blob: 6d49183fe013e3788ca9357d31570ffd31580b17 [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",
Marat Dukhan065b11e2020-05-22 09:49:41 -070060 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070061 "src/operators/convolution-nchw.c",
62 "src/operators/convolution-nhwc.c",
63 "src/operators/deconvolution-nhwc.c",
64 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
67 "src/operators/hardswish-nc.c",
68 "src/operators/leaky-relu-nc.c",
69 "src/operators/max-pooling-nhwc.c",
70 "src/operators/prelu-nc.c",
71 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
74 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070075]
76
Marat Dukhan3a77ea72019-12-23 12:10:24 -080077TABLE_SRCS = [
78 "src/tables/exp2-k-over-64.c",
79 "src/tables/exp2-k-over-2048.c",
80]
81
Marat Dukhan08c4a432019-10-03 09:29:21 -070082SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080083 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080084 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080085 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -070086 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
87 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -070088 "src/f32-clamp/gen/scalar-x1.c",
89 "src/f32-clamp/gen/scalar-x2.c",
90 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080091 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080092 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -070093 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
94 "src/f32-dwconv-chw/3x3p1-scalar.c",
95 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
96 "src/f32-dwconv-chw/5x5p2-scalar.c",
97 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -070098 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
99 "src/f32-dwconv/gen/up1x4-scalar.c",
100 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
101 "src/f32-dwconv/gen/up1x9-scalar.c",
102 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
103 "src/f32-dwconv/gen/up1x25-scalar.c",
104 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
105 "src/f32-dwconv/gen/up2x4-scalar.c",
106 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
107 "src/f32-dwconv/gen/up2x9-scalar.c",
108 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
109 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700110 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
111 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
112 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
113 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700114 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
115 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700116 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
117 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
118 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
119 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700120 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
121 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700122 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700123 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
124 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700125 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
126 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
127 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700128 "src/f32-gemm/gen/1x4-scalar.c",
129 "src/f32-gemm/gen/2x4-scalar.c",
130 "src/f32-gemm/gen/4x2-scalar.c",
131 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700132 "src/f32-gemm/gen/1x4-minmax-scalar.c",
133 "src/f32-gemm/gen/2x4-minmax-scalar.c",
134 "src/f32-gemm/gen/4x2-minmax-scalar.c",
135 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800136 "src/f32-hswish/gen/scalar-x1.c",
137 "src/f32-hswish/gen/scalar-x2.c",
138 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700139 "src/f32-ibilinear/gen/scalar-c1.c",
140 "src/f32-ibilinear/gen/scalar-c2.c",
141 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700142 "src/f32-igemm/gen/1x4-scalar.c",
143 "src/f32-igemm/gen/2x4-scalar.c",
144 "src/f32-igemm/gen/4x2-scalar.c",
145 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-igemm/gen/1x4-minmax-scalar.c",
147 "src/f32-igemm/gen/2x4-minmax-scalar.c",
148 "src/f32-igemm/gen/4x2-minmax-scalar.c",
149 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700150 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
151 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
152 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700153 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
154 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
155 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
156 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800157 "src/f32-prelu/gen/scalar-2x1.c",
158 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800159 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
160 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
161 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
162 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
163 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
164 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
165 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
166 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
167 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
168 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
169 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
170 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700171 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800172 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
173 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
174 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
175 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
176 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
177 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
178 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
179 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
180 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700181 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
182 "src/f32-spmm/gen/1x1-minmax-scalar.c",
183 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
184 "src/f32-spmm/gen/2x1-minmax-scalar.c",
185 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
186 "src/f32-spmm/gen/4x1-minmax-scalar.c",
187 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
188 "src/f32-spmm/gen/8x1-minmax-scalar.c",
189 "src/f32-spmm/gen/8x2-minmax-scalar.c",
190 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700191 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
192 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
193 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
194 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
195 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
196 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
197 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
198 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
199 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
200 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
201 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
202 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800203 "src/f32-vbinary/gen/vmax-scalar-x1.c",
204 "src/f32-vbinary/gen/vmax-scalar-x2.c",
205 "src/f32-vbinary/gen/vmax-scalar-x4.c",
206 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
207 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
208 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
209 "src/f32-vbinary/gen/vmin-scalar-x1.c",
210 "src/f32-vbinary/gen/vmin-scalar-x2.c",
211 "src/f32-vbinary/gen/vmin-scalar-x4.c",
212 "src/f32-vbinary/gen/vminc-scalar-x1.c",
213 "src/f32-vbinary/gen/vminc-scalar-x2.c",
214 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700215 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
216 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
217 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
218 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
219 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
220 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
221 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
222 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
223 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
224 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
225 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
226 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
227 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
228 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
229 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
230 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
231 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
232 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700233 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
234 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
235 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800236 "src/math/expminus-scalar-lut2048-p1.c",
237 "src/math/expminus-scalar-lut64-p2.c",
238 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700239 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700240 "src/math/roundne-scalar-nearbyint.c",
241 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700242 "src/math/roundd-scalar-addsub.c",
243 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700244 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700245 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700246 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700247 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700248 "src/math/roundz-scalar-addsub.c",
249 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700250 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800251 "src/math/sigmoid-scalar-lut2048-p1-div.c",
252 "src/math/sigmoid-scalar-lut64-p2-div.c",
253 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700254 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
255 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700256 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700257 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
258 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700259 "src/q8-gemm/2x2-minmax-scalar.c",
260 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700261 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700262 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700263 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700264 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700265 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700266 "src/x32-fill/scalar-float.c",
267 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700268 "src/x32-packx/x2-scalar.c",
269 "src/x32-packx/x3-scalar.c",
270 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700271 "src/x32-pad/scalar-int.c",
272 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700273 "src/x32-unpool/scalar.c",
274 "src/x32-zip/x2-scalar.c",
275 "src/x32-zip/x3-scalar.c",
276 "src/x32-zip/x4-scalar.c",
277 "src/x32-zip/xm-scalar.c",
278 "src/x8-lut/scalar.c",
279 "src/x8-zip/x2-scalar.c",
280 "src/x8-zip/x3-scalar.c",
281 "src/x8-zip/x4-scalar.c",
282 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700283 "src/requantization/precise-scalar.c",
284 "src/requantization/fp32-scalar.c",
285 "src/requantization/q31-scalar.c",
286 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700287]
288
Marat Dukhan436ebe62019-12-04 15:10:12 -0800289WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700290 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
291 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700292 "src/f32-clamp/gen/wasm-x1.c",
293 "src/f32-clamp/gen/wasm-x2.c",
294 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700295 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
296 "src/f32-dwconv/gen/up1x4-wasm.c",
297 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
298 "src/f32-dwconv/gen/up1x9-wasm.c",
299 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
300 "src/f32-dwconv/gen/up1x25-wasm.c",
301 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
302 "src/f32-dwconv/gen/up2x4-wasm.c",
303 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
304 "src/f32-dwconv/gen/up2x9-wasm.c",
305 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
306 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700307 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
308 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
309 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
310 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700311 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
312 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700313 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
314 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
315 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
316 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700317 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
318 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700319 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
320 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700321 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
322 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
323 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700324 "src/f32-gemm/gen/1x4-wasm.c",
325 "src/f32-gemm/gen/2x4-wasm.c",
326 "src/f32-gemm/gen/4x2-wasm.c",
327 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700328 "src/f32-gemm/gen/1x4-minmax-wasm.c",
329 "src/f32-gemm/gen/2x4-minmax-wasm.c",
330 "src/f32-gemm/gen/4x2-minmax-wasm.c",
331 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800332 "src/f32-hswish/gen/wasm-x1.c",
333 "src/f32-hswish/gen/wasm-x2.c",
334 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700335 "src/f32-igemm/gen/1x4-wasm.c",
336 "src/f32-igemm/gen/2x4-wasm.c",
337 "src/f32-igemm/gen/4x2-wasm.c",
338 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700339 "src/f32-igemm/gen/1x4-minmax-wasm.c",
340 "src/f32-igemm/gen/2x4-minmax-wasm.c",
341 "src/f32-igemm/gen/4x2-minmax-wasm.c",
342 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700343 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
344 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
345 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700346 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
347 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
348 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
349 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
350 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
351 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
352 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
353 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
354 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
355 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
356 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
357 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800358 "src/f32-vbinary/gen/vmax-wasm-x1.c",
359 "src/f32-vbinary/gen/vmax-wasm-x2.c",
360 "src/f32-vbinary/gen/vmax-wasm-x4.c",
361 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
362 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
363 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
364 "src/f32-vbinary/gen/vmin-wasm-x1.c",
365 "src/f32-vbinary/gen/vmin-wasm-x2.c",
366 "src/f32-vbinary/gen/vmin-wasm-x4.c",
367 "src/f32-vbinary/gen/vminc-wasm-x1.c",
368 "src/f32-vbinary/gen/vminc-wasm-x2.c",
369 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700370 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
371 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
372 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
373 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
374 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
375 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
376 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
377 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
378 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
379 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
380 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
381 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
382 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
383 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
384 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
385 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
386 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
387 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700388 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
389 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
390 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800391]
392
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800393PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800394 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800395 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800396 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700397 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
398 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700399 "src/f32-clamp/gen/psimd-x4.c",
400 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700401 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
402 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
403 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
404 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
405 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
406 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
407 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
408 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
409 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
410 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
411 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
412 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700413 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
414 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700415 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
416 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
417 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
418 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
419 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
420 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
421 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
422 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
423 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
424 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
425 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
426 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
427 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
428 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
429 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
430 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
431 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
432 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
433 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800434 "src/f32-hswish/gen/psimd-x4.c",
435 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700436 "src/f32-ibilinear/gen/psimd-c4.c",
437 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700438 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
439 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
440 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
441 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
442 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
443 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
444 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
445 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
446 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
447 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700448 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
449 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
450 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700451 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800452 "src/f32-prelu/gen/psimd-2x4.c",
453 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800454 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700455 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
456 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
457 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
458 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
459 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
460 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
461 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
462 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800463 "src/f32-vbinary/gen/vmax-psimd-x4.c",
464 "src/f32-vbinary/gen/vmax-psimd-x8.c",
465 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
466 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
467 "src/f32-vbinary/gen/vmin-psimd-x4.c",
468 "src/f32-vbinary/gen/vmin-psimd-x8.c",
469 "src/f32-vbinary/gen/vminc-psimd-x4.c",
470 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700471 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
472 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
473 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
474 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
475 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
476 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
477 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
478 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
479 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
480 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
481 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
482 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700483 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
484 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700485 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700486 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700487 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700488 "src/x32-unpool/psimd.c",
489 "src/x32-zip/x2-psimd.c",
490 "src/x32-zip/x3-psimd.c",
491 "src/x32-zip/x4-psimd.c",
492 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700493 "src/requantization/precise-psimd.c",
494 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700495]
496
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800497PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800498 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
499 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
500 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
501 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
502 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
503 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
504 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
505 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
506 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
507 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
508 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
509 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800510 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
511 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
512 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
513 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
514 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
515 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700516 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700517 "src/math/roundd-psimd-addsub.c",
518 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700519 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800520 "src/math/sigmoid-psimd-p5-div.c",
521]
522
Marat Dukhan08c4a432019-10-03 09:29:21 -0700523# ISA-specific micro-kernels
524NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700525 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
526 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700527 "src/f32-clamp/gen/neon-x4.c",
528 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700529 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
530 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700531 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
532 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700533 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
534 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700535 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
536 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700537 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
538 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
539 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
540 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700541 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
542 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
543 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
544 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700545 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
546 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
547 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
548 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700549 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700550 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
551 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700552 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
553 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
554 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
555 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
556 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
557 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
558 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
559 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
560 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
561 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
562 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
563 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
564 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
565 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
566 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
567 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
568 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
569 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
570 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
571 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
572 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
573 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
574 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
575 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
576 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
577 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
578 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
579 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
580 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
581 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
582 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800583 "src/f32-hswish/gen/neon-x4.c",
584 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700585 "src/f32-ibilinear/gen/neon-c4.c",
586 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700587 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
588 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
589 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
590 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
591 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
592 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
593 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
594 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
595 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
596 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
597 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
598 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
599 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
600 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
601 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
602 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700603 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
604 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
605 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700606 "src/f32-ppmm/gen/4x8-minmax-neon.c",
607 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800608 "src/f32-prelu/gen/neon-2x4.c",
609 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800610 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
611 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
612 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
613 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
614 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
615 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
616 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
617 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
618 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
619 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
620 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
621 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
622 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
623 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
624 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
625 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
626 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
627 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
628 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
629 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
630 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
631 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
632 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
633 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700634 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800635 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800636 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
637 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
638 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
639 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
640 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
641 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
642 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
643 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
644 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
645 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
646 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
647 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
648 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
649 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
650 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
651 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
652 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
653 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700654 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
655 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
656 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
657 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800658 "src/f32-vbinary/gen/vmax-neon-x4.c",
659 "src/f32-vbinary/gen/vmax-neon-x8.c",
660 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
661 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
662 "src/f32-vbinary/gen/vmin-neon-x4.c",
663 "src/f32-vbinary/gen/vmin-neon-x8.c",
664 "src/f32-vbinary/gen/vminc-neon-x4.c",
665 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700666 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
667 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
668 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
670 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
671 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
672 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
673 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
674 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
675 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700676 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
677 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
678 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
679 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700680 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700681 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
682 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700683 "src/q8-gemm/4x8-minmax-neon.c",
684 "src/q8-gemm/8x8-minmax-neon.c",
685 "src/q8-igemm/4x8-minmax-neon.c",
686 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700687 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700688 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700689 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700690 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700691 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700692 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700693 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700694 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700695 "src/x32-zip/x2-neon.c",
696 "src/x32-zip/x3-neon.c",
697 "src/x32-zip/x4-neon.c",
698 "src/x32-zip/xm-neon.c",
699 "src/x8-zip/x2-neon.c",
700 "src/x8-zip/x3-neon.c",
701 "src/x8-zip/x4-neon.c",
702 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700703 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700704 "src/math/roundd-neon-addsub.c",
705 "src/math/roundd-neon-cvt.c",
706 "src/math/roundu-neon-addsub.c",
707 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700708 "src/math/roundz-neon-addsub.c",
709 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800710 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800711 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
712 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
713 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
714 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
715 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
716 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700717 "src/requantization/precise-neon.c",
718 "src/requantization/fp32-neon.c",
719 "src/requantization/q31-neon.c",
720 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700721]
722
723NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700724 "src/f32-ibilinear/gen/neonfma-c4.c",
725 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700726 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
727 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
728 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
729 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
730 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
731 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
732 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
733 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
734 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700735 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
736 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
737 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
738 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700739 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
740 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
741 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
742 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700743 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
744 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
745 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
746 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700747 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
748 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
749 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
750 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
751 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
752 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
753 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
754 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
755 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
756 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
757 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
758 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
761 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
762 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
763 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
764 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800765 "src/f32-hswish/gen/neonfma-x4.c",
766 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700767 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
768 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800769 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
770 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
771 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
772 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
773 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
774 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
775 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
776 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
777 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
778 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
779 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
780 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
781 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
782 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
783 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
784 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
785 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
786 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
787 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
788 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
789 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
790 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
791 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
792 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800793 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
794 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
795 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
796 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
797 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
798 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
816 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
817 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
818 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
819 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
820 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
821 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
822 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
823 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
824 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
825 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
826 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
827 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
828 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
829 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
830 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
831 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
832 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
833 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
834 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
835 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
836 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
837 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
838 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
839 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
840 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
841 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
842 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
843 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
844 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
845 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
846 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700847 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
848 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800849 "src/math/exp-neonfma-lut64-p2.c",
850 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800851 "src/math/expminus-neonfma-lut2048-p1.c",
852 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800853 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800854 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
855 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
856 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
857 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
858 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
859 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
860 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
861 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
862 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
863 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
864 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
865 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
866 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
867 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
868 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
869 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
870 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
871 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700872]
873
874AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700875 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
876 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
877 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
878 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
879 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
880 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700881 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
882 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
883 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
884 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
885 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
886 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
887 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
888 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
889 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
890 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
891 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
892 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
893 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
894 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
895 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
896 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
897 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
898 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
899 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
900 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700903 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
904 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700905 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
906 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700907 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
908 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700909 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
910 "src/f32-dwconv-chw/3x3p1-neonfma.c",
911 "src/f32-dwconv-chw/5x5p2-neonfma.c",
912 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
913 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800914 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
915 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
916 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
917 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
918 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
919 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
926 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
927 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
928 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
929 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
930 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
931 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700932 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
933 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
934 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
935 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
936 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
937 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
938 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
939 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
940 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
941 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
942 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
943 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
944 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
945 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
946 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
947 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
948 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
949 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800950 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
951 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
952 "src/math/sigmoid-neonfma-rr1-p5-div.c",
953 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
954 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
955 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700956]
957
Marat Dukhan8853b822020-05-07 12:19:01 -0700958NEONV8_UKERNELS = [
959 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700960 "src/math/roundd-neonv8.c",
961 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700962 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -0700963]
964
Marat Dukhan08c4a432019-10-03 09:29:21 -0700965AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -0700966 "src/f16-clamp/gen/neonfp16arith-x16.c",
967 "src/f16-clamp/gen/neonfp16arith-x8.c",
968 "src/f16-hswish/gen/neonfp16arith-x16.c",
969 "src/f16-hswish/gen/neonfp16arith-x8.c",
970 "src/f16-prelu/gen/neonfp16arith-2x16.c",
971 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -0700972 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
973 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
974 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
975 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
976 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
977 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
978 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
979 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
980 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
981 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
982 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
983 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
984 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
985 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
986 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
987 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
988 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
989 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
990 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
991 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
992 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
993 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
994 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
995 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
996 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
997 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
998 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
999 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001000 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1001 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1002 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1003 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1004 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1005 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1006 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1007 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001008 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1009 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1010 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1011 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001012 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1013 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1014 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1015 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1016 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1017 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1018 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1019 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1020 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1021 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1022 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1023 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001024 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1025 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1026 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1027 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1028 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1029 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1030 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1031 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001032]
1033
1034SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001035 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1036 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001037 "src/f32-clamp/gen/sse-x4.c",
1038 "src/f32-clamp/gen/sse-x8.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001039 "src/f32-dwconv-chw/3x3p1-sse.c",
1040 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001041 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1042 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1043 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1044 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1045 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1046 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1047 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1048 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1049 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1050 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1051 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1052 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001053 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001054 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1055 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001056 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1057 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1058 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1059 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1060 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1061 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1062 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1063 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1064 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1065 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1066 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1067 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1068 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001069 "src/f32-hswish/gen/sse-x4.c",
1070 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001071 "src/f32-ibilinear/gen/sse-c4.c",
1072 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001073 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1074 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1075 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1076 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1077 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1078 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1079 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001080 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1081 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1082 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001083 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001084 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001085 "src/f32-spmm/gen/4x1-minmax-sse.c",
1086 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001087 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1088 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1089 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1090 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1091 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1092 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1093 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1094 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001095 "src/f32-vbinary/gen/vmax-sse-x4.c",
1096 "src/f32-vbinary/gen/vmax-sse-x8.c",
1097 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1098 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1099 "src/f32-vbinary/gen/vmin-sse-x4.c",
1100 "src/f32-vbinary/gen/vmin-sse-x8.c",
1101 "src/f32-vbinary/gen/vminc-sse-x4.c",
1102 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001103 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1104 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1105 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1106 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1107 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1108 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1109 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1110 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
1111 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1112 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1113 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1114 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001115 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1116 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001117 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001118 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001119 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001120 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001121 "src/math/roundd-sse-addsub.c",
1122 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001123 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001124]
1125
1126SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001127 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1128 "src/f32-argmaxpool/4x-sse2-c4.c",
1129 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001130 "src/f32-prelu/gen/sse2-2x4.c",
1131 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1133 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1134 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1135 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1136 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1137 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1138 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1139 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1140 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1141 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1142 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1143 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001144 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001145 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001146 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001147 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001148 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1149 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001150 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1151 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001152 "src/q8-igemm/4x4c2-minmax-sse2.c",
1153 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001154 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1155 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001156 "src/q8-gemm/2x4c8-minmax-sse2.c",
1157 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001158 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001159 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001160 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001161 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001162 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001163 "src/x32-zip/x2-sse2.c",
1164 "src/x32-zip/x3-sse2.c",
1165 "src/x32-zip/x4-sse2.c",
1166 "src/x32-zip/xm-sse2.c",
1167 "src/x8-zip/x2-sse2.c",
1168 "src/x8-zip/x3-sse2.c",
1169 "src/x8-zip/x4-sse2.c",
1170 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001171 "src/math/exp-sse2-p5.c",
1172 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001173 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001174 "src/math/roundd-sse2-cvt.c",
1175 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001176 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001177 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001178 "src/requantization/precise-sse2.c",
1179 "src/requantization/fp32-sse2.c",
1180 "src/requantization/q31-sse2.c",
1181 "src/requantization/gemmlowp-sse2.c",
1182]
1183
1184SSSE3_UKERNELS = [
1185 "src/requantization/precise-ssse3.c",
1186 "src/requantization/q31-ssse3.c",
1187 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001188]
1189
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001190SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001191 "src/f32-prelu/gen/sse41-2x4.c",
1192 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001193 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001194 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001195 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001196 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001197 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1198 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001199 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001200 "src/math/roundd-sse41.c",
1201 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001202 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001203 "src/requantization/precise-sse4.c",
1204 "src/requantization/q31-sse4.c",
1205 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001206]
1207
Marat Dukhan08c4a432019-10-03 09:29:21 -07001208AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001209 "src/f32-clamp/gen/avx-x8.c",
1210 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001211 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1212 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1213 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1214 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1215 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1216 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1217 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1218 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1219 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1220 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1221 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1222 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1223 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1224 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1225 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1226 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1227 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1228 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1229 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1230 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1231 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1232 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1233 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1234 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1235 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1236 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1237 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1238 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1239 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1240 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001241 "src/f32-hswish/gen/avx-x8.c",
1242 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001243 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1244 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1245 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1246 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1247 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1248 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1249 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1250 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1251 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001252 "src/f32-prelu/gen/avx-2x8.c",
1253 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001254 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001255 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1256 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1257 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1258 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1259 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1260 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1261 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1262 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001263 "src/f32-vbinary/gen/vmax-avx-x8.c",
1264 "src/f32-vbinary/gen/vmax-avx-x16.c",
1265 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1266 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1267 "src/f32-vbinary/gen/vmin-avx-x8.c",
1268 "src/f32-vbinary/gen/vmin-avx-x16.c",
1269 "src/f32-vbinary/gen/vminc-avx-x8.c",
1270 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001271 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1272 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1273 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1274 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1275 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1276 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1277 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1278 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
1279 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1280 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1281 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1282 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001283 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001284]
1285
Marat Dukhanfda12b82019-11-21 12:27:59 -08001286FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001287 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1288 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1289 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1290 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1291 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1292 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1293 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1294 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1295 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1296 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1297 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1298 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1299 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1300 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1301 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1302 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1303 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1304 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1305 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1306 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1307 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1308 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1309 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1310 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1311 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1312 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1313 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1314 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1315 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1316 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1317 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1318 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1319 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1320 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1321 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1322 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1323 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1324 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1325 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1326 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001327 "src/f32-hswish/gen/fma3-x8.c",
1328 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001329 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1330 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1331 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1332 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1333 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1334 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1335 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1336 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1337 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1338 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1339 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1340 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1341 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1342 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001343]
1344
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001345AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001346 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1347 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1348 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1349 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1350 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1351 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1352 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1353 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1354 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1355 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1356 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1357 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1358 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1359 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1360 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1361 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1362 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1363 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1364 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1365 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1366 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1367 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1368 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1369 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1370 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1373 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1374 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1375 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1376 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1377 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1378 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1379 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1380 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1381 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001382 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1383 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1384 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1385 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1386 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1387 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1388 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1389 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1390 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1391 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1403 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1404 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1405 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1406 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1407 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1408 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1409 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1410 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1411 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001412 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1415 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1416 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1417 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1418 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1419 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1420 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1421 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1422 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1423 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1424 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1425 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1426 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1427 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1428 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1429 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1430 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1431 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1432 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1433 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1434 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1435 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001436 "src/math/exp-avx2-p5.c",
1437 "src/math/exp-avx2-perm-p3.c",
1438 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001439 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001440 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001441 "src/math/sigmoid-avx2-rr2-p5-div.c",
1442 "src/math/sigmoid-avx2-rr1-p5-div.c",
1443 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1444 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1445 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1446 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001447]
1448
Marat Dukhan08c4a432019-10-03 09:29:21 -07001449AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001450 "src/f32-clamp/gen/avx512f-x16.c",
1451 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001452 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1453 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1454 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1455 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1456 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1457 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1458 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1459 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1460 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1461 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1462 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1463 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1464 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1465 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1466 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1467 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1468 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1469 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1470 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1471 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1472 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1473 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1474 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1475 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001476 "src/f32-hswish/gen/avx512f-x16.c",
1477 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001478 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1479 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1480 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1481 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1482 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1483 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001484 "src/f32-prelu/gen/avx512f-2x16.c",
1485 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001486 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1487 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1488 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1489 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1490 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1491 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1492 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1493 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1494 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1495 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1496 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1497 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1498 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1499 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1500 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1501 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1502 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1503 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1504 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1505 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1506 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1507 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1508 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1509 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1510 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1511 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1512 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1513 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1514 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1515 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1516 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1517 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1518 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1519 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1520 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1521 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001522 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001523 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1524 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1525 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1526 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1527 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1528 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1529 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1530 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001531 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1532 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1533 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1534 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1535 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1536 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1537 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1538 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001539 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1540 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1541 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1542 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1543 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1544 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1545 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1546 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
1547 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1548 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1549 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1550 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001551 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001552 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1553 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1554 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1555 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1556 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1557 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1558 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1559 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1560 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1561 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1562 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1563 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1564 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1565 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1566 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1567 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1568 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1569 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1570 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1571 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1572 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1573 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1574 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1575 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001576 "src/math/exp-avx512f-p5-scalef.c",
1577 "src/math/exp-avx512f-p5.c",
1578 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001579 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001580 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581]
1582
1583AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001584 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001585 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001586 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1587 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1588 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1589 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1590 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1591 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1592 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1593 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1594 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1595 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001596]
1597
1598AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001599 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001600 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1601 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001602 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001603 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1604 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001605 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1606 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1607 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001608 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001609 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1610 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1611 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001612 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001613 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1614 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001615 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001616 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1617 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1618 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1619 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1620 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1621 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1622 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1623 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1624 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1625 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1626 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1627 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1628 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1629 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1630 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1631 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1632 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1633 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1634 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1635 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1636 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001637 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001638 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1639 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1640 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1641 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1642 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1643 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1644 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1645 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1646 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1647 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1648 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1649 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1650 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1651 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1652 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1653 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1654 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1655 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1656 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1657 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1658 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1659 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1660 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1661 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1662 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1663 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1664 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1665 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1666 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1667 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1668 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1669 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1670 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1671 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1672 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1673 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1674 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1675 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001676]
1677
Marat Dukhan1b354632020-03-23 12:50:22 -07001678INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001679 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680 "src/xnnpack/argmaxpool.h",
1681 "src/xnnpack/avgpool.h",
1682 "src/xnnpack/clamp.h",
1683 "src/xnnpack/common.h",
1684 "src/xnnpack/conv.h",
1685 "src/xnnpack/dwconv.h",
1686 "src/xnnpack/gavgpool.h",
1687 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001688 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001689 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001690 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001691 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001692 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693 "src/xnnpack/lut.h",
1694 "src/xnnpack/math.h",
1695 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001696 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001697 "src/xnnpack/packx.h",
1698 "src/xnnpack/pad.h",
1699 "src/xnnpack/params.h",
1700 "src/xnnpack/pavgpool.h",
1701 "src/xnnpack/ppmm.h",
1702 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001703 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001704 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001705 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001706 "src/xnnpack/rmax.h",
1707 "src/xnnpack/scalar-utils.h",
1708 "src/xnnpack/spmm.h",
1709 "src/xnnpack/unpool.h",
1710 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001711 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001713 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001714 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001715 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001716 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001717 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001718]
1719
1720INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 "include/xnnpack.h",
1722 "src/xnnpack/allocator.h",
1723 "src/xnnpack/compute.h",
1724 "src/xnnpack/im2col.h",
1725 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001726 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001727 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001728 "src/xnnpack/operator.h",
1729 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001730 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001732 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001733 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001734]
1735
Marat Dukhan1b354632020-03-23 12:50:22 -07001736ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001737 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001738]
1739
Marat Dukhan1b354632020-03-23 12:50:22 -07001740MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001741 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001742 "include/xnnpack.h",
1743]
1744
Marat Dukhan1b354632020-03-23 12:50:22 -07001745MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001746 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001747 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001748 "src/xnnpack/requantization.h",
1749 "include/xnnpack.h",
1750]
1751
1752OPERATOR_TEST_PARAMS_HDRS = [
1753 "src/xnnpack/params.h",
1754 "src/xnnpack/common.h",
1755]
1756
1757WEIGHTS_PACK_HDRS = [
1758 "src/xnnpack/pack.h",
1759 "src/xnnpack/operator.h",
1760 "src/xnnpack/compute.h",
1761]
1762
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001763LOGGING_COPTS = select({
1764 # No logging in optimized mode
1765 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1766 # Full logging in debug mode
1767 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1768 # Error-only logging in default (fastbuild) mode
1769 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1770})
1771
1772LOGGING_HDRS = [
1773 "src/xnnpack/log.h",
1774]
1775
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001777 name = "tables",
1778 srcs = TABLE_SRCS,
1779 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001780 gcc_copts = xnnpack_gcc_std_copts(),
1781 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001782)
1783
1784xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001785 name = "scalar_ukernels",
1786 srcs = SCALAR_UKERNELS,
1787 hdrs = INTERNAL_HDRS,
1788 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001789 gcc_copts = xnnpack_gcc_std_copts(),
1790 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001791 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001792 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001793 "@FP16",
1794 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001795 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001796 ],
1797)
1798
1799xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001800 name = "wasm_ukernels",
1801 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001802 gcc_copts = xnnpack_gcc_std_copts(),
1803 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001804 wasm_srcs = WASM_UKERNELS,
1805 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001806 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001807 "@FP16",
1808 "@FXdiv",
1809 "@pthreadpool",
1810 ],
1811)
1812
1813xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001814 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001815 hdrs = INTERNAL_HDRS,
1816 aarch32_copts = [
1817 "-marm",
1818 "-mfpu=neon",
1819 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001820 gcc_copts = xnnpack_gcc_std_copts(),
1821 gcc_x86_copts = ["-msse2"],
1822 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001823 optimized_copts = [
1824 "-O3",
1825 "-ffast-math",
1826 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001827 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001829 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001830 "@FP16",
1831 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001832 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833 ],
1834)
1835
1836xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001837 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001838 hdrs = INTERNAL_HDRS,
1839 aarch32_copts = [
1840 "-marm",
1841 "-mfpu=neon",
1842 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001843 gcc_copts = xnnpack_gcc_std_copts(),
1844 gcc_x86_copts = ["-msse2"],
1845 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001846 optimized_copts = [
1847 "-O3",
1848 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07001849 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001850 deps = [
1851 ":tables",
1852 "@FP16",
1853 "@psimd",
1854 "@pthreadpool",
1855 ],
1856)
1857
1858xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001859 name = "neon_ukernels",
1860 hdrs = INTERNAL_HDRS,
1861 aarch32_copts = [
1862 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001863 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001864 "-mfpu=neon",
1865 ],
1866 aarch32_srcs = NEON_UKERNELS,
1867 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001868 gcc_copts = xnnpack_gcc_std_copts(),
1869 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001870 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001871 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001872 "@FP16",
1873 "@pthreadpool",
1874 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001875)
1876
1877xnnpack_cc_library(
1878 name = "neonfma_ukernels",
1879 hdrs = INTERNAL_HDRS,
1880 aarch32_copts = [
1881 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07001882 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001883 "-mfpu=neon-vfpv4",
1884 ],
1885 aarch32_srcs = NEONFMA_UKERNELS,
1886 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001887 gcc_copts = xnnpack_gcc_std_copts(),
1888 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001889 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001890 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001891 "@FP16",
1892 "@pthreadpool",
1893 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001894)
1895
1896xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07001897 name = "neonv8_ukernels",
1898 hdrs = INTERNAL_HDRS,
1899 aarch32_copts = [
1900 "-marm",
1901 "-march=armv8-a",
1902 "-mfpu=neon-fp-armv8",
1903 ],
1904 aarch32_srcs = NEONV8_UKERNELS,
1905 aarch64_srcs = NEONV8_UKERNELS,
1906 gcc_copts = xnnpack_gcc_std_copts(),
1907 msvc_copts = xnnpack_msvc_std_copts(),
1908 deps = [
1909 ":tables",
1910 "@FP16",
1911 "@pthreadpool",
1912 ],
1913)
1914
1915xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916 name = "neonfp16arith_ukernels",
1917 hdrs = INTERNAL_HDRS,
1918 aarch64_copts = ["-march=armv8.2-a+fp16"],
1919 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001920 gcc_copts = xnnpack_gcc_std_copts(),
1921 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001922 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001923 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001924 "@FP16",
1925 "@pthreadpool",
1926 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001927)
1928
1929xnnpack_cc_library(
1930 name = "sse2_ukernels",
1931 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001932 gcc_copts = xnnpack_gcc_std_copts(),
1933 gcc_x86_copts = ["-msse2"],
1934 msvc_copts = xnnpack_msvc_std_copts(),
1935 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001936 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001937 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001938 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001939 "@FP16",
1940 "@pthreadpool",
1941 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001942)
1943
1944xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001945 name = "ssse3_ukernels",
1946 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001947 gcc_copts = xnnpack_gcc_std_copts(),
1948 gcc_x86_copts = ["-mssse3"],
1949 msvc_copts = xnnpack_msvc_std_copts(),
1950 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001951 x86_srcs = SSSE3_UKERNELS,
1952 deps = [
1953 ":tables",
1954 "@FP16",
1955 "@pthreadpool",
1956 ],
1957)
1958
1959xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001960 name = "sse41_ukernels",
1961 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001962 gcc_copts = xnnpack_gcc_std_copts(),
1963 gcc_x86_copts = ["-msse4.1"],
1964 msvc_copts = xnnpack_msvc_std_copts(),
1965 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001966 x86_srcs = SSE41_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 Dukhan69c3f2c2019-11-06 12:30:01 -08001972)
1973
1974xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001975 name = "avx_ukernels",
1976 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001977 gcc_copts = xnnpack_gcc_std_copts(),
1978 gcc_x86_copts = ["-mavx"],
1979 msvc_copts = xnnpack_msvc_std_copts(),
1980 msvc_x86_32_copts = ["/arch:AVX"],
1981 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001982 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001983 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001984 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001985 "@FP16",
1986 "@pthreadpool",
1987 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001988)
1989
1990xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001991 name = "fma3_ukernels",
1992 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001993 gcc_copts = xnnpack_gcc_std_copts(),
1994 gcc_x86_copts = ["-mfma"],
1995 msvc_copts = xnnpack_msvc_std_copts(),
1996 msvc_x86_32_copts = ["/arch:AVX"],
1997 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08001998 x86_srcs = FMA3_UKERNELS,
1999 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002000 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002001 "@FP16",
2002 "@pthreadpool",
2003 ],
2004)
2005
2006xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002007 name = "avx2_ukernels",
2008 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002009 gcc_copts = xnnpack_gcc_std_copts(),
2010 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002011 "-mfma",
2012 "-mavx2",
2013 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002014 msvc_copts = xnnpack_msvc_std_copts(),
2015 msvc_x86_32_copts = ["/arch:AVX2"],
2016 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002017 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002018 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002019 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002020 "@FP16",
2021 "@pthreadpool",
2022 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002023)
2024
2025xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002026 name = "avx512f_ukernels",
2027 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002028 gcc_copts = xnnpack_gcc_std_copts(),
2029 gcc_x86_copts = ["-mavx512f"],
2030 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2031 msvc_copts = xnnpack_msvc_std_copts(),
2032 msvc_x86_32_copts = ["/arch:AVX512"],
2033 msvc_x86_64_copts = ["/arch:AVX512"],
2034 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002035 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002036 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002037 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002038 "@FP16",
2039 "@pthreadpool",
2040 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002041)
2042
2043xnnpack_cc_library(
2044 name = "asm_ukernels",
2045 hdrs = ["src/xnnpack/assembly.h"],
2046 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002047 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002048 aarch64_srcs = AARCH64_ASM_UKERNELS,
2049)
2050
2051xnnpack_aggregate_library(
2052 name = "ukernels",
2053 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 ":neon_ukernels",
2055 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002056 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057 ":asm_ukernels",
2058 ],
2059 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002060 ":neon_ukernels",
2061 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002062 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002063 ":neonfp16arith_ukernels",
2064 ":asm_ukernels",
2065 ],
2066 generic_deps = [":scalar_ukernels"],
Marat Dukhan500b8892020-04-15 17:09:50 -07002067 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002068 ":psimd_fastmath_ukernels",
2069 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002071 wasm_deps = [":wasm_ukernels"],
2072 wasmsimd_deps = [":wasm_ukernels"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002073 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002074 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002075 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002076 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002077 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002078 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002079 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080 ":avx512f_ukernels",
2081 ],
2082)
2083
2084xnnpack_cc_library(
2085 name = "im2col",
2086 srcs = ["src/im2col.c"],
2087 hdrs = [
2088 "src/xnnpack/common.h",
2089 "src/xnnpack/im2col.h",
2090 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002091 gcc_copts = xnnpack_gcc_std_copts(),
2092 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093)
2094
2095xnnpack_cc_library(
2096 name = "indirection",
2097 srcs = ["src/indirection.c"],
2098 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002099 gcc_copts = xnnpack_gcc_std_copts(),
2100 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002101 deps = [
2102 "@FP16",
2103 "@FXdiv",
2104 "@pthreadpool",
2105 ],
2106)
2107
2108xnnpack_cc_library(
2109 name = "operator_run",
2110 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002111 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002112 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002113 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2114 "//conditions:default": [],
2115 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002116 gcc_copts = xnnpack_gcc_std_copts(),
2117 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118 deps = [
2119 "@FP16",
2120 "@FXdiv",
2121 "@clog",
2122 "@pthreadpool",
2123 ],
2124)
2125
Chao Mei6ddfc602020-05-13 22:29:36 -07002126xnnpack_cc_library(
2127 name = "memory_planner",
2128 srcs = ["src/memory-planner.c"],
2129 hdrs = INTERNAL_HDRS,
2130 defines = select({
2131 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2132 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2133 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2134 }),
2135 gcc_copts = xnnpack_gcc_std_copts(),
2136 msvc_copts = xnnpack_msvc_std_copts(),
2137 deps = [
2138 "@pthreadpool",
2139 ],
2140)
2141
Marat Dukhan08c4a432019-10-03 09:29:21 -07002142cc_library(
2143 name = "enable_assembly",
2144 defines = select({
2145 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2146 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002147 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002148 }),
2149)
2150
Marat Dukhancf056b22019-10-07 10:26:29 -07002151xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002152 name = "operators",
2153 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002154 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002155 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002156 ],
2157 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002158 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002159 "-Isrc",
2160 "-Iinclude",
2161 ] + select({
2162 ":debug_build": [],
2163 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002164 }) + select({
2165 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2166 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002167 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002168 gcc_copts = xnnpack_gcc_std_copts(),
2169 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002170 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002171 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002172 "@FP16",
2173 "@FXdiv",
2174 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002175 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002176 ],
2177)
2178
Marat Dukhan10a38082020-04-17 03:58:35 -07002179xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002180 name = "XNNPACK",
2181 srcs = [
2182 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002183 "src/runtime.c",
2184 "src/subgraph.c",
2185 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002186 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002187 hdrs = ["include/xnnpack.h"],
2188 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002189 "-Isrc",
2190 "-Iinclude",
2191 ] + select({
2192 ":debug_build": [],
2193 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002194 }) + select({
2195 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2196 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002197 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002198 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002199 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002200 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002201 visibility = xnnpack_visibility(),
2202 deps = [
2203 ":enable_assembly",
Chao Mei6ddfc602020-05-13 22:29:36 -07002204 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002205 ":ukernels",
2206 ":operator_run",
2207 ":operators",
2208 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07002209 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002210 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002211 ] + select({
2212 ":emscripten": [],
2213 "//conditions:default": ["@cpuinfo"],
2214 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002215)
2216
Marat Dukhan10a38082020-04-17 03:58:35 -07002217xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002218 name = "xnnpack_operators_nhwc_f32",
2219 srcs = [
2220 "src/init.c",
2221 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002222 hdrs = ["include/xnnpack.h"],
2223 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002224 "-Isrc",
2225 "-Iinclude",
2226 ] + select({
2227 ":debug_build": [],
2228 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002229 }) + select({
2230 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2231 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002232 }),
2233 defines = [
2234 "XNN_NO_Q8_OPERATORS",
2235 "XNN_NO_U8_OPERATORS",
2236 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002237 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002238 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002239 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002240 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002241 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002242 visibility = xnnpack_visibility(),
2243 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002244 ":enable_assembly",
2245 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002246 ":operator_run",
2247 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002248 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002250 ] + select({
2251 ":emscripten": [],
2252 "//conditions:default": ["@cpuinfo"],
2253 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002254)
2255
Marat Dukhancf056b22019-10-07 10:26:29 -07002256xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257 name = "bench_utils",
2258 srcs = ["bench/utils.cc"],
2259 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002260 deps = [
2261 "@com_google_benchmark//:benchmark",
2262 "@cpuinfo",
2263 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264)
2265
Frank Barchard7e955972019-10-11 10:34:25 -07002266######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002267
2268xnnpack_benchmark(
2269 name = "q8_gemm_bench",
2270 srcs = [
2271 "bench/gemm.h",
2272 "bench/q8-gemm.cc",
2273 "src/xnnpack/AlignedAllocator.h",
2274 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002275 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002276 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2277)
2278
2279xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002280 name = "f16_igemm_bench",
2281 srcs = [
2282 "bench/f16-igemm.cc",
2283 "bench/conv.h",
2284 "bench/google/conv.h",
2285 "src/xnnpack/AlignedAllocator.h",
2286 ] + MICROKERNEL_BENCHMARK_HDRS,
2287 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2288)
2289
2290xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291 name = "f16_gemm_bench",
2292 srcs = [
2293 "bench/f16-gemm.cc",
2294 "bench/gemm.h",
2295 "src/xnnpack/AlignedAllocator.h",
2296 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002297 deps = MICROKERNEL_BENCHMARK_DEPS,
2298)
2299
2300xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002301 name = "f16_spmm_bench",
2302 srcs = [
2303 "bench/f16-spmm.cc",
2304 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002305 "src/xnnpack/AlignedAllocator.h",
2306 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002307 deps = MICROKERNEL_BENCHMARK_DEPS,
2308)
2309
2310xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002311 name = "f32_igemm_bench",
2312 srcs = [
2313 "bench/f32-igemm.cc",
2314 "bench/conv.h",
2315 "src/xnnpack/AlignedAllocator.h",
2316 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002317 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002318)
2319
2320xnnpack_benchmark(
2321 name = "f32_conv_hwc_bench",
2322 srcs = [
2323 "bench/f32-conv-hwc.cc",
2324 "bench/dconv.h",
2325 "src/xnnpack/AlignedAllocator.h",
2326 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002327 deps = MICROKERNEL_BENCHMARK_DEPS,
2328)
2329
2330xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002331 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07002332 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002333 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07002334 "bench/dconv.h",
2335 "src/xnnpack/AlignedAllocator.h",
2336 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07002337 deps = MICROKERNEL_BENCHMARK_DEPS,
2338)
2339
2340xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 name = "f32_dwconv_bench",
2342 srcs = [
2343 "bench/f32-dwconv.cc",
2344 "bench/dwconv.h",
2345 "src/xnnpack/AlignedAllocator.h",
2346 ] + MICROKERNEL_BENCHMARK_HDRS,
2347 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2348)
2349
2350xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002351 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002353 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002354 "bench/dwconv.h",
2355 "src/xnnpack/AlignedAllocator.h",
2356 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2357 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2358)
2359
2360xnnpack_benchmark(
2361 name = "f32_gemm_bench",
2362 srcs = [
2363 "bench/f32-gemm.cc",
2364 "bench/gemm.h",
2365 "src/xnnpack/AlignedAllocator.h",
2366 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002367 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002368 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002369)
2370
2371xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002372 name = "f32_raddexpminusmax_bench",
2373 srcs = [
2374 "bench/f32-raddexpminusmax.cc",
2375 "src/xnnpack/AlignedAllocator.h",
2376 ] + MICROKERNEL_BENCHMARK_HDRS,
2377 deps = MICROKERNEL_BENCHMARK_DEPS,
2378)
2379
2380xnnpack_benchmark(
2381 name = "f32_raddextexp_bench",
2382 srcs = [
2383 "bench/f32-raddextexp.cc",
2384 "src/xnnpack/AlignedAllocator.h",
2385 ] + MICROKERNEL_BENCHMARK_HDRS,
2386 deps = MICROKERNEL_BENCHMARK_DEPS,
2387)
2388
2389xnnpack_benchmark(
2390 name = "f32_raddstoreexpminusmax_bench",
2391 srcs = [
2392 "bench/f32-raddstoreexpminusmax.cc",
2393 "src/xnnpack/AlignedAllocator.h",
2394 ] + MICROKERNEL_BENCHMARK_HDRS,
2395 deps = MICROKERNEL_BENCHMARK_DEPS,
2396)
2397
2398xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002399 name = "f32_rmax_bench",
2400 srcs = [
2401 "bench/f32-rmax.cc",
2402 "src/xnnpack/AlignedAllocator.h",
2403 ] + MICROKERNEL_BENCHMARK_HDRS,
2404 deps = MICROKERNEL_BENCHMARK_DEPS,
2405)
2406
2407xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002408 name = "f32_sigmoid_bench",
2409 srcs = [
2410 "bench/f32-sigmoid.cc",
2411 "src/xnnpack/AlignedAllocator.h",
2412 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08002413 deps = MICROKERNEL_BENCHMARK_DEPS,
2414)
2415
2416xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002417 name = "f32_spmm_bench",
2418 srcs = [
2419 "bench/f32-spmm.cc",
2420 "bench/gemm.h",
2421 "src/xnnpack/AlignedAllocator.h",
2422 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002423 deps = MICROKERNEL_BENCHMARK_DEPS,
2424)
2425
2426xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002427 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002428 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002429 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002430 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002431 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08002432 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002433)
2434
2435xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002436 name = "f32_vscaleexpminusmax_bench",
2437 srcs = [
2438 "bench/f32-vscaleexpminusmax.cc",
2439 "src/xnnpack/AlignedAllocator.h",
2440 ] + MICROKERNEL_BENCHMARK_HDRS,
2441 deps = MICROKERNEL_BENCHMARK_DEPS,
2442)
2443
2444xnnpack_benchmark(
2445 name = "f32_vscaleextexp_bench",
2446 srcs = [
2447 "bench/f32-vscaleextexp.cc",
2448 "src/xnnpack/AlignedAllocator.h",
2449 ] + MICROKERNEL_BENCHMARK_HDRS,
2450 deps = MICROKERNEL_BENCHMARK_DEPS,
2451)
2452
2453xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454 name = "f32_im2col_gemm_bench",
2455 srcs = [
2456 "bench/f32-im2col-gemm.cc",
2457 "bench/conv.h",
2458 "src/xnnpack/AlignedAllocator.h",
2459 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2460 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2461)
2462
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002463xnnpack_benchmark(
2464 name = "requantization_bench",
2465 srcs = [
2466 "bench/requantization.cc",
2467 "src/xnnpack/requantization-stubs.h",
2468 "src/xnnpack/AlignedAllocator.h",
2469 ] + MICROKERNEL_BENCHMARK_HDRS,
2470 deps = MICROKERNEL_BENCHMARK_DEPS,
2471)
2472
Marat Dukhanffbf96a2020-05-14 02:59:08 -07002473xnnpack_benchmark(
2474 name = "rounding_bench",
2475 srcs = [
2476 "bench/rounding.cc",
2477 "src/xnnpack/math-stubs.h",
2478 "src/xnnpack/AlignedAllocator.h",
2479 ] + MICROKERNEL_BENCHMARK_HDRS,
2480 deps = MICROKERNEL_BENCHMARK_DEPS,
2481)
2482
Marat Dukhan08c4a432019-10-03 09:29:21 -07002483########################### Benchmarks for operators ###########################
2484
2485xnnpack_benchmark(
2486 name = "add_bench",
2487 srcs = ["bench/add.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002488 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002489)
2490
2491xnnpack_benchmark(
2492 name = "average_pooling_bench",
2493 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07002494 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002495 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002496 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002497)
2498
2499xnnpack_benchmark(
2500 name = "channel_shuffle_bench",
2501 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002502 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002503)
2504
2505xnnpack_benchmark(
2506 name = "convolution_bench",
2507 srcs = ["bench/convolution.cc"],
2508 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002509 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002510 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002511)
2512
2513xnnpack_benchmark(
2514 name = "deconvolution_bench",
2515 srcs = ["bench/deconvolution.cc"],
2516 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002517 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002518 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002519)
2520
2521xnnpack_benchmark(
2522 name = "global_average_pooling_bench",
2523 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002524 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002525)
2526
2527xnnpack_benchmark(
2528 name = "max_pooling_bench",
2529 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002530 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002531)
2532
2533xnnpack_benchmark(
2534 name = "sigmoid_bench",
2535 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002536 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002537 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002538 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002539)
2540
2541xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002542 name = "prelu_bench",
2543 srcs = ["bench/prelu.cc"],
2544 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07002545 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002546 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07002547)
2548
2549xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002550 name = "softmax_bench",
2551 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002552 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07002553 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07002554 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002555)
2556
Marat Dukhanc068bb62019-10-04 13:24:39 -07002557############################# End-to-end benchmarks ############################
2558
2559cc_library(
2560 name = "mobilenet_v1",
2561 srcs = ["models/mobilenet-v1.cc"],
2562 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002563 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002564 linkstatic = True,
2565 deps = [
2566 ":XNNPACK",
2567 "@pthreadpool",
2568 ],
2569)
2570
2571cc_library(
2572 name = "mobilenet_v2",
2573 srcs = ["models/mobilenet-v2.cc"],
2574 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002575 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002576 linkstatic = True,
2577 deps = [
2578 ":XNNPACK",
2579 "@pthreadpool",
2580 ],
2581)
2582
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002583cc_library(
2584 name = "mobilenet_v3_large",
2585 srcs = ["models/mobilenet-v3-large.cc"],
2586 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002587 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002588 linkstatic = True,
2589 deps = [
2590 ":XNNPACK",
2591 "@pthreadpool",
2592 ],
2593)
2594
2595cc_library(
2596 name = "mobilenet_v3_small",
2597 srcs = ["models/mobilenet-v3-small.cc"],
2598 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002599 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002600 linkstatic = True,
2601 deps = [
2602 ":XNNPACK",
2603 "@pthreadpool",
2604 ],
2605)
2606
Marat Dukhanc068bb62019-10-04 13:24:39 -07002607xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002608 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002609 srcs = [
2610 "bench/f32-dwconv-e2e.cc",
2611 "bench/end2end.h",
2612 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002613 deps = MICROKERNEL_BENCHMARK_DEPS + [
2614 ":XNNPACK",
2615 ":mobilenet_v1",
2616 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002617 ":mobilenet_v3_large",
2618 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002619 ],
2620)
2621
2622xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002623 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002624 srcs = [
2625 "bench/f32-gemm-e2e.cc",
2626 "bench/end2end.h",
2627 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002628 deps = MICROKERNEL_BENCHMARK_DEPS + [
2629 ":XNNPACK",
2630 ":mobilenet_v1",
2631 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002632 ":mobilenet_v3_large",
2633 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002634 ],
2635)
2636
2637xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002638 name = "end2end_bench",
2639 srcs = ["bench/end2end.cc"],
2640 deps = [
2641 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002642 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002643 ":mobilenet_v1",
2644 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002645 ":mobilenet_v3_large",
2646 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002647 "@pthreadpool",
2648 ],
2649)
2650
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002651#################### Accuracy evaluation for math functions ####################
2652
2653xnnpack_benchmark(
2654 name = "f32_exp_eval",
2655 srcs = [
2656 "eval/f32-exp.cc",
2657 "src/xnnpack/AlignedAllocator.h",
2658 ] + ACCURACY_EVAL_HDRS,
2659 deps = ACCURACY_EVAL_DEPS,
2660)
2661
Marat Dukhan515c9772019-10-17 18:07:57 -07002662xnnpack_benchmark(
2663 name = "f32_expminus_eval",
2664 srcs = [
2665 "eval/f32-expminus.cc",
2666 "src/xnnpack/AlignedAllocator.h",
2667 ] + ACCURACY_EVAL_HDRS,
2668 deps = ACCURACY_EVAL_DEPS,
2669)
2670
Marat Dukhan98ba4412019-10-23 02:14:28 -07002671xnnpack_benchmark(
2672 name = "f32_extexp_eval",
2673 srcs = [
2674 "eval/f32-extexp.cc",
2675 "src/xnnpack/AlignedAllocator.h",
2676 ] + ACCURACY_EVAL_HDRS,
2677 deps = ACCURACY_EVAL_DEPS,
2678)
2679
Marat Dukhan8853b822020-05-07 12:19:01 -07002680xnnpack_unit_test(
2681 name = "f32_roundne_eval",
2682 srcs = [
2683 "eval/f32-roundne.cc",
2684 "src/xnnpack/AlignedAllocator.h",
2685 "src/xnnpack/math-stubs.h",
2686 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07002687 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07002688 deps = MICROKERNEL_TEST_DEPS,
2689)
2690
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002691xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002692 name = "f32_roundd_eval",
2693 srcs = [
2694 "eval/f32-roundd.cc",
2695 "src/xnnpack/AlignedAllocator.h",
2696 "src/xnnpack/math-stubs.h",
2697 ] + MICROKERNEL_TEST_HDRS,
2698 automatic = False,
2699 deps = MICROKERNEL_TEST_DEPS,
2700)
2701
2702xnnpack_unit_test(
2703 name = "f32_roundu_eval",
2704 srcs = [
2705 "eval/f32-roundu.cc",
2706 "src/xnnpack/AlignedAllocator.h",
2707 "src/xnnpack/math-stubs.h",
2708 ] + MICROKERNEL_TEST_HDRS,
2709 automatic = False,
2710 deps = MICROKERNEL_TEST_DEPS,
2711)
2712
2713xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002714 name = "f32_roundz_eval",
2715 srcs = [
2716 "eval/f32-roundz.cc",
2717 "src/xnnpack/AlignedAllocator.h",
2718 "src/xnnpack/math-stubs.h",
2719 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002720 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002721 deps = MICROKERNEL_TEST_DEPS,
2722)
2723
Marat Dukhan346a9e52019-11-15 09:06:30 -08002724xnnpack_benchmark(
2725 name = "f32_sigmoid_eval",
2726 srcs = [
2727 "eval/f32-sigmoid.cc",
2728 "src/xnnpack/AlignedAllocator.h",
2729 ] + ACCURACY_EVAL_HDRS,
2730 deps = ACCURACY_EVAL_DEPS,
2731)
2732
Marat Dukhan08c4a432019-10-03 09:29:21 -07002733######################### Unit tests for micro-kernels #########################
2734
2735xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07002736 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07002738 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002739 "test/gemm-microkernel-tester.h",
2740 "src/xnnpack/AlignedAllocator.h",
2741 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2742 deps = MICROKERNEL_TEST_DEPS,
2743)
2744
2745xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07002746 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002747 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07002748 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002749 "test/spmm-microkernel-tester.h",
2750 "src/xnnpack/AlignedAllocator.h",
2751 ] + MICROKERNEL_TEST_HDRS,
2752 deps = MICROKERNEL_TEST_DEPS,
2753)
2754
2755xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002756 name = "f32_argmaxpool_test",
2757 srcs = [
2758 "test/f32-argmaxpool.cc",
2759 "test/argmaxpool-microkernel-tester.h",
2760 "src/xnnpack/AlignedAllocator.h",
2761 ] + MICROKERNEL_TEST_HDRS,
2762 deps = MICROKERNEL_TEST_DEPS,
2763)
2764
2765xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002766 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002768 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 "test/avgpool-microkernel-tester.h",
2770 "src/xnnpack/AlignedAllocator.h",
2771 ] + MICROKERNEL_TEST_HDRS,
2772 deps = MICROKERNEL_TEST_DEPS,
2773)
2774
2775xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07002776 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002777 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07002778 "test/f32-ibilinear.cc",
2779 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002780 "src/xnnpack/AlignedAllocator.h",
2781 ] + MICROKERNEL_TEST_HDRS,
2782 deps = MICROKERNEL_TEST_DEPS,
2783)
2784
2785xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002786 name = "f16_clamp_test",
2787 srcs = [
2788 "test/f16-clamp.cc",
2789 "test/clamp-microkernel-tester.h",
2790 ] + MICROKERNEL_TEST_HDRS,
2791 deps = MICROKERNEL_TEST_DEPS,
2792)
2793
2794xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 name = "f32_clamp_test",
2796 srcs = [
2797 "test/f32-clamp.cc",
2798 "test/clamp-microkernel-tester.h",
2799 ] + MICROKERNEL_TEST_HDRS,
2800 deps = MICROKERNEL_TEST_DEPS,
2801)
2802
2803xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002804 name = "f32_igemm_test",
2805 srcs = [
2806 "test/f32-igemm.cc",
2807 "test/gemm-microkernel-tester.h",
2808 "src/xnnpack/AlignedAllocator.h",
2809 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2810 deps = MICROKERNEL_TEST_DEPS,
2811)
2812
2813xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002814 name = "f32_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002815 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002816 "test/f32-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002817 "test/gemm-microkernel-tester.h",
2818 "src/xnnpack/AlignedAllocator.h",
2819 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002820 # Work-around for "too many sections" error
2821 mingw_copts = ["-Wa,-mbig-obj"],
2822 msys_copts = ["-Wa,-mbig-obj"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002823 deps = MICROKERNEL_TEST_DEPS,
2824)
2825
2826xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07002827 name = "f16_igemm_minmax_test",
2828 srcs = [
2829 "test/f16-igemm-minmax.cc",
2830 "test/gemm-microkernel-tester.h",
2831 "src/xnnpack/AlignedAllocator.h",
2832 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2833 deps = MICROKERNEL_TEST_DEPS,
2834)
2835
2836xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837 name = "f32_conv_hwc_test",
2838 srcs = [
2839 "test/f32-conv-hwc.cc",
2840 "test/conv-hwc-microkernel-tester.h",
2841 "src/xnnpack/AlignedAllocator.h",
2842 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2843 deps = MICROKERNEL_TEST_DEPS,
2844)
2845
2846xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002847 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002848 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002849 "test/f32-conv-hwc2chw.cc",
2850 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002851 "src/xnnpack/AlignedAllocator.h",
2852 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2853 deps = MICROKERNEL_TEST_DEPS,
2854)
2855
2856xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002857 name = "f32_dwconv_test",
2858 srcs = [
2859 "test/f32-dwconv.cc",
2860 "test/dwconv-microkernel-tester.h",
2861 "src/xnnpack/AlignedAllocator.h",
2862 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2863 deps = MICROKERNEL_TEST_DEPS,
2864)
2865
2866xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002867 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002868 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002869 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002870 "test/dwconv-microkernel-tester.h",
2871 "src/xnnpack/AlignedAllocator.h",
2872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2873 deps = MICROKERNEL_TEST_DEPS,
2874)
2875
2876xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002877 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002879 "test/f32-dwconv-chw.cc",
2880 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881 "src/xnnpack/AlignedAllocator.h",
2882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2883 deps = MICROKERNEL_TEST_DEPS,
2884)
2885
2886xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002887 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002888 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002889 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 "test/gavgpool-microkernel-tester.h",
2891 "src/xnnpack/AlignedAllocator.h",
2892 ] + MICROKERNEL_TEST_HDRS,
2893 deps = MICROKERNEL_TEST_DEPS,
2894)
2895
2896xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002897 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002899 "test/f32-gavgpool-cw.cc",
2900 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 "src/xnnpack/AlignedAllocator.h",
2902 ] + MICROKERNEL_TEST_HDRS,
2903 deps = MICROKERNEL_TEST_DEPS,
2904)
2905
2906xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07002907 name = "f32_gemm_test",
2908 srcs = [
2909 "test/f32-gemm.cc",
2910 "test/gemm-microkernel-tester.h",
2911 "src/xnnpack/AlignedAllocator.h",
2912 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2913 deps = MICROKERNEL_TEST_DEPS,
2914)
2915
2916xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002917 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002918 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002919 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002920 "test/gemm-microkernel-tester.h",
2921 "src/xnnpack/AlignedAllocator.h",
2922 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2923 deps = MICROKERNEL_TEST_DEPS,
2924)
2925
2926xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002927 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002929 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 "test/gemm-microkernel-tester.h",
2931 "src/xnnpack/AlignedAllocator.h",
2932 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2933 deps = MICROKERNEL_TEST_DEPS,
2934)
2935
2936xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002937 name = "f16_hswish_test",
2938 srcs = [
2939 "test/f16-hswish.cc",
2940 "test/hswish-microkernel-tester.h",
2941 ] + MICROKERNEL_TEST_HDRS,
2942 deps = MICROKERNEL_TEST_DEPS,
2943)
2944
2945xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 name = "f32_hswish_test",
2947 srcs = [
2948 "test/f32-hswish.cc",
2949 "test/hswish-microkernel-tester.h",
2950 ] + MICROKERNEL_TEST_HDRS,
2951 deps = MICROKERNEL_TEST_DEPS,
2952)
2953
2954xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002955 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002956 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002957 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958 "test/maxpool-microkernel-tester.h",
2959 ] + MICROKERNEL_TEST_HDRS,
2960 deps = MICROKERNEL_TEST_DEPS,
2961)
2962
2963xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07002964 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002965 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07002966 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 "test/avgpool-microkernel-tester.h",
2968 "src/xnnpack/AlignedAllocator.h",
2969 ] + MICROKERNEL_TEST_HDRS,
2970 deps = MICROKERNEL_TEST_DEPS,
2971)
2972
2973xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07002974 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002975 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002976 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 "test/gemm-microkernel-tester.h",
2978 "src/xnnpack/AlignedAllocator.h",
2979 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2980 deps = MICROKERNEL_TEST_DEPS,
2981)
2982
2983xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07002984 name = "f16_prelu_test",
2985 srcs = [
2986 "test/f16-prelu.cc",
2987 "test/prelu-microkernel-tester.h",
2988 "src/xnnpack/AlignedAllocator.h",
2989 ] + MICROKERNEL_TEST_HDRS,
2990 deps = MICROKERNEL_TEST_DEPS,
2991)
2992
2993xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002994 name = "f32_prelu_test",
2995 srcs = [
2996 "test/f32-prelu.cc",
2997 "test/prelu-microkernel-tester.h",
2998 "src/xnnpack/AlignedAllocator.h",
2999 ] + MICROKERNEL_TEST_HDRS,
3000 deps = MICROKERNEL_TEST_DEPS,
3001)
3002
3003xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003004 name = "f32_raddexpminusmax_test",
3005 srcs = [
3006 "test/f32-raddexpminusmax.cc",
3007 "test/raddexpminusmax-microkernel-tester.h",
3008 ] + MICROKERNEL_TEST_HDRS,
3009 deps = MICROKERNEL_TEST_DEPS,
3010)
3011
3012xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003013 name = "f32_raddextexp_test",
3014 srcs = [
3015 "test/f32-raddextexp.cc",
3016 "test/raddextexp-microkernel-tester.h",
3017 ] + MICROKERNEL_TEST_HDRS,
3018 deps = MICROKERNEL_TEST_DEPS,
3019)
3020
3021xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003022 name = "f32_raddstoreexpminusmax_test",
3023 srcs = [
3024 "test/f32-raddstoreexpminusmax.cc",
3025 "test/raddstoreexpminusmax-microkernel-tester.h",
3026 ] + MICROKERNEL_TEST_HDRS,
3027 deps = MICROKERNEL_TEST_DEPS,
3028)
3029
3030xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003031 name = "f32_rmax_test",
3032 srcs = [
3033 "test/f32-rmax.cc",
3034 "test/rmax-microkernel-tester.h",
3035 ] + MICROKERNEL_TEST_HDRS,
3036 deps = MICROKERNEL_TEST_DEPS,
3037)
3038
3039xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003040 name = "f32_sigmoid_test",
3041 srcs = [
3042 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003043 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003044 ] + MICROKERNEL_TEST_HDRS,
3045 deps = MICROKERNEL_TEST_DEPS,
3046)
3047
3048xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003049 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003050 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003051 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 "test/spmm-microkernel-tester.h",
3053 "src/xnnpack/AlignedAllocator.h",
3054 ] + MICROKERNEL_TEST_HDRS,
3055 deps = MICROKERNEL_TEST_DEPS,
3056)
3057
3058xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003059 name = "f16_vadd_minmax_test",
3060 srcs = [
3061 "test/f16-vadd-minmax.cc",
3062 "test/vbinary-microkernel-tester.h",
3063 ] + MICROKERNEL_TEST_HDRS,
3064 deps = MICROKERNEL_TEST_DEPS,
3065)
3066
3067xnnpack_unit_test(
3068 name = "f16_vaddc_minmax_test",
3069 srcs = [
3070 "test/f16-vaddc-minmax.cc",
3071 "test/vbinaryc-microkernel-tester.h",
3072 ] + MICROKERNEL_TEST_HDRS,
3073 deps = MICROKERNEL_TEST_DEPS,
3074)
3075
3076xnnpack_unit_test(
3077 name = "f16_vdiv_minmax_test",
3078 srcs = [
3079 "test/f16-vdiv-minmax.cc",
3080 "test/vbinary-microkernel-tester.h",
3081 ] + MICROKERNEL_TEST_HDRS,
3082 deps = MICROKERNEL_TEST_DEPS,
3083)
3084
3085xnnpack_unit_test(
3086 name = "f16_vdivc_minmax_test",
3087 srcs = [
3088 "test/f16-vdivc-minmax.cc",
3089 "test/vbinaryc-microkernel-tester.h",
3090 ] + MICROKERNEL_TEST_HDRS,
3091 deps = MICROKERNEL_TEST_DEPS,
3092)
3093
3094xnnpack_unit_test(
3095 name = "f16_vrdivc_minmax_test",
3096 srcs = [
3097 "test/f16-vrdivc-minmax.cc",
3098 "test/vbinaryc-microkernel-tester.h",
3099 ] + MICROKERNEL_TEST_HDRS,
3100 deps = MICROKERNEL_TEST_DEPS,
3101)
3102
3103xnnpack_unit_test(
3104 name = "f16_vmax_test",
3105 srcs = [
3106 "test/f16-vmax.cc",
3107 "test/vbinary-microkernel-tester.h",
3108 ] + MICROKERNEL_TEST_HDRS,
3109 deps = MICROKERNEL_TEST_DEPS,
3110)
3111
3112xnnpack_unit_test(
3113 name = "f16_vmaxc_test",
3114 srcs = [
3115 "test/f16-vmaxc.cc",
3116 "test/vbinaryc-microkernel-tester.h",
3117 ] + MICROKERNEL_TEST_HDRS,
3118 deps = MICROKERNEL_TEST_DEPS,
3119)
3120
3121xnnpack_unit_test(
3122 name = "f16_vmin_test",
3123 srcs = [
3124 "test/f16-vmin.cc",
3125 "test/vbinary-microkernel-tester.h",
3126 ] + MICROKERNEL_TEST_HDRS,
3127 deps = MICROKERNEL_TEST_DEPS,
3128)
3129
3130xnnpack_unit_test(
3131 name = "f16_vminc_test",
3132 srcs = [
3133 "test/f16-vminc.cc",
3134 "test/vbinaryc-microkernel-tester.h",
3135 ] + MICROKERNEL_TEST_HDRS,
3136 deps = MICROKERNEL_TEST_DEPS,
3137)
3138
3139xnnpack_unit_test(
3140 name = "f16_vmul_minmax_test",
3141 srcs = [
3142 "test/f16-vmul-minmax.cc",
3143 "test/vbinary-microkernel-tester.h",
3144 ] + MICROKERNEL_TEST_HDRS,
3145 deps = MICROKERNEL_TEST_DEPS,
3146)
3147
3148xnnpack_unit_test(
3149 name = "f16_vmulc_minmax_test",
3150 srcs = [
3151 "test/f16-vmulc-minmax.cc",
3152 "test/vbinaryc-microkernel-tester.h",
3153 ] + MICROKERNEL_TEST_HDRS,
3154 deps = MICROKERNEL_TEST_DEPS,
3155)
3156
3157xnnpack_unit_test(
3158 name = "f16_vsub_minmax_test",
3159 srcs = [
3160 "test/f16-vsub-minmax.cc",
3161 "test/vbinary-microkernel-tester.h",
3162 ] + MICROKERNEL_TEST_HDRS,
3163 deps = MICROKERNEL_TEST_DEPS,
3164)
3165
3166xnnpack_unit_test(
3167 name = "f16_vsubc_minmax_test",
3168 srcs = [
3169 "test/f16-vsubc-minmax.cc",
3170 "test/vbinaryc-microkernel-tester.h",
3171 ] + MICROKERNEL_TEST_HDRS,
3172 deps = MICROKERNEL_TEST_DEPS,
3173)
3174
3175xnnpack_unit_test(
3176 name = "f16_vrsubc_minmax_test",
3177 srcs = [
3178 "test/f16-vrsubc-minmax.cc",
3179 "test/vbinaryc-microkernel-tester.h",
3180 ] + MICROKERNEL_TEST_HDRS,
3181 deps = MICROKERNEL_TEST_DEPS,
3182)
3183
3184xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003185 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003186 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003187 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003188 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003189 ] + MICROKERNEL_TEST_HDRS,
3190 deps = MICROKERNEL_TEST_DEPS,
3191)
3192
3193xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003194 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003195 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003196 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003197 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 ] + MICROKERNEL_TEST_HDRS,
3199 deps = MICROKERNEL_TEST_DEPS,
3200)
3201
3202xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003203 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003204 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003205 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003206 "test/vbinary-microkernel-tester.h",
3207 ] + MICROKERNEL_TEST_HDRS,
3208 deps = MICROKERNEL_TEST_DEPS,
3209)
3210
3211xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003212 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003213 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003214 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003215 "test/vbinaryc-microkernel-tester.h",
3216 ] + MICROKERNEL_TEST_HDRS,
3217 deps = MICROKERNEL_TEST_DEPS,
3218)
3219
3220xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003221 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003222 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003223 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003224 "test/vbinaryc-microkernel-tester.h",
3225 ] + MICROKERNEL_TEST_HDRS,
3226 deps = MICROKERNEL_TEST_DEPS,
3227)
3228
3229xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003230 name = "f32_vmax_test",
3231 srcs = [
3232 "test/f32-vmax.cc",
3233 "test/vbinary-microkernel-tester.h",
3234 ] + MICROKERNEL_TEST_HDRS,
3235 deps = MICROKERNEL_TEST_DEPS,
3236)
3237
3238xnnpack_unit_test(
3239 name = "f32_vmaxc_test",
3240 srcs = [
3241 "test/f32-vmaxc.cc",
3242 "test/vbinaryc-microkernel-tester.h",
3243 ] + MICROKERNEL_TEST_HDRS,
3244 deps = MICROKERNEL_TEST_DEPS,
3245)
3246
3247xnnpack_unit_test(
3248 name = "f32_vmin_test",
3249 srcs = [
3250 "test/f32-vmin.cc",
3251 "test/vbinary-microkernel-tester.h",
3252 ] + MICROKERNEL_TEST_HDRS,
3253 deps = MICROKERNEL_TEST_DEPS,
3254)
3255
3256xnnpack_unit_test(
3257 name = "f32_vminc_test",
3258 srcs = [
3259 "test/f32-vminc.cc",
3260 "test/vbinaryc-microkernel-tester.h",
3261 ] + MICROKERNEL_TEST_HDRS,
3262 deps = MICROKERNEL_TEST_DEPS,
3263)
3264
3265xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003266 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003267 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003268 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003269 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003270 ] + MICROKERNEL_TEST_HDRS,
3271 deps = MICROKERNEL_TEST_DEPS,
3272)
3273
3274xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003275 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003276 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003277 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003278 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003279 ] + MICROKERNEL_TEST_HDRS,
3280 deps = MICROKERNEL_TEST_DEPS,
3281)
3282
3283xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003284 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003285 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003286 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003287 "test/vmulcaddc-microkernel-tester.h",
3288 "src/xnnpack/AlignedAllocator.h",
3289 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3290 deps = MICROKERNEL_TEST_DEPS,
3291)
3292
3293xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003294 name = "f32_vscale_test",
3295 srcs = [
3296 "test/f32-vscale.cc",
3297 "test/vscale-microkernel-tester.h",
3298 ] + MICROKERNEL_TEST_HDRS,
3299 deps = MICROKERNEL_TEST_DEPS,
3300)
3301
3302xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003303 name = "f32_vscaleexpminusmax_test",
3304 srcs = [
3305 "test/f32-vscaleexpminusmax.cc",
3306 "test/vscaleexpminusmax-microkernel-tester.h",
3307 ] + MICROKERNEL_TEST_HDRS,
3308 deps = MICROKERNEL_TEST_DEPS,
3309)
3310
3311xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003312 name = "f32_vscaleextexp_test",
3313 srcs = [
3314 "test/f32-vscaleextexp.cc",
3315 "test/vscaleextexp-microkernel-tester.h",
3316 ] + MICROKERNEL_TEST_HDRS,
3317 deps = MICROKERNEL_TEST_DEPS,
3318)
3319
3320xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003321 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07003322 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003323 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003324 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003325 ] + MICROKERNEL_TEST_HDRS,
3326 deps = MICROKERNEL_TEST_DEPS,
3327)
3328
3329xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003330 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003331 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003332 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003333 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003334 ] + MICROKERNEL_TEST_HDRS,
3335 deps = MICROKERNEL_TEST_DEPS,
3336)
3337
3338xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003339 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003340 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003341 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003342 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003343 ] + MICROKERNEL_TEST_HDRS,
3344 deps = MICROKERNEL_TEST_DEPS,
3345)
3346
3347xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003348 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003349 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003350 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003351 "test/avgpool-microkernel-tester.h",
3352 "src/xnnpack/AlignedAllocator.h",
3353 ] + MICROKERNEL_TEST_HDRS,
3354 deps = MICROKERNEL_TEST_DEPS,
3355)
3356
3357xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003358 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003360 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361 "test/gemm-microkernel-tester.h",
3362 "src/xnnpack/AlignedAllocator.h",
3363 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3364 deps = MICROKERNEL_TEST_DEPS,
3365)
3366
3367xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003368 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003370 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003371 "test/dwconv-microkernel-tester.h",
3372 "src/xnnpack/AlignedAllocator.h",
3373 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3374 deps = MICROKERNEL_TEST_DEPS,
3375)
3376
3377xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003378 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003379 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003380 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381 "test/gavgpool-microkernel-tester.h",
3382 "src/xnnpack/AlignedAllocator.h",
3383 ] + MICROKERNEL_TEST_HDRS,
3384 deps = MICROKERNEL_TEST_DEPS,
3385)
3386
3387xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003388 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003389 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003390 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003391 "test/gemm-microkernel-tester.h",
3392 "src/xnnpack/AlignedAllocator.h",
3393 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3394 deps = MICROKERNEL_TEST_DEPS,
3395)
3396
3397xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003398 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003399 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003400 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401 "test/vadd-microkernel-tester.h",
3402 ] + MICROKERNEL_TEST_HDRS,
3403 deps = MICROKERNEL_TEST_DEPS,
3404)
3405
3406xnnpack_unit_test(
3407 name = "u8_clamp_test",
3408 srcs = [
3409 "test/u8-clamp.cc",
3410 "test/clamp-microkernel-tester.h",
3411 ] + MICROKERNEL_TEST_HDRS,
3412 deps = MICROKERNEL_TEST_DEPS,
3413)
3414
3415xnnpack_unit_test(
3416 name = "u8_lut32norm_test",
3417 srcs = [
3418 "test/u8-lut32norm.cc",
3419 "test/lut-norm-microkernel-tester.h",
3420 ] + MICROKERNEL_TEST_HDRS,
3421 deps = MICROKERNEL_TEST_DEPS,
3422)
3423
3424xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003425 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003426 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003427 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003428 "test/maxpool-microkernel-tester.h",
3429 ] + MICROKERNEL_TEST_HDRS,
3430 deps = MICROKERNEL_TEST_DEPS,
3431)
3432
3433xnnpack_unit_test(
3434 name = "u8_rmax_test",
3435 srcs = [
3436 "test/u8-rmax.cc",
3437 "test/rmax-microkernel-tester.h",
3438 ] + MICROKERNEL_TEST_HDRS,
3439 deps = MICROKERNEL_TEST_DEPS,
3440)
3441
3442xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07003443 name = "x32_fill_test",
3444 srcs = [
3445 "test/x32-fill.cc",
3446 "test/fill-microkernel-tester.h",
3447 ] + MICROKERNEL_TEST_HDRS,
3448 deps = MICROKERNEL_TEST_DEPS,
3449)
3450
3451xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452 name = "x32_packx_test",
3453 srcs = [
3454 "test/x32-packx.cc",
3455 "test/pack-microkernel-tester.h",
3456 "src/xnnpack/AlignedAllocator.h",
3457 ] + MICROKERNEL_TEST_HDRS,
3458 deps = MICROKERNEL_TEST_DEPS,
3459)
3460
3461xnnpack_unit_test(
3462 name = "x32_pad_test",
3463 srcs = [
3464 "test/x32-pad.cc",
3465 "test/pad-microkernel-tester.h",
3466 ] + MICROKERNEL_TEST_HDRS,
3467 deps = MICROKERNEL_TEST_DEPS,
3468)
3469
3470xnnpack_unit_test(
3471 name = "x32_unpool_test",
3472 srcs = [
3473 "test/x32-unpool.cc",
3474 "test/unpool-microkernel-tester.h",
3475 ] + MICROKERNEL_TEST_HDRS,
3476 deps = MICROKERNEL_TEST_DEPS,
3477)
3478
3479xnnpack_unit_test(
3480 name = "x32_zip_test",
3481 srcs = [
3482 "test/x32-zip.cc",
3483 "test/zip-microkernel-tester.h",
3484 ] + MICROKERNEL_TEST_HDRS,
3485 deps = MICROKERNEL_TEST_DEPS,
3486)
3487
3488xnnpack_unit_test(
3489 name = "x8_lut_test",
3490 srcs = [
3491 "test/x8-lut.cc",
3492 "test/lut-microkernel-tester.h",
3493 ] + MICROKERNEL_TEST_HDRS,
3494 deps = MICROKERNEL_TEST_DEPS,
3495)
3496
3497xnnpack_unit_test(
3498 name = "x8_zip_test",
3499 srcs = [
3500 "test/x8-zip.cc",
3501 "test/zip-microkernel-tester.h",
3502 ] + MICROKERNEL_TEST_HDRS,
3503 deps = MICROKERNEL_TEST_DEPS,
3504)
3505
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003506xnnpack_unit_test(
3507 name = "requantization_test",
3508 srcs = [
3509 "src/xnnpack/requantization-stubs.h",
3510 "test/requantization.cc",
3511 "test/requantization-tester.h",
3512 ] + MICROKERNEL_TEST_HDRS,
3513 deps = MICROKERNEL_TEST_DEPS,
3514)
3515
Marat Dukhan20c3b922020-03-10 03:45:06 -07003516########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517
3518xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07003519 name = "operator_size_test",
3520 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003521 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003522)
3523
Marat Dukhan20c3b922020-03-10 03:45:06 -07003524xnnpack_binary(
3525 name = "subgraph_size_test",
3526 srcs = ["test/subgraph-size.c"],
3527 deps = [":XNNPACK"],
3528)
3529
3530########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003531
3532xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003533 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003534 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003535 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536 "test/add-operator-tester.h",
3537 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003538 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539)
3540
3541xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003542 name = "add_nd_test",
3543 srcs = [
3544 "test/add-nd.cc",
3545 "test/binary-elementwise-operator-tester.h",
3546 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003547 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003548)
3549
3550xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003551 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003552 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003553 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 "test/argmax-pooling-operator-tester.h",
3555 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003556 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557)
3558
3559xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003560 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003561 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003562 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003563 "test/average-pooling-operator-tester.h",
3564 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003565 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566)
3567
3568xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003569 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003570 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003571 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003572 "test/channel-pad-operator-tester.h",
3573 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003574 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003575)
3576
3577xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003578 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003580 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003581 "test/channel-shuffle-operator-tester.h",
3582 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003583 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584)
3585
3586xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003587 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003589 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590 "test/clamp-operator-tester.h",
3591 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003592 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593)
3594
3595xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07003596 name = "constant_pad_nd_test",
3597 srcs = [
3598 "test/constant-pad-nd.cc",
3599 "test/constant-pad-operator-tester.h",
3600 ],
3601 deps = OPERATOR_TEST_DEPS,
3602)
3603
3604xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003605 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003606 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003607 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003608 "test/convolution-operator-tester.h",
3609 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003610 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611)
3612
3613xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003614 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003616 "test/convolution-nchw.cc",
3617 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003619 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003620)
3621
3622xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003623 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003625 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003626 "test/deconvolution-operator-tester.h",
3627 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003628 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629)
3630
3631xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08003632 name = "divide_nd_test",
3633 srcs = [
3634 "test/binary-elementwise-operator-tester.h",
3635 "test/divide-nd.cc",
3636 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003637 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08003638)
3639
3640xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003641 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003643 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 "test/fully-connected-operator-tester.h",
3645 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003646 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647)
3648
3649xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003650 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003651 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003652 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 "test/global-average-pooling-operator-tester.h",
3654 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003655 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656)
3657
3658xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003659 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003661 "test/global-average-pooling-ncw.cc",
3662 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003664 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003665)
3666
3667xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003668 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003669 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003670 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671 "test/hardswish-operator-tester.h",
3672 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003673 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003674)
3675
3676xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003677 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003678 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003679 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003680 "test/leaky-relu-operator-tester.h",
3681 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003682 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003683)
3684
3685xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003686 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003687 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003688 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003689 "test/max-pooling-operator-tester.h",
3690 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003691 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692)
3693
3694xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003695 name = "maximum_nd_test",
3696 srcs = [
3697 "test/binary-elementwise-operator-tester.h",
3698 "test/maximum-nd.cc",
3699 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003700 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003701)
3702
3703xnnpack_unit_test(
3704 name = "minimum_nd_test",
3705 srcs = [
3706 "test/binary-elementwise-operator-tester.h",
3707 "test/minimum-nd.cc",
3708 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003709 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08003710)
3711
3712xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003713 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003714 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003715 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003716 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003717 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003718 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08003719)
3720
3721xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003722 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003723 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003724 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003725 "test/prelu-operator-tester.h",
3726 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003727 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728)
3729
3730xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003731 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003732 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003733 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003734 "test/resize-bilinear-operator-tester.h",
3735 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07003736 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08003737)
3738
3739xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003740 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003741 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003742 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003743 "test/sigmoid-operator-tester.h",
3744 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003745 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746)
3747
3748xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003749 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003751 "test/softmax-nc.cc",
3752 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003754 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003755)
3756
3757xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003758 name = "subtract_nd_test",
3759 srcs = [
3760 "test/binary-elementwise-operator-tester.h",
3761 "test/subtract-nd.cc",
3762 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003763 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003764)
3765
3766xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003767 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003768 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003769 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770 "test/unpooling-operator-tester.h",
3771 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07003772 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773)
3774
Chao Mei6ddfc602020-05-13 22:29:36 -07003775############################### Misc unit tests ###############################
3776
3777xnnpack_unit_test(
3778 name = "memory_planner_test",
3779 srcs = [
3780 "test/memory-planner-test.cc",
3781 ],
3782 deps = [
3783 ":XNNPACK",
3784 ":memory_planner",
3785 ],
3786)
3787
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788############################# Build configurations #############################
3789
Marat Dukhanb8642352019-10-30 15:43:02 -07003790# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003792 name = "xnn_enable_assembly_explicit_true",
3793 define_values = {"xnn_enable_assembly": "true"},
3794)
3795
3796# Disables usage of assembly kernels.
3797config_setting(
3798 name = "xnn_enable_assembly_explicit_false",
3799 define_values = {"xnn_enable_assembly": "false"},
3800)
3801
Marat Dukhan05702cf2020-03-26 15:41:33 -07003802# Disables usage of HMP-aware optimizations.
3803config_setting(
3804 name = "xnn_enable_hmp_explicit_false",
3805 define_values = {"xnn_enable_hmp": "false"},
3806)
3807
Chao Mei6ddfc602020-05-13 22:29:36 -07003808# Enable usage of optimized memory allocation
3809config_setting(
3810 name = "xnn_enable_memopt_explicit_true",
3811 define_values = {"xnnpack_enable_memopt": "true"},
3812)
3813
3814# Disable usage of optimized memory allocation
3815config_setting(
3816 name = "xnn_enable_memopt_explicit_false",
3817 define_values = {"xnnpack_enable_memopt": "false"},
3818)
3819
Marat Dukhanb8642352019-10-30 15:43:02 -07003820# Builds with -c dbg
3821config_setting(
3822 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003823 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003824 "compilation_mode": "dbg",
3825 },
3826)
3827
3828# Builds with -c opt
3829config_setting(
3830 name = "optimized_build",
3831 values = {
3832 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003833 },
3834)
3835
3836config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003837 name = "linux_k8",
3838 values = {"cpu": "k8"},
3839)
3840
3841config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003842 name = "linux_arm",
3843 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003844)
3845
3846config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07003847 name = "linux_armhf",
3848 values = {"cpu": "armhf"},
3849)
3850
3851config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07003852 name = "linux_aarch64",
3853 values = {"cpu": "aarch64"},
3854)
3855
3856config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003857 name = "android",
3858 values = {"crosstool_top": "//external:android/crosstool"},
3859)
3860
3861config_setting(
3862 name = "android_armv7",
3863 values = {
3864 "crosstool_top": "//external:android/crosstool",
3865 "cpu": "armeabi-v7a",
3866 },
3867)
3868
3869config_setting(
3870 name = "android_arm64",
3871 values = {
3872 "crosstool_top": "//external:android/crosstool",
3873 "cpu": "arm64-v8a",
3874 },
3875)
3876
3877config_setting(
3878 name = "android_x86",
3879 values = {
3880 "crosstool_top": "//external:android/crosstool",
3881 "cpu": "x86",
3882 },
3883)
3884
3885config_setting(
3886 name = "android_x86_64",
3887 values = {
3888 "crosstool_top": "//external:android/crosstool",
3889 "cpu": "x86_64",
3890 },
3891)
3892
3893config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003894 name = "windows_x86_64",
3895 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003896)
3897
3898config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07003899 name = "windows_x86_64_clang",
3900 values = {
3901 "compiler": "clang-cl",
3902 "cpu": "x64_windows",
3903 },
3904)
3905
3906config_setting(
3907 name = "windows_x86_64_mingw",
3908 values = {
3909 "compiler": "mingw-gcc",
3910 "cpu": "x64_windows",
3911 },
3912)
3913
3914config_setting(
3915 name = "windows_x86_64_msys",
3916 values = {
3917 "compiler": "msys-gcc",
3918 "cpu": "x64_windows",
3919 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07003920)
3921
3922config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003923 name = "macos_x86_64",
3924 values = {
3925 "apple_platform_type": "macos",
3926 "cpu": "darwin",
3927 },
3928)
3929
3930config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003932 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003933)
3934
3935config_setting(
3936 name = "emscripten_wasm",
3937 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003938 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003939 "cpu": "wasm",
3940 },
3941)
3942
3943config_setting(
3944 name = "emscripten_wasmsimd",
3945 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003946 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003947 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003948 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003949 },
3950)
3951
3952config_setting(
3953 name = "emscripten_asmjs",
3954 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003955 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003956 "cpu": "asmjs",
3957 },
3958)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003959
3960config_setting(
3961 name = "ios_armv7",
3962 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003963 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003964 "cpu": "ios_armv7",
3965 },
3966)
3967
3968config_setting(
3969 name = "ios_arm64",
3970 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003971 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003972 "cpu": "ios_arm64",
3973 },
3974)
3975
3976config_setting(
3977 name = "ios_arm64e",
3978 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003979 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003980 "cpu": "ios_arm64e",
3981 },
3982)
3983
3984config_setting(
3985 name = "ios_x86",
3986 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003987 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003988 "cpu": "ios_i386",
3989 },
3990)
3991
3992config_setting(
3993 name = "ios_x86_64",
3994 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08003995 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08003996 "cpu": "ios_x86_64",
3997 },
3998)
3999
4000config_setting(
4001 name = "watchos_armv7k",
4002 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004003 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004004 "cpu": "watchos_armv7k",
4005 },
4006)
4007
4008config_setting(
4009 name = "watchos_arm64_32",
4010 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004011 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004012 "cpu": "watchos_arm64_32",
4013 },
4014)
4015
4016config_setting(
4017 name = "watchos_x86",
4018 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004019 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004020 "cpu": "watchos_i386",
4021 },
4022)
4023
4024config_setting(
4025 name = "watchos_x86_64",
4026 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004027 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004028 "cpu": "watchos_x86_64",
4029 },
4030)
4031
4032config_setting(
4033 name = "tvos_arm64",
4034 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004035 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004036 "cpu": "tvos_arm64",
4037 },
4038)
4039
4040config_setting(
4041 name = "tvos_x86_64",
4042 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004043 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004044 "cpu": "tvos_x86_64",
4045 },
4046)