blob: 186a50c68f4f8096f1a9ee5d60b97cb3953fd10f [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 Dukhance7a3f82020-05-17 21:46:44 -0700524 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
525 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
526 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
527 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700528 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
529 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
530 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
531 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700532 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
533 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
534 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
535 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700536 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
537 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
538 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
539 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700540 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700541 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
542 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700543 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
544 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
545 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
546 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
547 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
548 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
549 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
550 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
551 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
552 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
553 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
554 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
555 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
556 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
557 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
558 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
559 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
560 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
561 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
562 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
563 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
564 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
565 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
566 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
567 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
568 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
569 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
570 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
571 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
572 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
573 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800574 "src/f32-hswish/gen/neon-x4.c",
575 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700576 "src/f32-ibilinear/gen/neon-c4.c",
577 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700578 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
579 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
580 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
581 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
582 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
583 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
584 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
585 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
586 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
587 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
588 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
589 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
590 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
591 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
592 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
593 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700594 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
595 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
596 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700597 "src/f32-ppmm/gen/4x8-minmax-neon.c",
598 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800599 "src/f32-prelu/gen/neon-2x4.c",
600 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800601 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
602 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
603 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
604 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
605 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
606 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
607 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
608 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
609 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
610 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
611 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
612 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
613 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
614 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
615 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
616 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
617 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
618 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
619 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
620 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
621 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
622 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
623 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
624 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700625 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800626 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800627 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
628 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
629 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
630 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
631 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
632 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
633 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
634 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
635 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
636 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
637 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
638 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
639 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
640 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
641 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
642 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
643 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
644 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700645 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
646 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
647 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
648 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800649 "src/f32-vbinary/gen/vmax-neon-x4.c",
650 "src/f32-vbinary/gen/vmax-neon-x8.c",
651 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
652 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
653 "src/f32-vbinary/gen/vmin-neon-x4.c",
654 "src/f32-vbinary/gen/vmin-neon-x8.c",
655 "src/f32-vbinary/gen/vminc-neon-x4.c",
656 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700657 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
658 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
659 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
660 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
661 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
662 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
663 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
664 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
665 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
666 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700667 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
668 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
669 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
670 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700671 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700672 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
673 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700674 "src/q8-gemm/4x8-minmax-neon.c",
675 "src/q8-gemm/8x8-minmax-neon.c",
676 "src/q8-igemm/4x8-minmax-neon.c",
677 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700678 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700679 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700680 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700681 "src/u8-rmax/neon.c",
682 "src/x32-packx/x4-neon-st4.c",
683 "src/x32-pad/x2-neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700684 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700685 "src/x32-zip/x2-neon.c",
686 "src/x32-zip/x3-neon.c",
687 "src/x32-zip/x4-neon.c",
688 "src/x32-zip/xm-neon.c",
689 "src/x8-zip/x2-neon.c",
690 "src/x8-zip/x3-neon.c",
691 "src/x8-zip/x4-neon.c",
692 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700693 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700694 "src/math/roundd-neon-addsub.c",
695 "src/math/roundd-neon-cvt.c",
696 "src/math/roundu-neon-addsub.c",
697 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700698 "src/math/roundz-neon-addsub.c",
699 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800700 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800701 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
702 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
703 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
704 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
705 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
706 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700707 "src/requantization/precise-neon.c",
708 "src/requantization/fp32-neon.c",
709 "src/requantization/q31-neon.c",
710 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700711]
712
713NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700714 "src/f32-ibilinear/gen/neonfma-c4.c",
715 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700716 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
717 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
718 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
719 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
720 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
721 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
722 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
723 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
724 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700725 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
726 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
727 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
728 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700729 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
730 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
731 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
732 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700733 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
734 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
735 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
736 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700737 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
738 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
739 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
740 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
741 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
742 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
743 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
744 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
745 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
746 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
747 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
748 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
749 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
750 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
751 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
752 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
753 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
754 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800755 "src/f32-hswish/gen/neonfma-x4.c",
756 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700757 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
758 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800759 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
760 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
761 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
762 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
763 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
764 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
765 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
766 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
767 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
768 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
769 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
770 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
771 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
772 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
773 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
774 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
775 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
776 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
777 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
778 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
779 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
780 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
781 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
782 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800783 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
784 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
785 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
786 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
787 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
788 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
789 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
790 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
791 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
792 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
832 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
833 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
834 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
835 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
836 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700837 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
838 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800839 "src/math/exp-neonfma-lut64-p2.c",
840 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800841 "src/math/expminus-neonfma-lut2048-p1.c",
842 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800843 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800844 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
845 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
846 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
847 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
848 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
849 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
850 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
851 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
852 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
853 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
854 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
855 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
856 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
857 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
858 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
859 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
860 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
861 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700862]
863
864AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700865 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
866 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
867 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
868 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
869 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
870 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700871 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
872 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
873 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
874 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
875 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
876 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
877 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
878 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
879 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
880 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
881 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
882 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
883 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
884 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
885 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
886 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
887 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
888 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
889 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
890 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700891 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
892 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
893 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
894 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700895 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
896 "src/f32-dwconv-chw/3x3p1-neonfma.c",
897 "src/f32-dwconv-chw/5x5p2-neonfma.c",
898 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
899 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800900 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
901 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
902 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
903 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
904 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
905 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
906 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
907 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
908 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
909 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
910 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
911 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
912 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
913 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
914 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
915 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
916 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
917 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700918 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
919 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
920 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
921 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
922 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
923 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
924 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
925 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
926 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
927 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
928 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
929 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
930 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
931 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
932 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
933 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
934 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
935 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800936 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
937 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
938 "src/math/sigmoid-neonfma-rr1-p5-div.c",
939 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
940 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
941 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700942]
943
Marat Dukhan8853b822020-05-07 12:19:01 -0700944NEONV8_UKERNELS = [
945 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700946 "src/math/roundd-neonv8.c",
947 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700948 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700949]
950
Marat Dukhan08c4a432019-10-03 09:29:21 -0700951AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -0700952 "src/f16-clamp/gen/neonfp16arith-x16.c",
953 "src/f16-clamp/gen/neonfp16arith-x8.c",
954 "src/f16-hswish/gen/neonfp16arith-x16.c",
955 "src/f16-hswish/gen/neonfp16arith-x8.c",
956 "src/f16-prelu/gen/neonfp16arith-2x16.c",
957 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -0700958 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
959 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
960 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
961 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
962 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
963 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
964 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
965 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
966 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
967 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
968 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
969 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
970 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
971 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
972 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
973 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
974 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
975 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
976 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
977 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
978 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
979 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
980 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
981 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
982 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
983 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
984 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
985 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -0700986 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
987 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
988 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
989 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
990 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
991 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
992 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
993 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -0700994 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
995 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
996 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
997 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -0700998 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
999 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1000 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1001 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1002 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1003 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1004 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1005 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1006 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1007 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1008 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1009 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001010 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1011 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1012 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1013 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1014 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1015 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1016 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1017 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001018]
1019
1020SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001021 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1022 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001023 "src/f32-clamp/gen/sse-x4.c",
1024 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001025 "src/f32-dwconv-chw/3x3p1-sse.c",
1026 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001027 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1028 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1029 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1030 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1031 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1032 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1033 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1034 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1035 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1036 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1037 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1038 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001039 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001040 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1041 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001042 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1043 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1044 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1045 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1046 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1047 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1048 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1049 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1050 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1051 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1052 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1053 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1054 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001055 "src/f32-hswish/gen/sse-x4.c",
1056 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001057 "src/f32-ibilinear/gen/sse-c4.c",
1058 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001059 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1060 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1061 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1062 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1063 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1064 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1065 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001066 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1067 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1068 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001069 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001070 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001071 "src/f32-spmm/gen/4x1-minmax-sse.c",
1072 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001073 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1074 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1075 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1076 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1077 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1078 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1079 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1080 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001081 "src/f32-vbinary/gen/vmax-sse-x4.c",
1082 "src/f32-vbinary/gen/vmax-sse-x8.c",
1083 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1084 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1085 "src/f32-vbinary/gen/vmin-sse-x4.c",
1086 "src/f32-vbinary/gen/vmin-sse-x8.c",
1087 "src/f32-vbinary/gen/vminc-sse-x4.c",
1088 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001089 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1090 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1091 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1092 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1093 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1094 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1095 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1096 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1097 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1098 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1099 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1100 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001101 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1102 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001103 "src/x32-packx/x4-sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001104 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001105 "src/math/roundd-sse-addsub.c",
1106 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001107 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001108]
1109
1110SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001111 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1112 "src/f32-argmaxpool/4x-sse2-c4.c",
1113 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001114 "src/f32-prelu/gen/sse2-2x4.c",
1115 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001116 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1117 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1118 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1119 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1120 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1121 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1122 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1123 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1124 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1125 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1126 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1127 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001128 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001129 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001130 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001131 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001132 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1133 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001134 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1135 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001136 "src/q8-igemm/4x4c2-minmax-sse2.c",
1137 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001138 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1139 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001140 "src/q8-gemm/2x4c8-minmax-sse2.c",
1141 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001142 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001143 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001144 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001145 "src/u8-rmax/sse2.c",
1146 "src/x32-pad/x2-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001147 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001148 "src/x32-zip/x2-sse2.c",
1149 "src/x32-zip/x3-sse2.c",
1150 "src/x32-zip/x4-sse2.c",
1151 "src/x32-zip/xm-sse2.c",
1152 "src/x8-zip/x2-sse2.c",
1153 "src/x8-zip/x3-sse2.c",
1154 "src/x8-zip/x4-sse2.c",
1155 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001156 "src/math/exp-sse2-p5.c",
1157 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001158 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001159 "src/math/roundd-sse2-cvt.c",
1160 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001161 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001162 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001163 "src/requantization/precise-sse2.c",
1164 "src/requantization/fp32-sse2.c",
1165 "src/requantization/q31-sse2.c",
1166 "src/requantization/gemmlowp-sse2.c",
1167]
1168
1169SSSE3_UKERNELS = [
1170 "src/requantization/precise-ssse3.c",
1171 "src/requantization/q31-ssse3.c",
1172 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001173]
1174
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001175SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001176 "src/f32-prelu/gen/sse41-2x4.c",
1177 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001178 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001179 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001180 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001181 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001182 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1183 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001184 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001185 "src/math/roundd-sse41.c",
1186 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001187 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001188 "src/requantization/precise-sse4.c",
1189 "src/requantization/q31-sse4.c",
1190 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001191]
1192
Marat Dukhan08c4a432019-10-03 09:29:21 -07001193AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001194 "src/f32-clamp/gen/avx-x8.c",
1195 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001196 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1197 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1198 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1199 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1200 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1201 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1202 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1203 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1204 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1205 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1206 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1207 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1208 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1209 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1210 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1211 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1212 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1213 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1214 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1215 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1216 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1217 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1218 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1219 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1220 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1221 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1222 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1223 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1224 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1225 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001226 "src/f32-hswish/gen/avx-x8.c",
1227 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001228 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1229 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1230 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1231 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1232 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1233 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1234 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1235 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1236 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001237 "src/f32-prelu/gen/avx-2x8.c",
1238 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001239 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001240 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1241 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1242 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1243 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1244 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1245 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1246 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1247 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001248 "src/f32-vbinary/gen/vmax-avx-x8.c",
1249 "src/f32-vbinary/gen/vmax-avx-x16.c",
1250 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1251 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1252 "src/f32-vbinary/gen/vmin-avx-x8.c",
1253 "src/f32-vbinary/gen/vmin-avx-x16.c",
1254 "src/f32-vbinary/gen/vminc-avx-x8.c",
1255 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001256 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1257 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1258 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1259 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1260 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1261 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1262 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1264 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1265 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1266 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1267 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001268 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001269]
1270
Marat Dukhanfda12b82019-11-21 12:27:59 -08001271FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001272 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1273 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1274 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1275 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1276 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1277 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1278 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1279 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1280 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1281 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1282 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1283 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1284 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1285 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1286 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1287 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1288 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1289 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1290 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1291 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1292 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1293 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1294 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1295 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1296 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1297 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1298 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1299 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1300 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1301 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1302 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1303 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1304 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1305 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1306 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1307 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1308 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1309 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1310 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1311 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001312 "src/f32-hswish/gen/fma3-x8.c",
1313 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001314 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1315 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1316 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1317 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1318 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1319 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1320 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1321 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1322 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1323 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1324 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1325 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1326 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1327 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001328]
1329
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001330AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001331 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1332 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1333 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1334 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1335 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1336 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1337 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1338 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1339 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1340 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1341 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1342 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1343 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1344 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1345 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1346 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1347 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1348 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1349 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1350 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1351 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1352 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1353 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1354 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1355 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1356 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1357 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1358 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1359 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1360 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1361 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1362 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1363 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1364 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1365 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1366 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001367 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1368 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1369 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1370 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1371 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1372 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1373 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1374 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1375 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1376 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1377 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1378 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1379 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1380 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1381 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1382 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1383 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1384 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1387 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1388 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1389 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1390 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001397 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1398 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1399 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1400 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1401 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1402 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1403 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1404 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1405 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1406 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1407 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1408 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1409 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1410 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1411 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1412 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1413 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1414 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1415 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1416 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1417 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1418 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1419 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1420 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001421 "src/math/exp-avx2-p5.c",
1422 "src/math/exp-avx2-perm-p3.c",
1423 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001424 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001425 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001426 "src/math/sigmoid-avx2-rr2-p5-div.c",
1427 "src/math/sigmoid-avx2-rr1-p5-div.c",
1428 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1429 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1430 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1431 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001432]
1433
Marat Dukhan08c4a432019-10-03 09:29:21 -07001434AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001435 "src/f32-clamp/gen/avx512f-x16.c",
1436 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001437 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1438 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1439 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1440 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1441 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1442 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1443 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1444 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1445 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1446 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1447 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1448 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1449 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1450 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1451 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1452 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1453 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1454 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1455 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1456 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1457 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1458 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1459 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1460 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001461 "src/f32-hswish/gen/avx512f-x16.c",
1462 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001463 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1464 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1465 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1466 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1467 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1468 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001469 "src/f32-prelu/gen/avx512f-2x16.c",
1470 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001471 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1472 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1473 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1474 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1475 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1476 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1477 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1478 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1479 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1480 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1481 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1482 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1483 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1484 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1485 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1486 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1487 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1488 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1489 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1490 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1491 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1492 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1493 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1494 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1495 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1496 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1497 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1498 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1499 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1500 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1501 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1502 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1503 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1504 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1505 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1506 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001507 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001508 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1509 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1510 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1511 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1512 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1513 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1514 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1515 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001516 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1517 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1518 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1519 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1520 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1521 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1522 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1523 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001524 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1525 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1526 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1527 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1528 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1529 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1530 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1531 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1532 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1533 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1534 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1535 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001536 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001537 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1538 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1539 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1540 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1541 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1542 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1543 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1544 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1545 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1546 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1547 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1548 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1549 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1550 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1551 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1552 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1553 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1554 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1555 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1556 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1557 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1558 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1559 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1560 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001561 "src/math/exp-avx512f-p5-scalef.c",
1562 "src/math/exp-avx512f-p5.c",
1563 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001564 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001565 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001566]
1567
1568AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001569 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001570 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001571 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1572 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1573 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1574 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1575 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1576 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1577 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1578 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1579 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1580 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581]
1582
1583AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001584 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001585 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1586 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001587 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001588 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1589 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001590 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1591 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1592 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001593 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001594 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1595 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1596 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001597 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001598 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1599 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001600 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001601 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1602 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1603 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1604 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1605 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1606 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1607 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1608 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1609 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1610 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1611 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1612 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1613 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1614 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1615 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1616 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1617 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1618 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1619 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1620 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1621 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001622 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001623 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1624 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1625 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1626 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1627 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1628 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1629 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1630 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1631 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1632 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1633 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1634 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1635 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1636 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1637 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1638 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1639 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1640 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1641 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1642 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1643 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1644 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1645 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1646 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1647 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1648 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1649 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1650 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1651 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1652 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1653 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1654 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1655 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1656 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1657 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1658 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1659 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1660 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001661]
1662
Marat Dukhan1b354632020-03-23 12:50:22 -07001663INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001664 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001665 "src/xnnpack/argmaxpool.h",
1666 "src/xnnpack/avgpool.h",
1667 "src/xnnpack/clamp.h",
1668 "src/xnnpack/common.h",
1669 "src/xnnpack/conv.h",
1670 "src/xnnpack/dwconv.h",
1671 "src/xnnpack/gavgpool.h",
1672 "src/xnnpack/gemm.h",
1673 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001674 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001675 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001676 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001677 "src/xnnpack/lut.h",
1678 "src/xnnpack/math.h",
1679 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001680 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681 "src/xnnpack/packx.h",
1682 "src/xnnpack/pad.h",
1683 "src/xnnpack/params.h",
1684 "src/xnnpack/pavgpool.h",
1685 "src/xnnpack/ppmm.h",
1686 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001687 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001688 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001689 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001690 "src/xnnpack/rmax.h",
1691 "src/xnnpack/scalar-utils.h",
1692 "src/xnnpack/spmm.h",
1693 "src/xnnpack/unpool.h",
1694 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001695 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001696 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001697 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001698 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001699 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001700 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001701 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001702]
1703
1704INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001705 "include/xnnpack.h",
1706 "src/xnnpack/allocator.h",
1707 "src/xnnpack/compute.h",
1708 "src/xnnpack/im2col.h",
1709 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001710 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001711 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 "src/xnnpack/operator.h",
1713 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001714 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001715 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001716 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001717 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001718]
1719
Marat Dukhan1b354632020-03-23 12:50:22 -07001720ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001721 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001722]
1723
Marat Dukhan1b354632020-03-23 12:50:22 -07001724MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001725 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001726 "include/xnnpack.h",
1727]
1728
Marat Dukhan1b354632020-03-23 12:50:22 -07001729MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001731 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001732 "src/xnnpack/requantization.h",
1733 "include/xnnpack.h",
1734]
1735
1736OPERATOR_TEST_PARAMS_HDRS = [
1737 "src/xnnpack/params.h",
1738 "src/xnnpack/common.h",
1739]
1740
1741WEIGHTS_PACK_HDRS = [
1742 "src/xnnpack/pack.h",
1743 "src/xnnpack/operator.h",
1744 "src/xnnpack/compute.h",
1745]
1746
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001747LOGGING_COPTS = select({
1748 # No logging in optimized mode
1749 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1750 # Full logging in debug mode
1751 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1752 # Error-only logging in default (fastbuild) mode
1753 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1754})
1755
1756LOGGING_HDRS = [
1757 "src/xnnpack/log.h",
1758]
1759
Marat Dukhan08c4a432019-10-03 09:29:21 -07001760xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001761 name = "tables",
1762 srcs = TABLE_SRCS,
1763 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001764 gcc_copts = xnnpack_gcc_std_copts(),
1765 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001766)
1767
1768xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 name = "scalar_ukernels",
1770 srcs = SCALAR_UKERNELS,
1771 hdrs = INTERNAL_HDRS,
1772 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001773 gcc_copts = xnnpack_gcc_std_copts(),
1774 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001775 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001776 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 "@FP16",
1778 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001779 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001780 ],
1781)
1782
1783xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001784 name = "wasm_ukernels",
1785 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001786 gcc_copts = xnnpack_gcc_std_copts(),
1787 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001788 wasm_srcs = WASM_UKERNELS,
1789 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001790 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001791 "@FP16",
1792 "@FXdiv",
1793 "@pthreadpool",
1794 ],
1795)
1796
1797xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001798 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001799 hdrs = INTERNAL_HDRS,
1800 aarch32_copts = [
1801 "-marm",
1802 "-mfpu=neon",
1803 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001804 gcc_copts = xnnpack_gcc_std_copts(),
1805 gcc_x86_copts = ["-msse2"],
1806 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807 optimized_copts = [
1808 "-O3",
1809 "-ffast-math",
1810 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001811 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001812 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001813 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 "@FP16",
1815 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001816 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001817 ],
1818)
1819
1820xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001821 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001822 hdrs = INTERNAL_HDRS,
1823 aarch32_copts = [
1824 "-marm",
1825 "-mfpu=neon",
1826 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001827 gcc_copts = xnnpack_gcc_std_copts(),
1828 gcc_x86_copts = ["-msse2"],
1829 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001830 optimized_copts = [
1831 "-O3",
1832 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001833 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001834 deps = [
1835 ":tables",
1836 "@FP16",
1837 "@psimd",
1838 "@pthreadpool",
1839 ],
1840)
1841
1842xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001843 name = "neon_ukernels",
1844 hdrs = INTERNAL_HDRS,
1845 aarch32_copts = [
1846 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001847 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848 "-mfpu=neon",
1849 ],
1850 aarch32_srcs = NEON_UKERNELS,
1851 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001852 gcc_copts = xnnpack_gcc_std_copts(),
1853 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001854 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001855 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001856 "@FP16",
1857 "@pthreadpool",
1858 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001859)
1860
1861xnnpack_cc_library(
1862 name = "neonfma_ukernels",
1863 hdrs = INTERNAL_HDRS,
1864 aarch32_copts = [
1865 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001866 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 "-mfpu=neon-vfpv4",
1868 ],
1869 aarch32_srcs = NEONFMA_UKERNELS,
1870 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001871 gcc_copts = xnnpack_gcc_std_copts(),
1872 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001873 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001874 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001875 "@FP16",
1876 "@pthreadpool",
1877 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001878)
1879
1880xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001881 name = "neonv8_ukernels",
1882 hdrs = INTERNAL_HDRS,
1883 aarch32_copts = [
1884 "-marm",
1885 "-march=armv8-a",
1886 "-mfpu=neon-fp-armv8",
1887 ],
1888 aarch32_srcs = NEONV8_UKERNELS,
1889 aarch64_srcs = NEONV8_UKERNELS,
1890 gcc_copts = xnnpack_gcc_std_copts(),
1891 msvc_copts = xnnpack_msvc_std_copts(),
1892 deps = [
1893 ":tables",
1894 "@FP16",
1895 "@pthreadpool",
1896 ],
1897)
1898
1899xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900 name = "neonfp16arith_ukernels",
1901 hdrs = INTERNAL_HDRS,
1902 aarch64_copts = ["-march=armv8.2-a+fp16"],
1903 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001904 gcc_copts = xnnpack_gcc_std_copts(),
1905 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001906 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001907 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001908 "@FP16",
1909 "@pthreadpool",
1910 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001911)
1912
1913xnnpack_cc_library(
1914 name = "sse2_ukernels",
1915 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001916 gcc_copts = xnnpack_gcc_std_copts(),
1917 gcc_x86_copts = ["-msse2"],
1918 msvc_copts = xnnpack_msvc_std_copts(),
1919 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001920 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001921 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001922 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001923 "@FP16",
1924 "@pthreadpool",
1925 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001926)
1927
1928xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001929 name = "ssse3_ukernels",
1930 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001931 gcc_copts = xnnpack_gcc_std_copts(),
1932 gcc_x86_copts = ["-mssse3"],
1933 msvc_copts = xnnpack_msvc_std_copts(),
1934 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001935 x86_srcs = SSSE3_UKERNELS,
1936 deps = [
1937 ":tables",
1938 "@FP16",
1939 "@pthreadpool",
1940 ],
1941)
1942
1943xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001944 name = "sse41_ukernels",
1945 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001946 gcc_copts = xnnpack_gcc_std_copts(),
1947 gcc_x86_copts = ["-msse4.1"],
1948 msvc_copts = xnnpack_msvc_std_copts(),
1949 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001950 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001951 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001952 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001953 "@FP16",
1954 "@pthreadpool",
1955 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001956)
1957
1958xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001959 name = "avx_ukernels",
1960 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001961 gcc_copts = xnnpack_gcc_std_copts(),
1962 gcc_x86_copts = ["-mavx"],
1963 msvc_copts = xnnpack_msvc_std_copts(),
1964 msvc_x86_32_copts = ["/arch:AVX"],
1965 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001966 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001967 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001968 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001969 "@FP16",
1970 "@pthreadpool",
1971 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001972)
1973
1974xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001975 name = "fma3_ukernels",
1976 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001977 gcc_copts = xnnpack_gcc_std_copts(),
1978 gcc_x86_copts = ["-mfma"],
1979 msvc_copts = xnnpack_msvc_std_copts(),
1980 msvc_x86_32_copts = ["/arch:AVX"],
1981 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001982 x86_srcs = FMA3_UKERNELS,
1983 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001984 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001985 "@FP16",
1986 "@pthreadpool",
1987 ],
1988)
1989
1990xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001991 name = "avx2_ukernels",
1992 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001993 gcc_copts = xnnpack_gcc_std_copts(),
1994 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001995 "-mfma",
1996 "-mavx2",
1997 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001998 msvc_copts = xnnpack_msvc_std_copts(),
1999 msvc_x86_32_copts = ["/arch:AVX2"],
2000 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002001 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002002 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002003 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002004 "@FP16",
2005 "@pthreadpool",
2006 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002007)
2008
2009xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002010 name = "avx512f_ukernels",
2011 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002012 gcc_copts = xnnpack_gcc_std_copts(),
2013 gcc_x86_copts = ["-mavx512f"],
2014 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2015 msvc_copts = xnnpack_msvc_std_copts(),
2016 msvc_x86_32_copts = ["/arch:AVX512"],
2017 msvc_x86_64_copts = ["/arch:AVX512"],
2018 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002019 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002020 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002021 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002022 "@FP16",
2023 "@pthreadpool",
2024 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025)
2026
2027xnnpack_cc_library(
2028 name = "asm_ukernels",
2029 hdrs = ["src/xnnpack/assembly.h"],
2030 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002031 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002032 aarch64_srcs = AARCH64_ASM_UKERNELS,
2033)
2034
2035xnnpack_aggregate_library(
2036 name = "ukernels",
2037 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002038 ":neon_ukernels",
2039 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002040 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002041 ":asm_ukernels",
2042 ],
2043 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002044 ":neon_ukernels",
2045 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002046 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002047 ":neonfp16arith_ukernels",
2048 ":asm_ukernels",
2049 ],
2050 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002051 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002052 ":psimd_fastmath_ukernels",
2053 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002055 wasm_deps = [":wasm_ukernels"],
2056 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002058 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002059 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002060 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002062 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002063 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002064 ":avx512f_ukernels",
2065 ],
2066)
2067
2068xnnpack_cc_library(
2069 name = "im2col",
2070 srcs = ["src/im2col.c"],
2071 hdrs = [
2072 "src/xnnpack/common.h",
2073 "src/xnnpack/im2col.h",
2074 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002075 gcc_copts = xnnpack_gcc_std_copts(),
2076 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002077)
2078
2079xnnpack_cc_library(
2080 name = "indirection",
2081 srcs = ["src/indirection.c"],
2082 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002083 gcc_copts = xnnpack_gcc_std_copts(),
2084 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085 deps = [
2086 "@FP16",
2087 "@FXdiv",
2088 "@pthreadpool",
2089 ],
2090)
2091
2092xnnpack_cc_library(
2093 name = "operator_run",
2094 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002095 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002096 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002097 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2098 "//conditions:default": [],
2099 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002100 gcc_copts = xnnpack_gcc_std_copts(),
2101 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002102 deps = [
2103 "@FP16",
2104 "@FXdiv",
2105 "@clog",
2106 "@pthreadpool",
2107 ],
2108)
2109
Chao Mei6ddfc602020-05-13 22:29:36 -07002110xnnpack_cc_library(
2111 name = "memory_planner",
2112 srcs = ["src/memory-planner.c"],
2113 hdrs = INTERNAL_HDRS,
2114 defines = select({
2115 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2116 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2117 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2118 }),
2119 gcc_copts = xnnpack_gcc_std_copts(),
2120 msvc_copts = xnnpack_msvc_std_copts(),
2121 deps = [
2122 "@pthreadpool",
2123 ],
2124)
2125
Marat Dukhan08c4a432019-10-03 09:29:21 -07002126cc_library(
2127 name = "enable_assembly",
2128 defines = select({
2129 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2130 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002131 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002132 }),
2133)
2134
Marat Dukhancf056b22019-10-07 10:26:29 -07002135xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002136 name = "operators",
2137 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002138 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002139 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002140 ],
2141 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002142 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002143 "-Isrc",
2144 "-Iinclude",
2145 ] + select({
2146 ":debug_build": [],
2147 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002148 }) + select({
2149 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2150 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002151 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002152 gcc_copts = xnnpack_gcc_std_copts(),
2153 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002155 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002156 "@FP16",
2157 "@FXdiv",
2158 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002160 ],
2161)
2162
Marat Dukhan10a38082020-04-17 03:58:35 -07002163xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002164 name = "XNNPACK",
2165 srcs = [
2166 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002167 "src/runtime.c",
2168 "src/subgraph.c",
2169 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002170 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002171 hdrs = ["include/xnnpack.h"],
2172 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002173 "-Isrc",
2174 "-Iinclude",
2175 ] + select({
2176 ":debug_build": [],
2177 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002178 }) + select({
2179 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2180 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002181 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002182 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002183 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002184 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002185 visibility = xnnpack_visibility(),
2186 deps = [
2187 ":enable_assembly",
Chao Mei6ddfc602020-05-13 22:29:36 -07002188 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002189 ":ukernels",
2190 ":operator_run",
2191 ":operators",
2192 "@clog",
2193 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002194 ] + select({
2195 ":emscripten": [],
2196 "//conditions:default": ["@cpuinfo"],
2197 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002198)
2199
Marat Dukhan10a38082020-04-17 03:58:35 -07002200xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002201 name = "xnnpack_operators_nhwc_f32",
2202 srcs = [
2203 "src/init.c",
2204 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002205 hdrs = ["include/xnnpack.h"],
2206 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002207 "-Isrc",
2208 "-Iinclude",
2209 ] + select({
2210 ":debug_build": [],
2211 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002212 }) + select({
2213 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2214 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002215 }),
2216 defines = [
2217 "XNN_NO_Q8_OPERATORS",
2218 "XNN_NO_U8_OPERATORS",
2219 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002220 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002221 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002222 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002223 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002224 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002225 visibility = xnnpack_visibility(),
2226 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002227 ":enable_assembly",
2228 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229 ":operator_run",
2230 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002231 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002232 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002233 ] + select({
2234 ":emscripten": [],
2235 "//conditions:default": ["@cpuinfo"],
2236 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237)
2238
Marat Dukhancf056b22019-10-07 10:26:29 -07002239xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002240 name = "bench_utils",
2241 srcs = ["bench/utils.cc"],
2242 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002243 deps = [
2244 "@com_google_benchmark//:benchmark",
2245 "@cpuinfo",
2246 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002247)
2248
Frank Barchard7e955972019-10-11 10:34:25 -07002249######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002250
2251xnnpack_benchmark(
2252 name = "q8_gemm_bench",
2253 srcs = [
2254 "bench/gemm.h",
2255 "bench/q8-gemm.cc",
2256 "src/xnnpack/AlignedAllocator.h",
2257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002258 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002259 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2260)
2261
2262xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002263 name = "f16_igemm_bench",
2264 srcs = [
2265 "bench/f16-igemm.cc",
2266 "bench/conv.h",
2267 "bench/google/conv.h",
2268 "src/xnnpack/AlignedAllocator.h",
2269 ] + MICROKERNEL_BENCHMARK_HDRS,
2270 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2271)
2272
2273xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002274 name = "f16_gemm_bench",
2275 srcs = [
2276 "bench/f16-gemm.cc",
2277 "bench/gemm.h",
2278 "src/xnnpack/AlignedAllocator.h",
2279 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002280 deps = MICROKERNEL_BENCHMARK_DEPS,
2281)
2282
2283xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002284 name = "f16_spmm_bench",
2285 srcs = [
2286 "bench/f16-spmm.cc",
2287 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002288 "src/xnnpack/AlignedAllocator.h",
2289 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002290 deps = MICROKERNEL_BENCHMARK_DEPS,
2291)
2292
2293xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002294 name = "f32_igemm_bench",
2295 srcs = [
2296 "bench/f32-igemm.cc",
2297 "bench/conv.h",
2298 "src/xnnpack/AlignedAllocator.h",
2299 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002300 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002301)
2302
2303xnnpack_benchmark(
2304 name = "f32_conv_hwc_bench",
2305 srcs = [
2306 "bench/f32-conv-hwc.cc",
2307 "bench/dconv.h",
2308 "src/xnnpack/AlignedAllocator.h",
2309 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002310 deps = MICROKERNEL_BENCHMARK_DEPS,
2311)
2312
2313xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002314 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07002315 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002316 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07002317 "bench/dconv.h",
2318 "src/xnnpack/AlignedAllocator.h",
2319 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002320 deps = MICROKERNEL_BENCHMARK_DEPS,
2321)
2322
2323xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002324 name = "f32_dwconv_bench",
2325 srcs = [
2326 "bench/f32-dwconv.cc",
2327 "bench/dwconv.h",
2328 "src/xnnpack/AlignedAllocator.h",
2329 ] + MICROKERNEL_BENCHMARK_HDRS,
2330 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2331)
2332
2333xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002334 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002335 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002336 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337 "bench/dwconv.h",
2338 "src/xnnpack/AlignedAllocator.h",
2339 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2340 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2341)
2342
2343xnnpack_benchmark(
2344 name = "f32_gemm_bench",
2345 srcs = [
2346 "bench/f32-gemm.cc",
2347 "bench/gemm.h",
2348 "src/xnnpack/AlignedAllocator.h",
2349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002350 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002351 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352)
2353
2354xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002355 name = "f32_raddexpminusmax_bench",
2356 srcs = [
2357 "bench/f32-raddexpminusmax.cc",
2358 "src/xnnpack/AlignedAllocator.h",
2359 ] + MICROKERNEL_BENCHMARK_HDRS,
2360 deps = MICROKERNEL_BENCHMARK_DEPS,
2361)
2362
2363xnnpack_benchmark(
2364 name = "f32_raddextexp_bench",
2365 srcs = [
2366 "bench/f32-raddextexp.cc",
2367 "src/xnnpack/AlignedAllocator.h",
2368 ] + MICROKERNEL_BENCHMARK_HDRS,
2369 deps = MICROKERNEL_BENCHMARK_DEPS,
2370)
2371
2372xnnpack_benchmark(
2373 name = "f32_raddstoreexpminusmax_bench",
2374 srcs = [
2375 "bench/f32-raddstoreexpminusmax.cc",
2376 "src/xnnpack/AlignedAllocator.h",
2377 ] + MICROKERNEL_BENCHMARK_HDRS,
2378 deps = MICROKERNEL_BENCHMARK_DEPS,
2379)
2380
2381xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002382 name = "f32_rmax_bench",
2383 srcs = [
2384 "bench/f32-rmax.cc",
2385 "src/xnnpack/AlignedAllocator.h",
2386 ] + MICROKERNEL_BENCHMARK_HDRS,
2387 deps = MICROKERNEL_BENCHMARK_DEPS,
2388)
2389
2390xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002391 name = "f32_sigmoid_bench",
2392 srcs = [
2393 "bench/f32-sigmoid.cc",
2394 "src/xnnpack/AlignedAllocator.h",
2395 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002396 deps = MICROKERNEL_BENCHMARK_DEPS,
2397)
2398
2399xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002400 name = "f32_spmm_bench",
2401 srcs = [
2402 "bench/f32-spmm.cc",
2403 "bench/gemm.h",
2404 "src/xnnpack/AlignedAllocator.h",
2405 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002406 deps = MICROKERNEL_BENCHMARK_DEPS,
2407)
2408
2409xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002410 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002411 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002412 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002413 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002414 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002415 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002416)
2417
2418xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002419 name = "f32_vscaleexpminusmax_bench",
2420 srcs = [
2421 "bench/f32-vscaleexpminusmax.cc",
2422 "src/xnnpack/AlignedAllocator.h",
2423 ] + MICROKERNEL_BENCHMARK_HDRS,
2424 deps = MICROKERNEL_BENCHMARK_DEPS,
2425)
2426
2427xnnpack_benchmark(
2428 name = "f32_vscaleextexp_bench",
2429 srcs = [
2430 "bench/f32-vscaleextexp.cc",
2431 "src/xnnpack/AlignedAllocator.h",
2432 ] + MICROKERNEL_BENCHMARK_HDRS,
2433 deps = MICROKERNEL_BENCHMARK_DEPS,
2434)
2435
2436xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002437 name = "f32_im2col_gemm_bench",
2438 srcs = [
2439 "bench/f32-im2col-gemm.cc",
2440 "bench/conv.h",
2441 "src/xnnpack/AlignedAllocator.h",
2442 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2443 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2444)
2445
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002446xnnpack_benchmark(
2447 name = "requantization_bench",
2448 srcs = [
2449 "bench/requantization.cc",
2450 "src/xnnpack/requantization-stubs.h",
2451 "src/xnnpack/AlignedAllocator.h",
2452 ] + MICROKERNEL_BENCHMARK_HDRS,
2453 deps = MICROKERNEL_BENCHMARK_DEPS,
2454)
2455
Marat Dukhanffbf96a2020-05-14 02:59:08 -07002456xnnpack_benchmark(
2457 name = "rounding_bench",
2458 srcs = [
2459 "bench/rounding.cc",
2460 "src/xnnpack/math-stubs.h",
2461 "src/xnnpack/AlignedAllocator.h",
2462 ] + MICROKERNEL_BENCHMARK_HDRS,
2463 deps = MICROKERNEL_BENCHMARK_DEPS,
2464)
2465
Marat Dukhan08c4a432019-10-03 09:29:21 -07002466########################### Benchmarks for operators ###########################
2467
2468xnnpack_benchmark(
2469 name = "add_bench",
2470 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002471 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002472)
2473
2474xnnpack_benchmark(
2475 name = "average_pooling_bench",
2476 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002477 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002478 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002479 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002480)
2481
2482xnnpack_benchmark(
2483 name = "channel_shuffle_bench",
2484 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002485 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002486)
2487
2488xnnpack_benchmark(
2489 name = "convolution_bench",
2490 srcs = ["bench/convolution.cc"],
2491 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002492 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002493 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002494)
2495
2496xnnpack_benchmark(
2497 name = "deconvolution_bench",
2498 srcs = ["bench/deconvolution.cc"],
2499 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002500 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002501 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002502)
2503
2504xnnpack_benchmark(
2505 name = "global_average_pooling_bench",
2506 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002507 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002508)
2509
2510xnnpack_benchmark(
2511 name = "max_pooling_bench",
2512 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002513 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002514)
2515
2516xnnpack_benchmark(
2517 name = "sigmoid_bench",
2518 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002519 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002520 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002521 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002522)
2523
2524xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002525 name = "prelu_bench",
2526 srcs = ["bench/prelu.cc"],
2527 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002528 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002529 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002530)
2531
2532xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002533 name = "softmax_bench",
2534 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002535 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002536 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002537 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002538)
2539
Marat Dukhanc068bb62019-10-04 13:24:39 -07002540############################# End-to-end benchmarks ############################
2541
2542cc_library(
2543 name = "mobilenet_v1",
2544 srcs = ["models/mobilenet-v1.cc"],
2545 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002546 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002547 linkstatic = True,
2548 deps = [
2549 ":XNNPACK",
2550 "@pthreadpool",
2551 ],
2552)
2553
2554cc_library(
2555 name = "mobilenet_v2",
2556 srcs = ["models/mobilenet-v2.cc"],
2557 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002558 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002559 linkstatic = True,
2560 deps = [
2561 ":XNNPACK",
2562 "@pthreadpool",
2563 ],
2564)
2565
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002566cc_library(
2567 name = "mobilenet_v3_large",
2568 srcs = ["models/mobilenet-v3-large.cc"],
2569 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002570 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002571 linkstatic = True,
2572 deps = [
2573 ":XNNPACK",
2574 "@pthreadpool",
2575 ],
2576)
2577
2578cc_library(
2579 name = "mobilenet_v3_small",
2580 srcs = ["models/mobilenet-v3-small.cc"],
2581 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002582 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002583 linkstatic = True,
2584 deps = [
2585 ":XNNPACK",
2586 "@pthreadpool",
2587 ],
2588)
2589
Marat Dukhanc068bb62019-10-04 13:24:39 -07002590xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002591 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002592 srcs = [
2593 "bench/f32-dwconv-e2e.cc",
2594 "bench/end2end.h",
2595 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002596 deps = MICROKERNEL_BENCHMARK_DEPS + [
2597 ":XNNPACK",
2598 ":mobilenet_v1",
2599 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002600 ":mobilenet_v3_large",
2601 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002602 ],
2603)
2604
2605xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002606 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002607 srcs = [
2608 "bench/f32-gemm-e2e.cc",
2609 "bench/end2end.h",
2610 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002611 deps = MICROKERNEL_BENCHMARK_DEPS + [
2612 ":XNNPACK",
2613 ":mobilenet_v1",
2614 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002615 ":mobilenet_v3_large",
2616 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002617 ],
2618)
2619
2620xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002621 name = "end2end_bench",
2622 srcs = ["bench/end2end.cc"],
2623 deps = [
2624 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002625 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002626 ":mobilenet_v1",
2627 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002628 ":mobilenet_v3_large",
2629 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002630 "@pthreadpool",
2631 ],
2632)
2633
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002634#################### Accuracy evaluation for math functions ####################
2635
2636xnnpack_benchmark(
2637 name = "f32_exp_eval",
2638 srcs = [
2639 "eval/f32-exp.cc",
2640 "src/xnnpack/AlignedAllocator.h",
2641 ] + ACCURACY_EVAL_HDRS,
2642 deps = ACCURACY_EVAL_DEPS,
2643)
2644
Marat Dukhan515c9772019-10-17 18:07:57 -07002645xnnpack_benchmark(
2646 name = "f32_expminus_eval",
2647 srcs = [
2648 "eval/f32-expminus.cc",
2649 "src/xnnpack/AlignedAllocator.h",
2650 ] + ACCURACY_EVAL_HDRS,
2651 deps = ACCURACY_EVAL_DEPS,
2652)
2653
Marat Dukhan98ba4412019-10-23 02:14:28 -07002654xnnpack_benchmark(
2655 name = "f32_extexp_eval",
2656 srcs = [
2657 "eval/f32-extexp.cc",
2658 "src/xnnpack/AlignedAllocator.h",
2659 ] + ACCURACY_EVAL_HDRS,
2660 deps = ACCURACY_EVAL_DEPS,
2661)
2662
Marat Dukhan8853b822020-05-07 12:19:01 -07002663xnnpack_unit_test(
2664 name = "f32_roundne_eval",
2665 srcs = [
2666 "eval/f32-roundne.cc",
2667 "src/xnnpack/AlignedAllocator.h",
2668 "src/xnnpack/math-stubs.h",
2669 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002670 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002671 deps = MICROKERNEL_TEST_DEPS,
2672)
2673
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002674xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002675 name = "f32_roundd_eval",
2676 srcs = [
2677 "eval/f32-roundd.cc",
2678 "src/xnnpack/AlignedAllocator.h",
2679 "src/xnnpack/math-stubs.h",
2680 ] + MICROKERNEL_TEST_HDRS,
2681 automatic = False,
2682 deps = MICROKERNEL_TEST_DEPS,
2683)
2684
2685xnnpack_unit_test(
2686 name = "f32_roundu_eval",
2687 srcs = [
2688 "eval/f32-roundu.cc",
2689 "src/xnnpack/AlignedAllocator.h",
2690 "src/xnnpack/math-stubs.h",
2691 ] + MICROKERNEL_TEST_HDRS,
2692 automatic = False,
2693 deps = MICROKERNEL_TEST_DEPS,
2694)
2695
2696xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002697 name = "f32_roundz_eval",
2698 srcs = [
2699 "eval/f32-roundz.cc",
2700 "src/xnnpack/AlignedAllocator.h",
2701 "src/xnnpack/math-stubs.h",
2702 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002703 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002704 deps = MICROKERNEL_TEST_DEPS,
2705)
2706
Marat Dukhan346a9e52019-11-15 09:06:30 -08002707xnnpack_benchmark(
2708 name = "f32_sigmoid_eval",
2709 srcs = [
2710 "eval/f32-sigmoid.cc",
2711 "src/xnnpack/AlignedAllocator.h",
2712 ] + ACCURACY_EVAL_HDRS,
2713 deps = ACCURACY_EVAL_DEPS,
2714)
2715
Marat Dukhan08c4a432019-10-03 09:29:21 -07002716######################### Unit tests for micro-kernels #########################
2717
2718xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002719 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002720 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002721 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002722 "test/gemm-microkernel-tester.h",
2723 "src/xnnpack/AlignedAllocator.h",
2724 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2725 deps = MICROKERNEL_TEST_DEPS,
2726)
2727
2728xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002729 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002730 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002731 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002732 "test/spmm-microkernel-tester.h",
2733 "src/xnnpack/AlignedAllocator.h",
2734 ] + MICROKERNEL_TEST_HDRS,
2735 deps = MICROKERNEL_TEST_DEPS,
2736)
2737
2738xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002739 name = "f32_argmaxpool_test",
2740 srcs = [
2741 "test/f32-argmaxpool.cc",
2742 "test/argmaxpool-microkernel-tester.h",
2743 "src/xnnpack/AlignedAllocator.h",
2744 ] + MICROKERNEL_TEST_HDRS,
2745 deps = MICROKERNEL_TEST_DEPS,
2746)
2747
2748xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002749 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002750 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002751 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002752 "test/avgpool-microkernel-tester.h",
2753 "src/xnnpack/AlignedAllocator.h",
2754 ] + MICROKERNEL_TEST_HDRS,
2755 deps = MICROKERNEL_TEST_DEPS,
2756)
2757
2758xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002759 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002760 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002761 "test/f32-ibilinear.cc",
2762 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002763 "src/xnnpack/AlignedAllocator.h",
2764 ] + MICROKERNEL_TEST_HDRS,
2765 deps = MICROKERNEL_TEST_DEPS,
2766)
2767
2768xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002769 name = "f16_clamp_test",
2770 srcs = [
2771 "test/f16-clamp.cc",
2772 "test/clamp-microkernel-tester.h",
2773 ] + MICROKERNEL_TEST_HDRS,
2774 deps = MICROKERNEL_TEST_DEPS,
2775)
2776
2777xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002778 name = "f32_clamp_test",
2779 srcs = [
2780 "test/f32-clamp.cc",
2781 "test/clamp-microkernel-tester.h",
2782 ] + MICROKERNEL_TEST_HDRS,
2783 deps = MICROKERNEL_TEST_DEPS,
2784)
2785
2786xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002787 name = "f32_igemm_test",
2788 srcs = [
2789 "test/f32-igemm.cc",
2790 "test/gemm-microkernel-tester.h",
2791 "src/xnnpack/AlignedAllocator.h",
2792 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2793 deps = MICROKERNEL_TEST_DEPS,
2794)
2795
2796xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002797 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002800 "test/gemm-microkernel-tester.h",
2801 "src/xnnpack/AlignedAllocator.h",
2802 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002803 # Work-around for "too many sections" error
2804 mingw_copts = ["-Wa,-mbig-obj"],
2805 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806 deps = MICROKERNEL_TEST_DEPS,
2807)
2808
2809xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002810 name = "f16_igemm_minmax_test",
2811 srcs = [
2812 "test/f16-igemm-minmax.cc",
2813 "test/gemm-microkernel-tester.h",
2814 "src/xnnpack/AlignedAllocator.h",
2815 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2816 deps = MICROKERNEL_TEST_DEPS,
2817)
2818
2819xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820 name = "f32_conv_hwc_test",
2821 srcs = [
2822 "test/f32-conv-hwc.cc",
2823 "test/conv-hwc-microkernel-tester.h",
2824 "src/xnnpack/AlignedAllocator.h",
2825 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2826 deps = MICROKERNEL_TEST_DEPS,
2827)
2828
2829xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002830 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002831 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002832 "test/f32-conv-hwc2chw.cc",
2833 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002834 "src/xnnpack/AlignedAllocator.h",
2835 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2836 deps = MICROKERNEL_TEST_DEPS,
2837)
2838
2839xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002840 name = "f32_dwconv_test",
2841 srcs = [
2842 "test/f32-dwconv.cc",
2843 "test/dwconv-microkernel-tester.h",
2844 "src/xnnpack/AlignedAllocator.h",
2845 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2846 deps = MICROKERNEL_TEST_DEPS,
2847)
2848
2849xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002850 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002851 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002852 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002853 "test/dwconv-microkernel-tester.h",
2854 "src/xnnpack/AlignedAllocator.h",
2855 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2856 deps = MICROKERNEL_TEST_DEPS,
2857)
2858
2859xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002860 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002861 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002862 "test/f32-dwconv-chw.cc",
2863 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002864 "src/xnnpack/AlignedAllocator.h",
2865 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2866 deps = MICROKERNEL_TEST_DEPS,
2867)
2868
2869xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002870 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002872 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002873 "test/gavgpool-microkernel-tester.h",
2874 "src/xnnpack/AlignedAllocator.h",
2875 ] + MICROKERNEL_TEST_HDRS,
2876 deps = MICROKERNEL_TEST_DEPS,
2877)
2878
2879xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002880 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002882 "test/f32-gavgpool-cw.cc",
2883 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002884 "src/xnnpack/AlignedAllocator.h",
2885 ] + MICROKERNEL_TEST_HDRS,
2886 deps = MICROKERNEL_TEST_DEPS,
2887)
2888
2889xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002890 name = "f32_gemm_test",
2891 srcs = [
2892 "test/f32-gemm.cc",
2893 "test/gemm-microkernel-tester.h",
2894 "src/xnnpack/AlignedAllocator.h",
2895 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2896 deps = MICROKERNEL_TEST_DEPS,
2897)
2898
2899xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002900 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002902 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 "test/gemm-microkernel-tester.h",
2904 "src/xnnpack/AlignedAllocator.h",
2905 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2906 deps = MICROKERNEL_TEST_DEPS,
2907)
2908
2909xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002910 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002912 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913 "test/gemm-microkernel-tester.h",
2914 "src/xnnpack/AlignedAllocator.h",
2915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2916 deps = MICROKERNEL_TEST_DEPS,
2917)
2918
2919xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002920 name = "f16_hswish_test",
2921 srcs = [
2922 "test/f16-hswish.cc",
2923 "test/hswish-microkernel-tester.h",
2924 ] + MICROKERNEL_TEST_HDRS,
2925 deps = MICROKERNEL_TEST_DEPS,
2926)
2927
2928xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002929 name = "f32_hswish_test",
2930 srcs = [
2931 "test/f32-hswish.cc",
2932 "test/hswish-microkernel-tester.h",
2933 ] + MICROKERNEL_TEST_HDRS,
2934 deps = MICROKERNEL_TEST_DEPS,
2935)
2936
2937xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002938 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002939 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002940 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 "test/maxpool-microkernel-tester.h",
2942 ] + MICROKERNEL_TEST_HDRS,
2943 deps = MICROKERNEL_TEST_DEPS,
2944)
2945
2946xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002947 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002949 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002950 "test/avgpool-microkernel-tester.h",
2951 "src/xnnpack/AlignedAllocator.h",
2952 ] + MICROKERNEL_TEST_HDRS,
2953 deps = MICROKERNEL_TEST_DEPS,
2954)
2955
2956xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002957 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002959 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960 "test/gemm-microkernel-tester.h",
2961 "src/xnnpack/AlignedAllocator.h",
2962 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2963 deps = MICROKERNEL_TEST_DEPS,
2964)
2965
2966xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002967 name = "f16_prelu_test",
2968 srcs = [
2969 "test/f16-prelu.cc",
2970 "test/prelu-microkernel-tester.h",
2971 "src/xnnpack/AlignedAllocator.h",
2972 ] + MICROKERNEL_TEST_HDRS,
2973 deps = MICROKERNEL_TEST_DEPS,
2974)
2975
2976xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 name = "f32_prelu_test",
2978 srcs = [
2979 "test/f32-prelu.cc",
2980 "test/prelu-microkernel-tester.h",
2981 "src/xnnpack/AlignedAllocator.h",
2982 ] + MICROKERNEL_TEST_HDRS,
2983 deps = MICROKERNEL_TEST_DEPS,
2984)
2985
2986xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002987 name = "f32_raddexpminusmax_test",
2988 srcs = [
2989 "test/f32-raddexpminusmax.cc",
2990 "test/raddexpminusmax-microkernel-tester.h",
2991 ] + MICROKERNEL_TEST_HDRS,
2992 deps = MICROKERNEL_TEST_DEPS,
2993)
2994
2995xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002996 name = "f32_raddextexp_test",
2997 srcs = [
2998 "test/f32-raddextexp.cc",
2999 "test/raddextexp-microkernel-tester.h",
3000 ] + MICROKERNEL_TEST_HDRS,
3001 deps = MICROKERNEL_TEST_DEPS,
3002)
3003
3004xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003005 name = "f32_raddstoreexpminusmax_test",
3006 srcs = [
3007 "test/f32-raddstoreexpminusmax.cc",
3008 "test/raddstoreexpminusmax-microkernel-tester.h",
3009 ] + MICROKERNEL_TEST_HDRS,
3010 deps = MICROKERNEL_TEST_DEPS,
3011)
3012
3013xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003014 name = "f32_rmax_test",
3015 srcs = [
3016 "test/f32-rmax.cc",
3017 "test/rmax-microkernel-tester.h",
3018 ] + MICROKERNEL_TEST_HDRS,
3019 deps = MICROKERNEL_TEST_DEPS,
3020)
3021
3022xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003023 name = "f32_sigmoid_test",
3024 srcs = [
3025 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003026 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003027 ] + MICROKERNEL_TEST_HDRS,
3028 deps = MICROKERNEL_TEST_DEPS,
3029)
3030
3031xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003032 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003033 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003034 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003035 "test/spmm-microkernel-tester.h",
3036 "src/xnnpack/AlignedAllocator.h",
3037 ] + MICROKERNEL_TEST_HDRS,
3038 deps = MICROKERNEL_TEST_DEPS,
3039)
3040
3041xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003042 name = "f16_vadd_minmax_test",
3043 srcs = [
3044 "test/f16-vadd-minmax.cc",
3045 "test/vbinary-microkernel-tester.h",
3046 ] + MICROKERNEL_TEST_HDRS,
3047 deps = MICROKERNEL_TEST_DEPS,
3048)
3049
3050xnnpack_unit_test(
3051 name = "f16_vaddc_minmax_test",
3052 srcs = [
3053 "test/f16-vaddc-minmax.cc",
3054 "test/vbinaryc-microkernel-tester.h",
3055 ] + MICROKERNEL_TEST_HDRS,
3056 deps = MICROKERNEL_TEST_DEPS,
3057)
3058
3059xnnpack_unit_test(
3060 name = "f16_vdiv_minmax_test",
3061 srcs = [
3062 "test/f16-vdiv-minmax.cc",
3063 "test/vbinary-microkernel-tester.h",
3064 ] + MICROKERNEL_TEST_HDRS,
3065 deps = MICROKERNEL_TEST_DEPS,
3066)
3067
3068xnnpack_unit_test(
3069 name = "f16_vdivc_minmax_test",
3070 srcs = [
3071 "test/f16-vdivc-minmax.cc",
3072 "test/vbinaryc-microkernel-tester.h",
3073 ] + MICROKERNEL_TEST_HDRS,
3074 deps = MICROKERNEL_TEST_DEPS,
3075)
3076
3077xnnpack_unit_test(
3078 name = "f16_vrdivc_minmax_test",
3079 srcs = [
3080 "test/f16-vrdivc-minmax.cc",
3081 "test/vbinaryc-microkernel-tester.h",
3082 ] + MICROKERNEL_TEST_HDRS,
3083 deps = MICROKERNEL_TEST_DEPS,
3084)
3085
3086xnnpack_unit_test(
3087 name = "f16_vmax_test",
3088 srcs = [
3089 "test/f16-vmax.cc",
3090 "test/vbinary-microkernel-tester.h",
3091 ] + MICROKERNEL_TEST_HDRS,
3092 deps = MICROKERNEL_TEST_DEPS,
3093)
3094
3095xnnpack_unit_test(
3096 name = "f16_vmaxc_test",
3097 srcs = [
3098 "test/f16-vmaxc.cc",
3099 "test/vbinaryc-microkernel-tester.h",
3100 ] + MICROKERNEL_TEST_HDRS,
3101 deps = MICROKERNEL_TEST_DEPS,
3102)
3103
3104xnnpack_unit_test(
3105 name = "f16_vmin_test",
3106 srcs = [
3107 "test/f16-vmin.cc",
3108 "test/vbinary-microkernel-tester.h",
3109 ] + MICROKERNEL_TEST_HDRS,
3110 deps = MICROKERNEL_TEST_DEPS,
3111)
3112
3113xnnpack_unit_test(
3114 name = "f16_vminc_test",
3115 srcs = [
3116 "test/f16-vminc.cc",
3117 "test/vbinaryc-microkernel-tester.h",
3118 ] + MICROKERNEL_TEST_HDRS,
3119 deps = MICROKERNEL_TEST_DEPS,
3120)
3121
3122xnnpack_unit_test(
3123 name = "f16_vmul_minmax_test",
3124 srcs = [
3125 "test/f16-vmul-minmax.cc",
3126 "test/vbinary-microkernel-tester.h",
3127 ] + MICROKERNEL_TEST_HDRS,
3128 deps = MICROKERNEL_TEST_DEPS,
3129)
3130
3131xnnpack_unit_test(
3132 name = "f16_vmulc_minmax_test",
3133 srcs = [
3134 "test/f16-vmulc-minmax.cc",
3135 "test/vbinaryc-microkernel-tester.h",
3136 ] + MICROKERNEL_TEST_HDRS,
3137 deps = MICROKERNEL_TEST_DEPS,
3138)
3139
3140xnnpack_unit_test(
3141 name = "f16_vsub_minmax_test",
3142 srcs = [
3143 "test/f16-vsub-minmax.cc",
3144 "test/vbinary-microkernel-tester.h",
3145 ] + MICROKERNEL_TEST_HDRS,
3146 deps = MICROKERNEL_TEST_DEPS,
3147)
3148
3149xnnpack_unit_test(
3150 name = "f16_vsubc_minmax_test",
3151 srcs = [
3152 "test/f16-vsubc-minmax.cc",
3153 "test/vbinaryc-microkernel-tester.h",
3154 ] + MICROKERNEL_TEST_HDRS,
3155 deps = MICROKERNEL_TEST_DEPS,
3156)
3157
3158xnnpack_unit_test(
3159 name = "f16_vrsubc_minmax_test",
3160 srcs = [
3161 "test/f16-vrsubc-minmax.cc",
3162 "test/vbinaryc-microkernel-tester.h",
3163 ] + MICROKERNEL_TEST_HDRS,
3164 deps = MICROKERNEL_TEST_DEPS,
3165)
3166
3167xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003168 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003170 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003171 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003172 ] + MICROKERNEL_TEST_HDRS,
3173 deps = MICROKERNEL_TEST_DEPS,
3174)
3175
3176xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003177 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003178 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003179 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003180 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181 ] + MICROKERNEL_TEST_HDRS,
3182 deps = MICROKERNEL_TEST_DEPS,
3183)
3184
3185xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003186 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003187 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003188 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003189 "test/vbinary-microkernel-tester.h",
3190 ] + MICROKERNEL_TEST_HDRS,
3191 deps = MICROKERNEL_TEST_DEPS,
3192)
3193
3194xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003195 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003196 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003197 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003198 "test/vbinaryc-microkernel-tester.h",
3199 ] + MICROKERNEL_TEST_HDRS,
3200 deps = MICROKERNEL_TEST_DEPS,
3201)
3202
3203xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003204 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003205 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003206 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003207 "test/vbinaryc-microkernel-tester.h",
3208 ] + MICROKERNEL_TEST_HDRS,
3209 deps = MICROKERNEL_TEST_DEPS,
3210)
3211
3212xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003213 name = "f32_vmax_test",
3214 srcs = [
3215 "test/f32-vmax.cc",
3216 "test/vbinary-microkernel-tester.h",
3217 ] + MICROKERNEL_TEST_HDRS,
3218 deps = MICROKERNEL_TEST_DEPS,
3219)
3220
3221xnnpack_unit_test(
3222 name = "f32_vmaxc_test",
3223 srcs = [
3224 "test/f32-vmaxc.cc",
3225 "test/vbinaryc-microkernel-tester.h",
3226 ] + MICROKERNEL_TEST_HDRS,
3227 deps = MICROKERNEL_TEST_DEPS,
3228)
3229
3230xnnpack_unit_test(
3231 name = "f32_vmin_test",
3232 srcs = [
3233 "test/f32-vmin.cc",
3234 "test/vbinary-microkernel-tester.h",
3235 ] + MICROKERNEL_TEST_HDRS,
3236 deps = MICROKERNEL_TEST_DEPS,
3237)
3238
3239xnnpack_unit_test(
3240 name = "f32_vminc_test",
3241 srcs = [
3242 "test/f32-vminc.cc",
3243 "test/vbinaryc-microkernel-tester.h",
3244 ] + MICROKERNEL_TEST_HDRS,
3245 deps = MICROKERNEL_TEST_DEPS,
3246)
3247
3248xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003249 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003251 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003252 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003253 ] + MICROKERNEL_TEST_HDRS,
3254 deps = MICROKERNEL_TEST_DEPS,
3255)
3256
3257xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003258 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003259 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003260 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003261 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 ] + MICROKERNEL_TEST_HDRS,
3263 deps = MICROKERNEL_TEST_DEPS,
3264)
3265
3266xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003267 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003269 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270 "test/vmulcaddc-microkernel-tester.h",
3271 "src/xnnpack/AlignedAllocator.h",
3272 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3273 deps = MICROKERNEL_TEST_DEPS,
3274)
3275
3276xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003277 name = "f32_vscale_test",
3278 srcs = [
3279 "test/f32-vscale.cc",
3280 "test/vscale-microkernel-tester.h",
3281 ] + MICROKERNEL_TEST_HDRS,
3282 deps = MICROKERNEL_TEST_DEPS,
3283)
3284
3285xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003286 name = "f32_vscaleexpminusmax_test",
3287 srcs = [
3288 "test/f32-vscaleexpminusmax.cc",
3289 "test/vscaleexpminusmax-microkernel-tester.h",
3290 ] + MICROKERNEL_TEST_HDRS,
3291 deps = MICROKERNEL_TEST_DEPS,
3292)
3293
3294xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003295 name = "f32_vscaleextexp_test",
3296 srcs = [
3297 "test/f32-vscaleextexp.cc",
3298 "test/vscaleextexp-microkernel-tester.h",
3299 ] + MICROKERNEL_TEST_HDRS,
3300 deps = MICROKERNEL_TEST_DEPS,
3301)
3302
3303xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003304 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003305 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003306 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003307 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003308 ] + MICROKERNEL_TEST_HDRS,
3309 deps = MICROKERNEL_TEST_DEPS,
3310)
3311
3312xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003313 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003314 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003315 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003316 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003317 ] + MICROKERNEL_TEST_HDRS,
3318 deps = MICROKERNEL_TEST_DEPS,
3319)
3320
3321xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003322 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003323 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003324 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003325 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003326 ] + MICROKERNEL_TEST_HDRS,
3327 deps = MICROKERNEL_TEST_DEPS,
3328)
3329
3330xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003331 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003332 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003333 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003334 "test/avgpool-microkernel-tester.h",
3335 "src/xnnpack/AlignedAllocator.h",
3336 ] + MICROKERNEL_TEST_HDRS,
3337 deps = MICROKERNEL_TEST_DEPS,
3338)
3339
3340xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003341 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003343 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344 "test/gemm-microkernel-tester.h",
3345 "src/xnnpack/AlignedAllocator.h",
3346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3347 deps = MICROKERNEL_TEST_DEPS,
3348)
3349
3350xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003351 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003352 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003353 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003354 "test/dwconv-microkernel-tester.h",
3355 "src/xnnpack/AlignedAllocator.h",
3356 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3357 deps = MICROKERNEL_TEST_DEPS,
3358)
3359
3360xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003361 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003363 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "test/gavgpool-microkernel-tester.h",
3365 "src/xnnpack/AlignedAllocator.h",
3366 ] + MICROKERNEL_TEST_HDRS,
3367 deps = MICROKERNEL_TEST_DEPS,
3368)
3369
3370xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003371 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003372 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003373 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003374 "test/gemm-microkernel-tester.h",
3375 "src/xnnpack/AlignedAllocator.h",
3376 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3377 deps = MICROKERNEL_TEST_DEPS,
3378)
3379
3380xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003381 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003382 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003383 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003384 "test/vadd-microkernel-tester.h",
3385 ] + MICROKERNEL_TEST_HDRS,
3386 deps = MICROKERNEL_TEST_DEPS,
3387)
3388
3389xnnpack_unit_test(
3390 name = "u8_clamp_test",
3391 srcs = [
3392 "test/u8-clamp.cc",
3393 "test/clamp-microkernel-tester.h",
3394 ] + MICROKERNEL_TEST_HDRS,
3395 deps = MICROKERNEL_TEST_DEPS,
3396)
3397
3398xnnpack_unit_test(
3399 name = "u8_lut32norm_test",
3400 srcs = [
3401 "test/u8-lut32norm.cc",
3402 "test/lut-norm-microkernel-tester.h",
3403 ] + MICROKERNEL_TEST_HDRS,
3404 deps = MICROKERNEL_TEST_DEPS,
3405)
3406
3407xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003408 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003410 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003411 "test/maxpool-microkernel-tester.h",
3412 ] + MICROKERNEL_TEST_HDRS,
3413 deps = MICROKERNEL_TEST_DEPS,
3414)
3415
3416xnnpack_unit_test(
3417 name = "u8_rmax_test",
3418 srcs = [
3419 "test/u8-rmax.cc",
3420 "test/rmax-microkernel-tester.h",
3421 ] + MICROKERNEL_TEST_HDRS,
3422 deps = MICROKERNEL_TEST_DEPS,
3423)
3424
3425xnnpack_unit_test(
3426 name = "x32_packx_test",
3427 srcs = [
3428 "test/x32-packx.cc",
3429 "test/pack-microkernel-tester.h",
3430 "src/xnnpack/AlignedAllocator.h",
3431 ] + MICROKERNEL_TEST_HDRS,
3432 deps = MICROKERNEL_TEST_DEPS,
3433)
3434
3435xnnpack_unit_test(
3436 name = "x32_pad_test",
3437 srcs = [
3438 "test/x32-pad.cc",
3439 "test/pad-microkernel-tester.h",
3440 ] + MICROKERNEL_TEST_HDRS,
3441 deps = MICROKERNEL_TEST_DEPS,
3442)
3443
3444xnnpack_unit_test(
3445 name = "x32_unpool_test",
3446 srcs = [
3447 "test/x32-unpool.cc",
3448 "test/unpool-microkernel-tester.h",
3449 ] + MICROKERNEL_TEST_HDRS,
3450 deps = MICROKERNEL_TEST_DEPS,
3451)
3452
3453xnnpack_unit_test(
3454 name = "x32_zip_test",
3455 srcs = [
3456 "test/x32-zip.cc",
3457 "test/zip-microkernel-tester.h",
3458 ] + MICROKERNEL_TEST_HDRS,
3459 deps = MICROKERNEL_TEST_DEPS,
3460)
3461
3462xnnpack_unit_test(
3463 name = "x8_lut_test",
3464 srcs = [
3465 "test/x8-lut.cc",
3466 "test/lut-microkernel-tester.h",
3467 ] + MICROKERNEL_TEST_HDRS,
3468 deps = MICROKERNEL_TEST_DEPS,
3469)
3470
3471xnnpack_unit_test(
3472 name = "x8_zip_test",
3473 srcs = [
3474 "test/x8-zip.cc",
3475 "test/zip-microkernel-tester.h",
3476 ] + MICROKERNEL_TEST_HDRS,
3477 deps = MICROKERNEL_TEST_DEPS,
3478)
3479
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003480xnnpack_unit_test(
3481 name = "requantization_test",
3482 srcs = [
3483 "src/xnnpack/requantization-stubs.h",
3484 "test/requantization.cc",
3485 "test/requantization-tester.h",
3486 ] + MICROKERNEL_TEST_HDRS,
3487 deps = MICROKERNEL_TEST_DEPS,
3488)
3489
Marat Dukhan20c3b922020-03-10 03:45:06 -07003490########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491
3492xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003493 name = "operator_size_test",
3494 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003495 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003496)
3497
Marat Dukhan20c3b922020-03-10 03:45:06 -07003498xnnpack_binary(
3499 name = "subgraph_size_test",
3500 srcs = ["test/subgraph-size.c"],
3501 deps = [":XNNPACK"],
3502)
3503
3504########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505
3506xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003507 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003509 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003510 "test/add-operator-tester.h",
3511 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003512 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513)
3514
3515xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003516 name = "add_nd_test",
3517 srcs = [
3518 "test/add-nd.cc",
3519 "test/binary-elementwise-operator-tester.h",
3520 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003521 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003522)
3523
3524xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003525 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003526 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003527 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003528 "test/argmax-pooling-operator-tester.h",
3529 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003530 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003531)
3532
3533xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003534 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003535 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003536 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003537 "test/average-pooling-operator-tester.h",
3538 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003539 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003540)
3541
3542xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003543 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003544 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003545 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546 "test/channel-pad-operator-tester.h",
3547 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003548 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003549)
3550
3551xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003552 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003554 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 "test/channel-shuffle-operator-tester.h",
3556 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003557 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558)
3559
3560xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003561 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003562 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003563 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 "test/clamp-operator-tester.h",
3565 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003566 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003567)
3568
3569xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003570 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003572 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 "test/convolution-operator-tester.h",
3574 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003575 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576)
3577
3578xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003579 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003581 "test/convolution-nchw.cc",
3582 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003584 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585)
3586
3587xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003588 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003590 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003591 "test/deconvolution-operator-tester.h",
3592 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003593 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594)
3595
3596xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003597 name = "divide_nd_test",
3598 srcs = [
3599 "test/binary-elementwise-operator-tester.h",
3600 "test/divide-nd.cc",
3601 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003602 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003603)
3604
3605xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003606 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003608 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 "test/fully-connected-operator-tester.h",
3610 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003611 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612)
3613
3614xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003615 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003616 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003617 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 "test/global-average-pooling-operator-tester.h",
3619 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003620 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003621)
3622
3623xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003624 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003625 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003626 "test/global-average-pooling-ncw.cc",
3627 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003628 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003629 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630)
3631
3632xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003633 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003635 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 "test/hardswish-operator-tester.h",
3637 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003638 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003639)
3640
3641xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003642 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003644 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 "test/leaky-relu-operator-tester.h",
3646 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003647 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648)
3649
3650xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003651 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003653 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654 "test/max-pooling-operator-tester.h",
3655 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003656 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657)
3658
3659xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003660 name = "maximum_nd_test",
3661 srcs = [
3662 "test/binary-elementwise-operator-tester.h",
3663 "test/maximum-nd.cc",
3664 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003665 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003666)
3667
3668xnnpack_unit_test(
3669 name = "minimum_nd_test",
3670 srcs = [
3671 "test/binary-elementwise-operator-tester.h",
3672 "test/minimum-nd.cc",
3673 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003674 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003675)
3676
3677xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003678 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003679 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003680 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003681 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003682 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003683 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003684)
3685
3686xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003687 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003688 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003689 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003690 "test/prelu-operator-tester.h",
3691 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003692 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693)
3694
3695xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003696 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003697 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003698 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003699 "test/resize-bilinear-operator-tester.h",
3700 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003701 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003702)
3703
3704xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003705 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003707 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003708 "test/sigmoid-operator-tester.h",
3709 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003710 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711)
3712
3713xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003714 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003716 "test/softmax-nc.cc",
3717 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003719 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003720)
3721
3722xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003723 name = "subtract_nd_test",
3724 srcs = [
3725 "test/binary-elementwise-operator-tester.h",
3726 "test/subtract-nd.cc",
3727 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003728 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003729)
3730
3731xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003732 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003733 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003734 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 "test/unpooling-operator-tester.h",
3736 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003737 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738)
3739
Chao Mei6ddfc602020-05-13 22:29:36 -07003740############################### Misc unit tests ###############################
3741
3742xnnpack_unit_test(
3743 name = "memory_planner_test",
3744 srcs = [
3745 "test/memory-planner-test.cc",
3746 ],
3747 deps = [
3748 ":XNNPACK",
3749 ":memory_planner",
3750 ],
3751)
3752
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753############################# Build configurations #############################
3754
Marat Dukhanb8642352019-10-30 15:43:02 -07003755# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003757 name = "xnn_enable_assembly_explicit_true",
3758 define_values = {"xnn_enable_assembly": "true"},
3759)
3760
3761# Disables usage of assembly kernels.
3762config_setting(
3763 name = "xnn_enable_assembly_explicit_false",
3764 define_values = {"xnn_enable_assembly": "false"},
3765)
3766
Marat Dukhan05702cf2020-03-26 15:41:33 -07003767# Disables usage of HMP-aware optimizations.
3768config_setting(
3769 name = "xnn_enable_hmp_explicit_false",
3770 define_values = {"xnn_enable_hmp": "false"},
3771)
3772
Chao Mei6ddfc602020-05-13 22:29:36 -07003773# Enable usage of optimized memory allocation
3774config_setting(
3775 name = "xnn_enable_memopt_explicit_true",
3776 define_values = {"xnnpack_enable_memopt": "true"},
3777)
3778
3779# Disable usage of optimized memory allocation
3780config_setting(
3781 name = "xnn_enable_memopt_explicit_false",
3782 define_values = {"xnnpack_enable_memopt": "false"},
3783)
3784
Marat Dukhanb8642352019-10-30 15:43:02 -07003785# Builds with -c dbg
3786config_setting(
3787 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003789 "compilation_mode": "dbg",
3790 },
3791)
3792
3793# Builds with -c opt
3794config_setting(
3795 name = "optimized_build",
3796 values = {
3797 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003798 },
3799)
3800
3801config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003802 name = "linux_k8",
3803 values = {"cpu": "k8"},
3804)
3805
3806config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003807 name = "linux_arm",
3808 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003809)
3810
3811config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003812 name = "linux_armhf",
3813 values = {"cpu": "armhf"},
3814)
3815
3816config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003817 name = "linux_aarch64",
3818 values = {"cpu": "aarch64"},
3819)
3820
3821config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822 name = "android",
3823 values = {"crosstool_top": "//external:android/crosstool"},
3824)
3825
3826config_setting(
3827 name = "android_armv7",
3828 values = {
3829 "crosstool_top": "//external:android/crosstool",
3830 "cpu": "armeabi-v7a",
3831 },
3832)
3833
3834config_setting(
3835 name = "android_arm64",
3836 values = {
3837 "crosstool_top": "//external:android/crosstool",
3838 "cpu": "arm64-v8a",
3839 },
3840)
3841
3842config_setting(
3843 name = "android_x86",
3844 values = {
3845 "crosstool_top": "//external:android/crosstool",
3846 "cpu": "x86",
3847 },
3848)
3849
3850config_setting(
3851 name = "android_x86_64",
3852 values = {
3853 "crosstool_top": "//external:android/crosstool",
3854 "cpu": "x86_64",
3855 },
3856)
3857
3858config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003859 name = "windows_x86_64",
3860 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003861)
3862
3863config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003864 name = "windows_x86_64_clang",
3865 values = {
3866 "compiler": "clang-cl",
3867 "cpu": "x64_windows",
3868 },
3869)
3870
3871config_setting(
3872 name = "windows_x86_64_mingw",
3873 values = {
3874 "compiler": "mingw-gcc",
3875 "cpu": "x64_windows",
3876 },
3877)
3878
3879config_setting(
3880 name = "windows_x86_64_msys",
3881 values = {
3882 "compiler": "msys-gcc",
3883 "cpu": "x64_windows",
3884 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003885)
3886
3887config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003888 name = "macos_x86_64",
3889 values = {
3890 "apple_platform_type": "macos",
3891 "cpu": "darwin",
3892 },
3893)
3894
3895config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003896 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003897 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898)
3899
3900config_setting(
3901 name = "emscripten_wasm",
3902 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003903 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904 "cpu": "wasm",
3905 },
3906)
3907
3908config_setting(
3909 name = "emscripten_wasmsimd",
3910 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003911 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003912 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003913 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 },
3915)
3916
3917config_setting(
3918 name = "emscripten_asmjs",
3919 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003920 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 "cpu": "asmjs",
3922 },
3923)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003924
3925config_setting(
3926 name = "ios_armv7",
3927 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003928 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003929 "cpu": "ios_armv7",
3930 },
3931)
3932
3933config_setting(
3934 name = "ios_arm64",
3935 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003936 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003937 "cpu": "ios_arm64",
3938 },
3939)
3940
3941config_setting(
3942 name = "ios_arm64e",
3943 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003944 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003945 "cpu": "ios_arm64e",
3946 },
3947)
3948
3949config_setting(
3950 name = "ios_x86",
3951 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003952 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003953 "cpu": "ios_i386",
3954 },
3955)
3956
3957config_setting(
3958 name = "ios_x86_64",
3959 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003960 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003961 "cpu": "ios_x86_64",
3962 },
3963)
3964
3965config_setting(
3966 name = "watchos_armv7k",
3967 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003968 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003969 "cpu": "watchos_armv7k",
3970 },
3971)
3972
3973config_setting(
3974 name = "watchos_arm64_32",
3975 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003976 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003977 "cpu": "watchos_arm64_32",
3978 },
3979)
3980
3981config_setting(
3982 name = "watchos_x86",
3983 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003984 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003985 "cpu": "watchos_i386",
3986 },
3987)
3988
3989config_setting(
3990 name = "watchos_x86_64",
3991 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003992 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003993 "cpu": "watchos_x86_64",
3994 },
3995)
3996
3997config_setting(
3998 name = "tvos_arm64",
3999 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004000 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004001 "cpu": "tvos_arm64",
4002 },
4003)
4004
4005config_setting(
4006 name = "tvos_x86_64",
4007 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004008 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004009 "cpu": "tvos_x86_64",
4010 },
4011)