2684 ] + MICROKERNEL_TEST_HDRS,
2685 automatic = False,
2686 deps = MICROKERNEL_TEST_DEPS,
2687)
2688
2689xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002690 name = "f32_roundz_eval",
2691 srcs = [
2692 "eval/f32-roundz.cc",
2693 "src/xnnpack/AlignedAllocator.h",
2694 "src/xnnpack/math-stubs.h",
2695 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002696 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002697 deps = MICROKERNEL_TEST_DEPS,
2698)
2699
Marat Dukhan346a9e52019-11-15 09:06:30 -08002700xnnpack_benchmark(
2701 name = "f32_sigmoid_eval",
2702 srcs = [
2703 "eval/f32-sigmoid.cc",
2704 "src/xnnpack/AlignedAllocator.h",
2705 ] + ACCURACY_EVAL_HDRS,
2706 deps = ACCURACY_EVAL_DEPS,
2707)
2708
Marat Dukhan08c4a432019-10-03 09:29:21 -07002709######################### Unit tests for micro-kernels #########################
2710
2711xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002712 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002713 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002714 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002715 "test/gemm-microkernel-tester.h",
2716 "src/xnnpack/AlignedAllocator.h",
2717 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2718 deps = MICROKERNEL_TEST_DEPS,
2719)
2720
2721xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002722 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002723 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002724 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002725 "test/spmm-microkernel-tester.h",
2726 "src/xnnpack/AlignedAllocator.h",
2727 ] + MICROKERNEL_TEST_HDRS,
2728 deps = MICROKERNEL_TEST_DEPS,
2729)
2730
2731xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732 name = "f32_argmaxpool_test",
2733 srcs = [
2734 "test/f32-argmaxpool.cc",
2735 "test/argmaxpool-microkernel-tester.h",
2736 "src/xnnpack/AlignedAllocator.h",
2737 ] + MICROKERNEL_TEST_HDRS,
2738 deps = MICROKERNEL_TEST_DEPS,
2739)
2740
2741xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002742 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002744 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002745 "test/avgpool-microkernel-tester.h",
2746 "src/xnnpack/AlignedAllocator.h",
2747 ] + MICROKERNEL_TEST_HDRS,
2748 deps = MICROKERNEL_TEST_DEPS,
2749)
2750
2751xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002752 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002753 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002754 "test/f32-ibilinear.cc",
2755 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002756 "src/xnnpack/AlignedAllocator.h",
2757 ] + MICROKERNEL_TEST_HDRS,
2758 deps = MICROKERNEL_TEST_DEPS,
2759)
2760
2761xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002762 name = "f16_clamp_test",
2763 srcs = [
2764 "test/f16-clamp.cc",
2765 "test/clamp-microkernel-tester.h",
2766 ] + MICROKERNEL_TEST_HDRS,
2767 deps = MICROKERNEL_TEST_DEPS,
2768)
2769
2770xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002771 name = "f32_clamp_test",
2772 srcs = [
2773 "test/f32-clamp.cc",
2774 "test/clamp-microkernel-tester.h",
2775 ] + MICROKERNEL_TEST_HDRS,
2776 deps = MICROKERNEL_TEST_DEPS,
2777)
2778
2779xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002780 name = "f32_igemm_test",
2781 srcs = [
2782 "test/f32-igemm.cc",
2783 "test/gemm-microkernel-tester.h",
2784 "src/xnnpack/AlignedAllocator.h",
2785 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2786 deps = MICROKERNEL_TEST_DEPS,
2787)
2788
2789xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002790 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002791 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002792 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002793 "test/gemm-microkernel-tester.h",
2794 "src/xnnpack/AlignedAllocator.h",
2795 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002796 # Work-around for "too many sections" error
2797 mingw_copts = ["-Wa,-mbig-obj"],
2798 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002799 deps = MICROKERNEL_TEST_DEPS,
2800)
2801
2802xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002803 name = "f16_igemm_minmax_test",
2804 srcs = [
2805 "test/f16-igemm-minmax.cc",
2806 "test/gemm-microkernel-tester.h",
2807 "src/xnnpack/AlignedAllocator.h",
2808 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2809 deps = MICROKERNEL_TEST_DEPS,
2810)
2811
2812xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002813 name = "f32_conv_hwc_test",
2814 srcs = [
2815 "test/f32-conv-hwc.cc",
2816 "test/conv-hwc-microkernel-tester.h",
2817 "src/xnnpack/AlignedAllocator.h",
2818 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2819 deps = MICROKERNEL_TEST_DEPS,
2820)
2821
2822xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002823 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002825 "test/f32-conv-hwc2chw.cc",
2826 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002827 "src/xnnpack/AlignedAllocator.h",
2828 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2829 deps = MICROKERNEL_TEST_DEPS,
2830)
2831
2832xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002833 name = "f32_dwconv_test",
2834 srcs = [
2835 "test/f32-dwconv.cc",
2836 "test/dwconv-microkernel-tester.h",
2837 "src/xnnpack/AlignedAllocator.h",
2838 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2839 deps = MICROKERNEL_TEST_DEPS,
2840)
2841
2842xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002843 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002844 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002845 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002846 "test/dwconv-microkernel-tester.h",
2847 "src/xnnpack/AlignedAllocator.h",
2848 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2849 deps = MICROKERNEL_TEST_DEPS,
2850)
2851
2852xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002853 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002854 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002855 "test/f32-dwconv-chw.cc",
2856 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002857 "src/xnnpack/AlignedAllocator.h",
2858 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2859 deps = MICROKERNEL_TEST_DEPS,
2860)
2861
2862xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002863 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002864 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002865 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002866 "test/gavgpool-microkernel-tester.h",
2867 "src/xnnpack/AlignedAllocator.h",
2868 ] + MICROKERNEL_TEST_HDRS,
2869 deps = MICROKERNEL_TEST_DEPS,
2870)
2871
2872xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002873 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002874 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002875 "test/f32-gavgpool-cw.cc",
2876 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002877 "src/xnnpack/AlignedAllocator.h",
2878 ] + MICROKERNEL_TEST_HDRS,
2879 deps = MICROKERNEL_TEST_DEPS,
2880)
2881
2882xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002883 name = "f32_gemm_test",
2884 srcs = [
2885 "test/f32-gemm.cc",
2886 "test/gemm-microkernel-tester.h",
2887 "src/xnnpack/AlignedAllocator.h",
2888 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2889 deps = MICROKERNEL_TEST_DEPS,
2890)
2891
2892xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002893 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002894 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002895 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002896 "test/gemm-microkernel-tester.h",
2897 "src/xnnpack/AlignedAllocator.h",
2898 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2899 deps = MICROKERNEL_TEST_DEPS,
2900)
2901
2902xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002903 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002904 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002905 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002906 "test/gemm-microkernel-tester.h",
2907 "src/xnnpack/AlignedAllocator.h",
2908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2909 deps = MICROKERNEL_TEST_DEPS,
2910)
2911
2912xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002913 name = "f16_hswish_test",
2914 srcs = [
2915 "test/f16-hswish.cc",
2916 "test/hswish-microkernel-tester.h",
2917 ] + MICROKERNEL_TEST_HDRS,
2918 deps = MICROKERNEL_TEST_DEPS,
2919)
2920
2921xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 name = "f32_hswish_test",
2923 srcs = [
2924 "test/f32-hswish.cc",
2925 "test/hswish-microkernel-tester.h",
2926 ] + MICROKERNEL_TEST_HDRS,
2927 deps = MICROKERNEL_TEST_DEPS,
2928)
2929
2930xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002931 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002932 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002933 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 "test/maxpool-microkernel-tester.h",
2935 ] + MICROKERNEL_TEST_HDRS,
2936 deps = MICROKERNEL_TEST_DEPS,
2937)
2938
2939xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002940 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002942 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943 "test/avgpool-microkernel-tester.h",
2944 "src/xnnpack/AlignedAllocator.h",
2945 ] + MICROKERNEL_TEST_HDRS,
2946 deps = MICROKERNEL_TEST_DEPS,
2947)
2948
2949xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002950 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002952 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002953 "test/gemm-microkernel-tester.h",
2954 "src/xnnpack/AlignedAllocator.h",
2955 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2956 deps = MICROKERNEL_TEST_DEPS,
2957)
2958
2959xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002960 name = "f16_prelu_test",
2961 srcs = [
2962 "test/f16-prelu.cc",
2963 "test/prelu-microkernel-tester.h",
2964 "src/xnnpack/AlignedAllocator.h",
2965 ] + MICROKERNEL_TEST_HDRS,
2966 deps = MICROKERNEL_TEST_DEPS,
2967)
2968
2969xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 name = "f32_prelu_test",
2971 srcs = [
2972 "test/f32-prelu.cc",
2973 "test/prelu-microkernel-tester.h",
2974 "src/xnnpack/AlignedAllocator.h",
2975 ] + MICROKERNEL_TEST_HDRS,
2976 deps = MICROKERNEL_TEST_DEPS,
2977)
2978
2979xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002980 name = "f32_raddexpminusmax_test",
2981 srcs = [
2982 "test/f32-raddexpminusmax.cc",
2983 "test/raddexpminusmax-microkernel-tester.h",
2984 ] + MICROKERNEL_TEST_HDRS,
2985 deps = MICROKERNEL_TEST_DEPS,
2986)
2987
2988xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002989 name = "f32_raddextexp_test",
2990 srcs = [
2991 "test/f32-raddextexp.cc",
2992 "test/raddextexp-microkernel-tester.h",
2993 ] + MICROKERNEL_TEST_HDRS,
2994 deps = MICROKERNEL_TEST_DEPS,
2995)
2996
2997xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002998 name = "f32_raddstoreexpminusmax_test",
2999 srcs = [
3000 "test/f32-raddstoreexpminusmax.cc",
3001 "test/raddstoreexpminusmax-microkernel-tester.h",
3002 ] + MICROKERNEL_TEST_HDRS,
3003 deps = MICROKERNEL_TEST_DEPS,
3004)
3005
3006xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 name = "f32_rmax_test",
3008 srcs = [
3009 "test/f32-rmax.cc",
3010 "test/rmax-microkernel-tester.h",
3011 ] + MICROKERNEL_TEST_HDRS,
3012 deps = MICROKERNEL_TEST_DEPS,
3013)
3014
3015xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003016 name = "f32_sigmoid_test",
3017 srcs = [
3018 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003019 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003020 ] + MICROKERNEL_TEST_HDRS,
3021 deps = MICROKERNEL_TEST_DEPS,
3022)
3023
3024xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003025 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003026 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003027 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003028 "test/spmm-microkernel-tester.h",
3029 "src/xnnpack/AlignedAllocator.h",
3030 ] + MICROKERNEL_TEST_HDRS,
3031 deps = MICROKERNEL_TEST_DEPS,
3032)
3033
3034xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003035 name = "f16_vadd_minmax_test",
3036 srcs = [
3037 "test/f16-vadd-minmax.cc",
3038 "test/vbinary-microkernel-tester.h",
3039 ] + MICROKERNEL_TEST_HDRS,
3040 deps = MICROKERNEL_TEST_DEPS,
3041)
3042
3043xnnpack_unit_test(
3044 name = "f16_vaddc_minmax_test",
3045 srcs = [
3046 "test/f16-vaddc-minmax.cc",
3047 "test/vbinaryc-microkernel-tester.h",
3048 ] + MICROKERNEL_TEST_HDRS,
3049 deps = MICROKERNEL_TEST_DEPS,
3050)
3051
3052xnnpack_unit_test(
3053 name = "f16_vdiv_minmax_test",
3054 srcs = [
3055 "test/f16-vdiv-minmax.cc",
3056 "test/vbinary-microkernel-tester.h",
3057 ] + MICROKERNEL_TEST_HDRS,
3058 deps = MICROKERNEL_TEST_DEPS,
3059)
3060
3061xnnpack_unit_test(
3062 name = "f16_vdivc_minmax_test",
3063 srcs = [
3064 "test/f16-vdivc-minmax.cc",
3065 "test/vbinaryc-microkernel-tester.h",
3066 ] + MICROKERNEL_TEST_HDRS,
3067 deps = MICROKERNEL_TEST_DEPS,
3068)
3069
3070xnnpack_unit_test(
3071 name = "f16_vrdivc_minmax_test",
3072 srcs = [
3073 "test/f16-vrdivc-minmax.cc",
3074 "test/vbinaryc-microkernel-tester.h",
3075 ] + MICROKERNEL_TEST_HDRS,
3076 deps = MICROKERNEL_TEST_DEPS,
3077)
3078
3079xnnpack_unit_test(
3080 name = "f16_vmax_test",
3081 srcs = [
3082 "test/f16-vmax.cc",
3083 "test/vbinary-microkernel-tester.h",
3084 ] + MICROKERNEL_TEST_HDRS,
3085 deps = MICROKERNEL_TEST_DEPS,
3086)
3087
3088xnnpack_unit_test(
3089 name = "f16_vmaxc_test",
3090 srcs = [
3091 "test/f16-vmaxc.cc",
3092 "test/vbinaryc-microkernel-tester.h",
3093 ] + MICROKERNEL_TEST_HDRS,
3094 deps = MICROKERNEL_TEST_DEPS,
3095)
3096
3097xnnpack_unit_test(
3098 name = "f16_vmin_test",
3099 srcs = [
3100 "test/f16-vmin.cc",
3101 "test/vbinary-microkernel-tester.h",
3102 ] + MICROKERNEL_TEST_HDRS,
3103 deps = MICROKERNEL_TEST_DEPS,
3104)
3105
3106xnnpack_unit_test(
3107 name = "f16_vminc_test",
3108 srcs = [
3109 "test/f16-vminc.cc",
3110 "test/vbinaryc-microkernel-tester.h",
3111 ] + MICROKERNEL_TEST_HDRS,
3112 deps = MICROKERNEL_TEST_DEPS,
3113)
3114
3115xnnpack_unit_test(
3116 name = "f16_vmul_minmax_test",
3117 srcs = [
3118 "test/f16-vmul-minmax.cc",
3119 "test/vbinary-microkernel-tester.h",
3120 ] + MICROKERNEL_TEST_HDRS,
3121 deps = MICROKERNEL_TEST_DEPS,
3122)
3123
3124xnnpack_unit_test(
3125 name = "f16_vmulc_minmax_test",
3126 srcs = [
3127 "test/f16-vmulc-minmax.cc",
3128 "test/vbinaryc-microkernel-tester.h",
3129 ] + MICROKERNEL_TEST_HDRS,
3130 deps = MICROKERNEL_TEST_DEPS,
3131)
3132
3133xnnpack_unit_test(
3134 name = "f16_vsub_minmax_test",
3135 srcs = [
3136 "test/f16-vsub-minmax.cc",
3137 "test/vbinary-microkernel-tester.h",
3138 ] + MICROKERNEL_TEST_HDRS,
3139 deps = MICROKERNEL_TEST_DEPS,
3140)
3141
3142xnnpack_unit_test(
3143 name = "f16_vsubc_minmax_test",
3144 srcs = [
3145 "test/f16-vsubc-minmax.cc",
3146 "test/vbinaryc-microkernel-tester.h",
3147 ] + MICROKERNEL_TEST_HDRS,
3148 deps = MICROKERNEL_TEST_DEPS,
3149)
3150
3151xnnpack_unit_test(
3152 name = "f16_vrsubc_minmax_test",
3153 srcs = [
3154 "test/f16-vrsubc-minmax.cc",
3155 "test/vbinaryc-microkernel-tester.h",
3156 ] + MICROKERNEL_TEST_HDRS,
3157 deps = MICROKERNEL_TEST_DEPS,
3158)
3159
3160xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003161 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003163 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003164 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003165 ] + MICROKERNEL_TEST_HDRS,
3166 deps = MICROKERNEL_TEST_DEPS,
3167)
3168
3169xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003170 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003171 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003172 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003173 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174 ] + MICROKERNEL_TEST_HDRS,
3175 deps = MICROKERNEL_TEST_DEPS,
3176)
3177
3178xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003179 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003180 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003181 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003182 "test/vbinary-microkernel-tester.h",
3183 ] + MICROKERNEL_TEST_HDRS,
3184 deps = MICROKERNEL_TEST_DEPS,
3185)
3186
3187xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003188 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003189 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003190 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003191 "test/vbinaryc-microkernel-tester.h",
3192 ] + MICROKERNEL_TEST_HDRS,
3193 deps = MICROKERNEL_TEST_DEPS,
3194)
3195
3196xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003197 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003198 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003199 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003200 "test/vbinaryc-microkernel-tester.h",
3201 ] + MICROKERNEL_TEST_HDRS,
3202 deps = MICROKERNEL_TEST_DEPS,
3203)
3204
3205xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003206 name = "f32_vmax_test",
3207 srcs = [
3208 "test/f32-vmax.cc",
3209 "test/vbinary-microkernel-tester.h",
3210 ] + MICROKERNEL_TEST_HDRS,
3211 deps = MICROKERNEL_TEST_DEPS,
3212)
3213
3214xnnpack_unit_test(
3215 name = "f32_vmaxc_test",
3216 srcs = [
3217 "test/f32-vmaxc.cc",
3218 "test/vbinaryc-microkernel-tester.h",
3219 ] + MICROKERNEL_TEST_HDRS,
3220 deps = MICROKERNEL_TEST_DEPS,
3221)
3222
3223xnnpack_unit_test(
3224 name = "f32_vmin_test",
3225 srcs = [
3226 "test/f32-vmin.cc",
3227 "test/vbinary-microkernel-tester.h",
3228 ] + MICROKERNEL_TEST_HDRS,
3229 deps = MICROKERNEL_TEST_DEPS,
3230)
3231
3232xnnpack_unit_test(
3233 name = "f32_vminc_test",
3234 srcs = [
3235 "test/f32-vminc.cc",
3236 "test/vbinaryc-microkernel-tester.h",
3237 ] + MICROKERNEL_TEST_HDRS,
3238 deps = MICROKERNEL_TEST_DEPS,
3239)
3240
3241xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003242 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003243 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003244 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003245 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003246 ] + MICROKERNEL_TEST_HDRS,
3247 deps = MICROKERNEL_TEST_DEPS,
3248)
3249
3250xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003251 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003252 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003253 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003254 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003255 ] + MICROKERNEL_TEST_HDRS,
3256 deps = MICROKERNEL_TEST_DEPS,
3257)
3258
3259xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003260 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003261 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003262 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003263 "test/vmulcaddc-microkernel-tester.h",
3264 "src/xnnpack/AlignedAllocator.h",
3265 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3266 deps = MICROKERNEL_TEST_DEPS,
3267)
3268
3269xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003270 name = "f32_vscale_test",
3271 srcs = [
3272 "test/f32-vscale.cc",
3273 "test/vscale-microkernel-tester.h",
3274 ] + MICROKERNEL_TEST_HDRS,
3275 deps = MICROKERNEL_TEST_DEPS,
3276)
3277
3278xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003279 name = "f32_vscaleexpminusmax_test",
3280 srcs = [
3281 "test/f32-vscaleexpminusmax.cc",
3282 "test/vscaleexpminusmax-microkernel-tester.h",
3283 ] + MICROKERNEL_TEST_HDRS,
3284 deps = MICROKERNEL_TEST_DEPS,
3285)
3286
3287xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003288 name = "f32_vscaleextexp_test",
3289 srcs = [
3290 "test/f32-vscaleextexp.cc",
3291 "test/vscaleextexp-microkernel-tester.h",
3292 ] + MICROKERNEL_TEST_HDRS,
3293 deps = MICROKERNEL_TEST_DEPS,
3294)
3295
3296xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003297 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003298 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003299 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003300 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003301 ] + MICROKERNEL_TEST_HDRS,
3302 deps = MICROKERNEL_TEST_DEPS,
3303)
3304
3305xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003306 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003307 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003308 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003309 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003310 ] + MICROKERNEL_TEST_HDRS,
3311 deps = MICROKERNEL_TEST_DEPS,
3312)
3313
3314xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003315 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003316 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003317 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003318 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003319 ] + MICROKERNEL_TEST_HDRS,
3320 deps = MICROKERNEL_TEST_DEPS,
3321)
3322
3323xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003324 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003326 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003327 "test/avgpool-microkernel-tester.h",
3328 "src/xnnpack/AlignedAllocator.h",
3329 ] + MICROKERNEL_TEST_HDRS,
3330 deps = MICROKERNEL_TEST_DEPS,
3331)
3332
3333xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003334 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003335 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003336 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003337 "test/gemm-microkernel-tester.h",
3338 "src/xnnpack/AlignedAllocator.h",
3339 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3340 deps = MICROKERNEL_TEST_DEPS,
3341)
3342
3343xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003344 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003345 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003346 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 "test/dwconv-microkernel-tester.h",
3348 "src/xnnpack/AlignedAllocator.h",
3349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3350 deps = MICROKERNEL_TEST_DEPS,
3351)
3352
3353xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003354 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003355 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003356 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003357 "test/gavgpool-microkernel-tester.h",
3358 "src/xnnpack/AlignedAllocator.h",
3359 ] + MICROKERNEL_TEST_HDRS,
3360 deps = MICROKERNEL_TEST_DEPS,
3361)
3362
3363xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003364 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003365 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003366 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003367 "test/gemm-microkernel-tester.h",
3368 "src/xnnpack/AlignedAllocator.h",
3369 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3370 deps = MICROKERNEL_TEST_DEPS,
3371)
3372
3373xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003374 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003376 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 "test/vadd-microkernel-tester.h",
3378 ] + MICROKERNEL_TEST_HDRS,
3379 deps = MICROKERNEL_TEST_DEPS,
3380)
3381
3382xnnpack_unit_test(
3383 name = "u8_clamp_test",
3384 srcs = [
3385 "test/u8-clamp.cc",
3386 "test/clamp-microkernel-tester.h",
3387 ] + MICROKERNEL_TEST_HDRS,
3388 deps = MICROKERNEL_TEST_DEPS,
3389)
3390
3391xnnpack_unit_test(
3392 name = "u8_lut32norm_test",
3393 srcs = [
3394 "test/u8-lut32norm.cc",
3395 "test/lut-norm-microkernel-tester.h",
3396 ] + MICROKERNEL_TEST_HDRS,
3397 deps = MICROKERNEL_TEST_DEPS,
3398)
3399
3400xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003401 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003403 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404 "test/maxpool-microkernel-tester.h",
3405 ] + MICROKERNEL_TEST_HDRS,
3406 deps = MICROKERNEL_TEST_DEPS,
3407)
3408
3409xnnpack_unit_test(
3410 name = "u8_rmax_test",
3411 srcs = [
3412 "test/u8-rmax.cc",
3413 "test/rmax-microkernel-tester.h",
3414 ] + MICROKERNEL_TEST_HDRS,
3415 deps = MICROKERNEL_TEST_DEPS,
3416)
3417
3418xnnpack_unit_test(
3419 name = "x32_packx_test",
3420 srcs = [
3421 "test/x32-packx.cc",
3422 "test/pack-microkernel-tester.h",
3423 "src/xnnpack/AlignedAllocator.h",
3424 ] + MICROKERNEL_TEST_HDRS,
3425 deps = MICROKERNEL_TEST_DEPS,
3426)
3427
3428xnnpack_unit_test(
3429 name = "x32_pad_test",
3430 srcs = [
3431 "test/x32-pad.cc",
3432 "test/pad-microkernel-tester.h",
3433 ] + MICROKERNEL_TEST_HDRS,
3434 deps = MICROKERNEL_TEST_DEPS,
3435)
3436
3437xnnpack_unit_test(
3438 name = "x32_unpool_test",
3439 srcs = [
3440 "test/x32-unpool.cc",
3441 "test/unpool-microkernel-tester.h",
3442 ] + MICROKERNEL_TEST_HDRS,
3443 deps = MICROKERNEL_TEST_DEPS,
3444)
3445
3446xnnpack_unit_test(
3447 name = "x32_zip_test",
3448 srcs = [
3449 "test/x32-zip.cc",
3450 "test/zip-microkernel-tester.h",
3451 ] + MICROKERNEL_TEST_HDRS,
3452 deps = MICROKERNEL_TEST_DEPS,
3453)
3454
3455xnnpack_unit_test(
3456 name = "x8_lut_test",
3457 srcs = [
3458 "test/x8-lut.cc",
3459 "test/lut-microkernel-tester.h",
3460 ] + MICROKERNEL_TEST_HDRS,
3461 deps = MICROKERNEL_TEST_DEPS,
3462)
3463
3464xnnpack_unit_test(
3465 name = "x8_zip_test",
3466 srcs = [
3467 "test/x8-zip.cc",
3468 "test/zip-microkernel-tester.h",
3469 ] + MICROKERNEL_TEST_HDRS,
3470 deps = MICROKERNEL_TEST_DEPS,
3471)
3472
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003473xnnpack_unit_test(
3474 name = "requantization_test",
3475 srcs = [
3476 "src/xnnpack/requantization-stubs.h",
3477 "test/requantization.cc",
3478 "test/requantization-tester.h",
3479 ] + MICROKERNEL_TEST_HDRS,
3480 deps = MICROKERNEL_TEST_DEPS,
3481)
3482
Marat Dukhan20c3b922020-03-10 03:45:06 -07003483########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484
3485xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003486 name = "operator_size_test",
3487 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003488 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489)
3490
Marat Dukhan20c3b922020-03-10 03:45:06 -07003491xnnpack_binary(
3492 name = "subgraph_size_test",
3493 srcs = ["test/subgraph-size.c"],
3494 deps = [":XNNPACK"],
3495)
3496
3497########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003498
3499xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003500 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003502 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003503 "test/add-operator-tester.h",
3504 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003505 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003506)
3507
3508xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003509 name = "add_nd_test",
3510 srcs = [
3511 "test/add-nd.cc",
3512 "test/binary-elementwise-operator-tester.h",
3513 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003514 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003515)
3516
3517xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003518 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003519 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003520 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521 "test/argmax-pooling-operator-tester.h",
3522 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003523 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003524)
3525
3526xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003527 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003528 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003529 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530 "test/average-pooling-operator-tester.h",
3531 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003533)
3534
3535xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003536 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003537 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003538 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539 "test/channel-pad-operator-tester.h",
3540 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003541 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003542)
3543
3544xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003545 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003547 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003548 "test/channel-shuffle-operator-tester.h",
3549 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003550 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003551)
3552
3553xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003554 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003556 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 "test/clamp-operator-tester.h",
3558 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003559 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003560)
3561
3562xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003563 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003565 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566 "test/convolution-operator-tester.h",
3567 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003568 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569)
3570
3571xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003572 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003574 "test/convolution-nchw.cc",
3575 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003577 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003578)
3579
3580xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003581 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003583 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584 "test/deconvolution-operator-tester.h",
3585 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003586 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587)
3588
3589xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003590 name = "divide_nd_test",
3591 srcs = [
3592 "test/binary-elementwise-operator-tester.h",
3593 "test/divide-nd.cc",
3594 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003595 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003596)
3597
3598xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003599 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003600 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003601 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 "test/fully-connected-operator-tester.h",
3603 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003604 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605)
3606
3607xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003608 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003610 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611 "test/global-average-pooling-operator-tester.h",
3612 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003613 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003614)
3615
3616xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003617 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003619 "test/global-average-pooling-ncw.cc",
3620 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003622 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003623)
3624
3625xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003626 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003628 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629 "test/hardswish-operator-tester.h",
3630 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003631 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632)
3633
3634xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003635 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003637 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003638 "test/leaky-relu-operator-tester.h",
3639 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003640 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641)
3642
3643xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003644 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003646 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647 "test/max-pooling-operator-tester.h",
3648 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003649 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003650)
3651
3652xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003653 name = "maximum_nd_test",
3654 srcs = [
3655 "test/binary-elementwise-operator-tester.h",
3656 "test/maximum-nd.cc",
3657 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003658 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003659)
3660
3661xnnpack_unit_test(
3662 name = "minimum_nd_test",
3663 srcs = [
3664 "test/binary-elementwise-operator-tester.h",
3665 "test/minimum-nd.cc",
3666 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003667 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003668)
3669
3670xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003671 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003672 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003673 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003674 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003675 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003676 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003677)
3678
3679xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003680 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003681 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003682 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003683 "test/prelu-operator-tester.h",
3684 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003685 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686)
3687
3688xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003689 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003690 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003691 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003692 "test/resize-bilinear-operator-tester.h",
3693 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003694 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003695)
3696
3697xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003698 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003700 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701 "test/sigmoid-operator-tester.h",
3702 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003703 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003704)
3705
3706xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003707 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003708 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003709 "test/softmax-nc.cc",
3710 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003712 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003713)
3714
3715xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003716 name = "subtract_nd_test",
3717 srcs = [
3718 "test/binary-elementwise-operator-tester.h",
3719 "test/subtract-nd.cc",
3720 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003721 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003722)
3723
3724xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003725 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003726 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003727 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 "test/unpooling-operator-tester.h",
3729 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003730 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003731)
3732
Chao Mei6ddfc602020-05-13 22:29:36 -07003733############################### Misc unit tests ###############################
3734
3735xnnpack_unit_test(
3736 name = "memory_planner_test",
3737 srcs = [
3738 "test/memory-planner-test.cc",
3739 ],
3740 deps = [
3741 ":XNNPACK",
3742 ":memory_planner",
3743 ],
3744)
3745
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746############################# Build configurations #############################
3747
Marat Dukhanb8642352019-10-30 15:43:02 -07003748# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003750 name = "xnn_enable_assembly_explicit_true",
3751 define_values = {"xnn_enable_assembly": "true"},
3752)
3753
3754# Disables usage of assembly kernels.
3755config_setting(
3756 name = "xnn_enable_assembly_explicit_false",
3757 define_values = {"xnn_enable_assembly": "false"},
3758)
3759
Marat Dukhan05702cf2020-03-26 15:41:33 -07003760# Disables usage of HMP-aware optimizations.
3761config_setting(
3762 name = "xnn_enable_hmp_explicit_false",
3763 define_values = {"xnn_enable_hmp": "false"},
3764)
3765
Chao Mei6ddfc602020-05-13 22:29:36 -07003766# Enable usage of optimized memory allocation
3767config_setting(
3768 name = "xnn_enable_memopt_explicit_true",
3769 define_values = {"xnnpack_enable_memopt": "true"},
3770)
3771
3772# Disable usage of optimized memory allocation
3773config_setting(
3774 name = "xnn_enable_memopt_explicit_false",
3775 define_values = {"xnnpack_enable_memopt": "false"},
3776)
3777
Marat Dukhanb8642352019-10-30 15:43:02 -07003778# Builds with -c dbg
3779config_setting(
3780 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003781 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003782 "compilation_mode": "dbg",
3783 },
3784)
3785
3786# Builds with -c opt
3787config_setting(
3788 name = "optimized_build",
3789 values = {
3790 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791 },
3792)
3793
3794config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003795 name = "linux_k8",
3796 values = {"cpu": "k8"},
3797)
3798
3799config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003800 name = "linux_arm",
3801 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003802)
3803
3804config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003805 name = "linux_armhf",
3806 values = {"cpu": "armhf"},
3807)
3808
3809config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003810 name = "linux_aarch64",
3811 values = {"cpu": "aarch64"},
3812)
3813
3814config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003815 name = "android",
3816 values = {"crosstool_top": "//external:android/crosstool"},
3817)
3818
3819config_setting(
3820 name = "android_armv7",
3821 values = {
3822 "crosstool_top": "//external:android/crosstool",
3823 "cpu": "armeabi-v7a",
3824 },
3825)
3826
3827config_setting(
3828 name = "android_arm64",
3829 values = {
3830 "crosstool_top": "//external:android/crosstool",
3831 "cpu": "arm64-v8a",
3832 },
3833)
3834
3835config_setting(
3836 name = "android_x86",
3837 values = {
3838 "crosstool_top": "//external:android/crosstool",
3839 "cpu": "x86",
3840 },
3841)
3842
3843config_setting(
3844 name = "android_x86_64",
3845 values = {
3846 "crosstool_top": "//external:android/crosstool",
3847 "cpu": "x86_64",
3848 },
3849)
3850
3851config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003852 name = "windows_x86_64",
3853 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003854)
3855
3856config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003857 name = "windows_x86_64_clang",
3858 values = {
3859 "compiler": "clang-cl",
3860 "cpu": "x64_windows",
3861 },
3862)
3863
3864config_setting(
3865 name = "windows_x86_64_mingw",
3866 values = {
3867 "compiler": "mingw-gcc",
3868 "cpu": "x64_windows",
3869 },
3870)
3871
3872config_setting(
3873 name = "windows_x86_64_msys",
3874 values = {
3875 "compiler": "msys-gcc",
3876 "cpu": "x64_windows",
3877 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003878)
3879
3880config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003881 name = "macos_x86_64",
3882 values = {
3883 "apple_platform_type": "macos",
3884 "cpu": "darwin",
3885 },
3886)
3887
3888config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003889 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003890 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003891)
3892
3893config_setting(
3894 name = "emscripten_wasm",
3895 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003896 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003897 "cpu": "wasm",
3898 },
3899)
3900
3901config_setting(
3902 name = "emscripten_wasmsimd",
3903 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003904 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003906 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003907 },
3908)
3909
3910config_setting(
3911 name = "emscripten_asmjs",
3912 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003913 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 "cpu": "asmjs",
3915 },
3916)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003917
3918config_setting(
3919 name = "ios_armv7",
3920 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003921 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003922 "cpu": "ios_armv7",
3923 },
3924)
3925
3926config_setting(
3927 name = "ios_arm64",
3928 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003929 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003930 "cpu": "ios_arm64",
3931 },
3932)
3933
3934config_setting(
3935 name = "ios_arm64e",
3936 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003937 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003938 "cpu": "ios_arm64e",
3939 },
3940)
3941
3942config_setting(
3943 name = "ios_x86",
3944 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003945 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003946 "cpu": "ios_i386",
3947 },
3948)
3949
3950config_setting(
3951 name = "ios_x86_64",
3952 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003953 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003954 "cpu": "ios_x86_64",
3955 },
3956)
3957
3958config_setting(
3959 name = "watchos_armv7k",
3960 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003961 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003962 "cpu": "watchos_armv7k",
3963 },
3964)
3965
3966config_setting(
3967 name = "watchos_arm64_32",
3968 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003969 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003970 "cpu": "watchos_arm64_32",
3971 },
3972)
3973
3974config_setting(
3975 name = "watchos_x86",
3976 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003977 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003978 "cpu": "watchos_i386",
3979 },
3980)
3981
3982config_setting(
3983 name = "watchos_x86_64",
3984 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003985 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003986 "cpu": "watchos_x86_64",
3987 },
3988)
3989
3990config_setting(
3991 name = "tvos_arm64",
3992 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003993 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003994 "cpu": "tvos_arm64",
3995 },
3996)
3997
3998config_setting(
3999 name = "tvos_x86_64",
4000 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004001 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004002 "cpu": "tvos_x86_64",
4003 },
4004)