blob: 095b25967d10547e2706285d1d2aaf21b12a52d9 [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 Dukhana84e40b2019-12-11 15:38:03 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_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_copts", "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 Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_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
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-avgpool/mp9p8q-scalar.c",
86 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080087 "src/f32-bilinear/gen/scalar-c1.c",
88 "src/f32-bilinear/gen/scalar-c2.c",
89 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070090 "src/f32-clamp/scalar.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",
Erich Elsen563df5f2019-10-23 08:02:21 -070093 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070094 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070095 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080096 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080097 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080098 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
99 "src/f32-dwconv/gen/up1x25-scalar.c",
100 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
101 "src/f32-dwconv/gen/up1x4-scalar.c",
102 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
103 "src/f32-dwconv/gen/up1x9-scalar.c",
104 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
105 "src/f32-dwconv/gen/up2x25-scalar.c",
106 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
107 "src/f32-dwconv/gen/up2x4-scalar.c",
108 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
109 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700110 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700111 "src/f32-gavgpool/mp7p7q-scalar.c",
112 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800113 "src/f32-gemm/gen-inc/1x4-scalar.c",
114 "src/f32-gemm/gen-inc/2x4-scalar.c",
115 "src/f32-gemm/gen-inc/4x4-scalar.c",
116 "src/f32-gemm/gen/1x4-scalar.c",
117 "src/f32-gemm/gen/2x4-scalar.c",
118 "src/f32-gemm/gen/4x2-scalar.c",
119 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800120 "src/f32-hswish/gen/scalar-x1.c",
121 "src/f32-hswish/gen/scalar-x2.c",
122 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800123 "src/f32-igemm/gen/1x4-scalar.c",
124 "src/f32-igemm/gen/2x4-scalar.c",
125 "src/f32-igemm/gen/4x2-scalar.c",
126 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800127 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700128 "src/f32-pavgpool/mp9p8q-scalar.c",
129 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800130 "src/f32-ppmm/gen/2x4-scalar.c",
131 "src/f32-ppmm/gen/3x3-scalar.c",
132 "src/f32-ppmm/gen/4x2-scalar.c",
133 "src/f32-ppmm/gen/4x4-scalar.c",
134 "src/f32-prelu/gen/scalar-2x1.c",
135 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800136 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
137 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
138 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
139 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
140 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
141 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
142 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
143 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
144 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
145 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
146 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
147 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700148 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800149 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
150 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
151 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
152 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
153 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
154 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
155 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
156 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
157 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800158 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
159 "src/f32-spmm/gen/1x1-scalar.c",
160 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
161 "src/f32-spmm/gen/2x1-scalar.c",
162 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
163 "src/f32-spmm/gen/4x1-scalar.c",
164 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
165 "src/f32-spmm/gen/8x1-scalar.c",
166 "src/f32-spmm/gen/8x2-scalar.c",
167 "src/f32-spmm/gen/8x4-scalar.c",
168 "src/f32-vbinary/gen/vadd-scalar-x1.c",
169 "src/f32-vbinary/gen/vadd-scalar-x2.c",
170 "src/f32-vbinary/gen/vadd-scalar-x4.c",
171 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
172 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
173 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800174 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
175 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
176 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
177 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
178 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
179 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800180 "src/f32-vbinary/gen/vmax-scalar-x1.c",
181 "src/f32-vbinary/gen/vmax-scalar-x2.c",
182 "src/f32-vbinary/gen/vmax-scalar-x4.c",
183 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
184 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
185 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
186 "src/f32-vbinary/gen/vmin-scalar-x1.c",
187 "src/f32-vbinary/gen/vmin-scalar-x2.c",
188 "src/f32-vbinary/gen/vmin-scalar-x4.c",
189 "src/f32-vbinary/gen/vminc-scalar-x1.c",
190 "src/f32-vbinary/gen/vminc-scalar-x2.c",
191 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800192 "src/f32-vbinary/gen/vmul-scalar-x1.c",
193 "src/f32-vbinary/gen/vmul-scalar-x2.c",
194 "src/f32-vbinary/gen/vmul-scalar-x4.c",
195 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
196 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
197 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800198 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
199 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
200 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
202 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
203 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
204 "src/f32-vbinary/gen/vsub-scalar-x1.c",
205 "src/f32-vbinary/gen/vsub-scalar-x2.c",
206 "src/f32-vbinary/gen/vsub-scalar-x4.c",
207 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
208 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
209 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
210 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
211 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
212 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800213 "src/math/expminus-scalar-lut2048-p1.c",
214 "src/math/expminus-scalar-lut64-p2.c",
215 "src/math/expminus-scalar-p5.c",
216 "src/math/sigmoid-scalar-lut2048-p1-div.c",
217 "src/math/sigmoid-scalar-lut64-p2-div.c",
218 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700219 "src/q8-avgpool/mp9p8q-scalar.c",
220 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700221 "src/q8-dwconv/up1x9-scalar.c",
222 "src/q8-gavgpool/mp7p7q-scalar.c",
223 "src/q8-gavgpool/up7-scalar.c",
224 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800225 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700226 "src/q8-vadd/scalar.c",
227 "src/u8-clamp/scalar.c",
228 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800229 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/u8-rmax/scalar.c",
231 "src/x32-packx/x2-scalar.c",
232 "src/x32-packx/x3-scalar.c",
233 "src/x32-packx/x4-scalar.c",
234 "src/x32-pad/x2-scalar.c",
235 "src/x32-unpool/scalar.c",
236 "src/x32-zip/x2-scalar.c",
237 "src/x32-zip/x3-scalar.c",
238 "src/x32-zip/x4-scalar.c",
239 "src/x32-zip/xm-scalar.c",
240 "src/x8-lut/scalar.c",
241 "src/x8-zip/x2-scalar.c",
242 "src/x8-zip/x3-scalar.c",
243 "src/x8-zip/x4-scalar.c",
244 "src/x8-zip/xm-scalar.c",
245]
246
Marat Dukhan436ebe62019-12-04 15:10:12 -0800247WASM_UKERNELS = [
248 "src/f32-avgpool/mp9p8q-wasm.c",
249 "src/f32-avgpool/up9-wasm.c",
250 "src/f32-clamp/wasm.c",
251 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
252 "src/f32-dwconv/gen/up1x25-wasm.c",
253 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
254 "src/f32-dwconv/gen/up1x4-wasm.c",
255 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
256 "src/f32-dwconv/gen/up1x9-wasm.c",
257 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
258 "src/f32-dwconv/gen/up2x25-wasm.c",
259 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
260 "src/f32-dwconv/gen/up2x4-wasm.c",
261 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
262 "src/f32-dwconv/gen/up2x9-wasm.c",
263 "src/f32-gavgpool/mp7p7q-wasm.c",
264 "src/f32-gavgpool/up7-wasm.c",
265 "src/f32-gemm/gen-inc/1x4-wasm.c",
266 "src/f32-gemm/gen-inc/2x4-wasm.c",
267 "src/f32-gemm/gen-inc/4x4-wasm.c",
268 "src/f32-gemm/gen/1x4-wasm.c",
269 "src/f32-gemm/gen/2x4-wasm.c",
270 "src/f32-gemm/gen/4x2-wasm.c",
271 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800272 "src/f32-hswish/gen/wasm-x1.c",
273 "src/f32-hswish/gen/wasm-x2.c",
274 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800275 "src/f32-igemm/gen/1x4-wasm.c",
276 "src/f32-igemm/gen/2x4-wasm.c",
277 "src/f32-igemm/gen/4x2-wasm.c",
278 "src/f32-igemm/gen/4x4-wasm.c",
279 "src/f32-maxpool/9p8x-wasm-c1.c",
280 "src/f32-pavgpool/mp9p8q-wasm.c",
281 "src/f32-pavgpool/up9-wasm.c",
282 "src/f32-prelu/gen/wasm-2x1.c",
283 "src/f32-prelu/gen/wasm-2x4.c",
284 "src/f32-vbinary/gen/vadd-wasm-x1.c",
285 "src/f32-vbinary/gen/vadd-wasm-x2.c",
286 "src/f32-vbinary/gen/vadd-wasm-x4.c",
287 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
288 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
289 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800290 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
291 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
292 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
293 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
294 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
295 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800296 "src/f32-vbinary/gen/vmax-wasm-x1.c",
297 "src/f32-vbinary/gen/vmax-wasm-x2.c",
298 "src/f32-vbinary/gen/vmax-wasm-x4.c",
299 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
300 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
301 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
302 "src/f32-vbinary/gen/vmin-wasm-x1.c",
303 "src/f32-vbinary/gen/vmin-wasm-x2.c",
304 "src/f32-vbinary/gen/vmin-wasm-x4.c",
305 "src/f32-vbinary/gen/vminc-wasm-x1.c",
306 "src/f32-vbinary/gen/vminc-wasm-x2.c",
307 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800308 "src/f32-vbinary/gen/vmul-wasm-x1.c",
309 "src/f32-vbinary/gen/vmul-wasm-x2.c",
310 "src/f32-vbinary/gen/vmul-wasm-x4.c",
311 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
312 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
313 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800314 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
315 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
316 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800317 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
318 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
319 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
320 "src/f32-vbinary/gen/vsub-wasm-x1.c",
321 "src/f32-vbinary/gen/vsub-wasm-x2.c",
322 "src/f32-vbinary/gen/vsub-wasm-x4.c",
323 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
324 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
325 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
326 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
327 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
328 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
329]
330
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800331PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800332 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800333 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800334 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700335 "src/f32-avgpool/mp9p8q-psimd.c",
336 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800337 "src/f32-bilinear/gen/psimd-c4.c",
338 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700339 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800340 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
341 "src/f32-dwconv/gen/up4x25-psimd.c",
342 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
343 "src/f32-dwconv/gen/up4x4-psimd.c",
344 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
345 "src/f32-dwconv/gen/up4x9-psimd.c",
346 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
347 "src/f32-dwconv/gen/up8x25-psimd.c",
348 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
349 "src/f32-dwconv/gen/up8x4-psimd.c",
350 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
351 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700352 "src/f32-gavgpool/mp7p7q-psimd.c",
353 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800354 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
355 "src/f32-gemm/gen/1x8-psimd-splat.c",
356 "src/f32-gemm/gen/1x8s4-psimd.c",
357 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
358 "src/f32-gemm/gen/4x8-psimd-splat.c",
359 "src/f32-gemm/gen/4x8s4-psimd.c",
360 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
361 "src/f32-gemm/gen/6x8-psimd-splat.c",
362 "src/f32-gemm/gen/6x8s4-psimd.c",
363 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
364 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
365 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
366 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
367 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
368 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
369 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
370 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
371 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800372 "src/f32-hswish/gen/psimd-x4.c",
373 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800374 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
375 "src/f32-igemm/gen/1x8-psimd-splat.c",
376 "src/f32-igemm/gen/1x8s4-psimd.c",
377 "src/f32-igemm/gen/4x2c4-psimd.c",
378 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
379 "src/f32-igemm/gen/4x8-psimd-splat.c",
380 "src/f32-igemm/gen/4x8s4-psimd.c",
381 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
382 "src/f32-igemm/gen/6x8-psimd-splat.c",
383 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800384 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700385 "src/f32-pavgpool/mp9p8q-psimd.c",
386 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800387 "src/f32-ppmm/gen/4x8-psimd.c",
388 "src/f32-prelu/gen/psimd-2x4.c",
389 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800390 "src/f32-rmax/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800391 "src/f32-vbinary/gen/vadd-psimd-x4.c",
392 "src/f32-vbinary/gen/vadd-psimd-x8.c",
393 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
394 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800395 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
396 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
397 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
398 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800399 "src/f32-vbinary/gen/vmax-psimd-x4.c",
400 "src/f32-vbinary/gen/vmax-psimd-x8.c",
401 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
402 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
403 "src/f32-vbinary/gen/vmin-psimd-x4.c",
404 "src/f32-vbinary/gen/vmin-psimd-x8.c",
405 "src/f32-vbinary/gen/vminc-psimd-x4.c",
406 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800407 "src/f32-vbinary/gen/vmul-psimd-x4.c",
408 "src/f32-vbinary/gen/vmul-psimd-x8.c",
409 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
410 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800411 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
412 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800413 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
414 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
415 "src/f32-vbinary/gen/vsub-psimd-x4.c",
416 "src/f32-vbinary/gen/vsub-psimd-x8.c",
417 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
418 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
419 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
420 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421 "src/x32-packx/x4-psimd.c",
422 "src/x32-pad/x2-psimd.c",
423 "src/x32-unpool/psimd.c",
424 "src/x32-zip/x2-psimd.c",
425 "src/x32-zip/x3-psimd.c",
426 "src/x32-zip/x4-psimd.c",
427 "src/x32-zip/xm-psimd.c",
428]
429
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800430PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800431 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
432 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
433 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
434 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
435 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
436 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
437 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
438 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
439 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
440 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
441 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
442 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800443 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
444 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
445 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
446 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
447 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
448 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
449 "src/math/sigmoid-psimd-p5-div.c",
450]
451
Marat Dukhan08c4a432019-10-03 09:29:21 -0700452# ISA-specific micro-kernels
453NEON_UKERNELS = [
454 "src/f32-avgpool/mp9p8q-neon.c",
455 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800456 "src/f32-bilinear/gen/neon-c4.c",
457 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800459 "src/f32-dwconv/gen/up4x9-neon.c",
460 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
461 "src/f32-dwconv/gen/up8x9-neon.c",
462 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/f32-gavgpool-spchw/neon-x4.c",
464 "src/f32-gavgpool/mp7p7q-neon.c",
465 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800466 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
467 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
468 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
469 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
470 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
471 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800472 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800473 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
474 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
475 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
476 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800477 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800478 "src/f32-gemm/gen/1x8s4-neon.c",
479 "src/f32-gemm/gen/4x8s4-neon.c",
480 "src/f32-gemm/gen/6x8s4-neon.c",
481 "src/f32-gemm/gen/8x8s4-neon.c",
482 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
483 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
484 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
485 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
486 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800487 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800488 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
489 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
490 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
491 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800492 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800493 "src/f32-gemm/gen-inc/1x8s4-neon.c",
494 "src/f32-gemm/gen-inc/4x8s4-neon.c",
495 "src/f32-gemm/gen-inc/6x8s4-neon.c",
496 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800497 "src/f32-hswish/gen/neon-x4.c",
498 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800499 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
500 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
501 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
502 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
503 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
504 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800505 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800506 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
507 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
508 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
509 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800510 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800511 "src/f32-igemm/gen/1x8s4-neon.c",
512 "src/f32-igemm/gen/4x8s4-neon.c",
513 "src/f32-igemm/gen/6x8s4-neon.c",
514 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/f32-pavgpool/mp9p8q-neon.c",
516 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800517 "src/f32-ppmm/gen/4x8-neon.c",
518 "src/f32-ppmm/gen/8x8-neon.c",
519 "src/f32-prelu/gen/neon-2x4.c",
520 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800521 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
522 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
523 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
524 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
525 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
526 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
527 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
528 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
529 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
530 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
531 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
532 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
533 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
534 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
535 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
536 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
537 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
538 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
539 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
540 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
541 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
542 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
543 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
544 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800546 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800547 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
548 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
549 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
550 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
551 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
552 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
553 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
554 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
555 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
556 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
557 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
558 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
559 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
560 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
561 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
562 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
563 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
564 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800565 "src/f32-vbinary/gen/vadd-neon-x4.c",
566 "src/f32-vbinary/gen/vadd-neon-x8.c",
567 "src/f32-vbinary/gen/vaddc-neon-x4.c",
568 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800569 "src/f32-vbinary/gen/vmax-neon-x4.c",
570 "src/f32-vbinary/gen/vmax-neon-x8.c",
571 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
572 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
573 "src/f32-vbinary/gen/vmin-neon-x4.c",
574 "src/f32-vbinary/gen/vmin-neon-x8.c",
575 "src/f32-vbinary/gen/vminc-neon-x4.c",
576 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800577 "src/f32-vbinary/gen/vmul-neon-x4.c",
578 "src/f32-vbinary/gen/vmul-neon-x8.c",
579 "src/f32-vbinary/gen/vmulc-neon-x4.c",
580 "src/f32-vbinary/gen/vmulc-neon-x8.c",
581 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
582 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
583 "src/f32-vbinary/gen/vsub-neon-x4.c",
584 "src/f32-vbinary/gen/vsub-neon-x8.c",
585 "src/f32-vbinary/gen/vsubc-neon-x4.c",
586 "src/f32-vbinary/gen/vsubc-neon-x8.c",
587 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
588 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700589 "src/q8-avgpool/mp9p8q-neon.c",
590 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/q8-dwconv/up8x9-neon.c",
592 "src/q8-gavgpool/mp7p7q-neon.c",
593 "src/q8-gavgpool/up7-neon.c",
594 "src/q8-gemm/4x8-neon.c",
595 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800596 "src/q8-igemm/4x8-neon.c",
597 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700598 "src/q8-vadd/neon.c",
599 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800600 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700601 "src/u8-rmax/neon.c",
602 "src/x32-packx/x4-neon-st4.c",
603 "src/x32-pad/x2-neon.c",
604 "src/x32-zip/x2-neon.c",
605 "src/x32-zip/x3-neon.c",
606 "src/x32-zip/x4-neon.c",
607 "src/x32-zip/xm-neon.c",
608 "src/x8-zip/x2-neon.c",
609 "src/x8-zip/x3-neon.c",
610 "src/x8-zip/x4-neon.c",
611 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800612 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800613 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
614 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
615 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
616 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
617 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
618 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700619]
620
621NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800622 "src/f32-bilinear/gen/neonfma-c4.c",
623 "src/f32-bilinear/gen/neonfma-c8.c",
624 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
625 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
626 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
627 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800628 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800629 "src/f32-igemm/gen/1x8s4-neonfma.c",
630 "src/f32-igemm/gen/4x8s4-neonfma.c",
631 "src/f32-igemm/gen/6x8s4-neonfma.c",
632 "src/f32-igemm/gen/8x8s4-neonfma.c",
633 "src/f32-dwconv/gen/up4x9-neonfma.c",
634 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
635 "src/f32-dwconv/gen/up8x9-neonfma.c",
636 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
637 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
638 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
639 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
640 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800641 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800642 "src/f32-gemm/gen/1x8s4-neonfma.c",
643 "src/f32-gemm/gen/4x8s4-neonfma.c",
644 "src/f32-gemm/gen/6x8s4-neonfma.c",
645 "src/f32-gemm/gen/8x8s4-neonfma.c",
646 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
647 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
648 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
649 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800650 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800651 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
652 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
653 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
654 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800655 "src/f32-hswish/gen/neonfma-x4.c",
656 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800657 "src/f32-ppmm/gen/4x8-neonfma.c",
658 "src/f32-ppmm/gen/8x8-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800659 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
660 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
661 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
662 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
663 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
664 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
665 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
666 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
667 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
668 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
669 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
670 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
671 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
672 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
673 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
674 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
675 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
676 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
677 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
678 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
679 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
680 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
681 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
682 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800683 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
684 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
685 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
686 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
687 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
688 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
689 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
690 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
691 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
692 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
693 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
694 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
695 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
696 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
697 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
698 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
699 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
700 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
701 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
702 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
703 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
704 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
705 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
706 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
707 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
708 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
709 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
710 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
711 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
712 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
713 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
714 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
715 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
716 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
718 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
719 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
720 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
721 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
722 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
723 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
724 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
725 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
726 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
727 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
728 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
729 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
730 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
731 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
732 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
733 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
734 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
735 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
736 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800737 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
738 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800739 "src/math/exp-neonfma-lut64-p2.c",
740 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800741 "src/math/expminus-neonfma-lut2048-p1.c",
742 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800743 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800744 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
745 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
746 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
747 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
748 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
749 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
750 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
751 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
752 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
753 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
754 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
755 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
756 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
757 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
758 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
759 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
760 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
761 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700762]
763
764AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800765 "src/f32-vbinary/gen/vdiv-neon-x4.c",
766 "src/f32-vbinary/gen/vdiv-neon-x8.c",
767 "src/f32-vbinary/gen/vdivc-neon-x4.c",
768 "src/f32-vbinary/gen/vdivc-neon-x8.c",
769 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
770 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800771 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
772 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
773 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
774 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
775 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
776 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800777 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800778 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
779 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
780 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
781 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
782 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800783 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800784 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
785 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
786 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
787 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
788 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
789 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800790 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
792 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
793 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
794 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
795 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
796 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
797 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800798 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800816 "src/f32-spmm/gen/12x1-neonfma.c",
817 "src/f32-spmm/gen/12x2-neonfma.c",
818 "src/f32-spmm/gen/12x4-neonfma.c",
819 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
820 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
821 "src/f32-spmm/gen/16x1-neonfma.c",
822 "src/f32-spmm/gen/16x2-neonfma.c",
823 "src/f32-spmm/gen/16x4-neonfma.c",
824 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
825 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
826 "src/f32-spmm/gen/4x1-neonfma.c",
827 "src/f32-spmm/gen/4x2-neonfma.c",
828 "src/f32-spmm/gen/4x4-neonfma.c",
829 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
830 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
831 "src/f32-spmm/gen/8x1-neonfma.c",
832 "src/f32-spmm/gen/8x2-neonfma.c",
833 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800834 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
835 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
836 "src/math/sigmoid-neonfma-rr1-p5-div.c",
837 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
838 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
839 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700840]
841
842AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800843 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
844 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
845 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhanbdb56f52020-02-05 21:42:49 -0800846 "src/f16-spmm/gen/8x1-neonfp16arith.c",
847 "src/f16-spmm/gen/8x1-neonfp16arith-unroll2.c",
848 "src/f16-spmm/gen/16x1-neonfp16arith.c",
849 "src/f16-spmm/gen/16x1-neonfp16arith-unroll2.c",
850 "src/f16-spmm/gen/24x1-neonfp16arith.c",
851 "src/f16-spmm/gen/24x1-neonfp16arith-unroll2.c",
852 "src/f16-spmm/gen/32x1-neonfp16arith.c",
853 "src/f16-spmm/gen/32x1-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700854]
855
856SSE_UKERNELS = [
857 "src/f32-avgpool/mp9p8q-sse.c",
858 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800859 "src/f32-bilinear/gen/sse-c4.c",
860 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700861 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800862 "src/f32-dwconv-spchw/3x3p1-sse.c",
863 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800864 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
865 "src/f32-dwconv/gen/up4x25-sse.c",
866 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
867 "src/f32-dwconv/gen/up4x4-sse.c",
868 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
869 "src/f32-dwconv/gen/up4x9-sse.c",
870 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
871 "src/f32-dwconv/gen/up8x25-sse.c",
872 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
873 "src/f32-dwconv/gen/up8x4-sse.c",
874 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
875 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876 "src/f32-gavgpool-spchw/sse-x4.c",
877 "src/f32-gavgpool/mp7p7q-sse.c",
878 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800879 "src/f32-gemm/gen/1x8-sse-dup.c",
880 "src/f32-gemm/gen/1x8-sse-load1.c",
881 "src/f32-gemm/gen/1x8s4-sse.c",
882 "src/f32-gemm/gen/4x8-sse-dup.c",
883 "src/f32-gemm/gen/4x8-sse-load1.c",
884 "src/f32-gemm/gen/4x8s4-sse.c",
885 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
886 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
887 "src/f32-gemm/gen-inc/1x8s4-sse.c",
888 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
889 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
890 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800891 "src/f32-hswish/gen/sse-x4.c",
892 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800893 "src/f32-igemm/gen/1x8-sse-dup.c",
894 "src/f32-igemm/gen/1x8-sse-load1.c",
895 "src/f32-igemm/gen/1x8s4-sse.c",
896 "src/f32-igemm/gen/4x2c4-sse.c",
897 "src/f32-igemm/gen/4x8-sse-dup.c",
898 "src/f32-igemm/gen/4x8-sse-load1.c",
899 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800900 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700901 "src/f32-pavgpool/mp9p8q-sse.c",
902 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800903 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700904 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800905 "src/f32-spmm/gen/4x1-sse.c",
906 "src/f32-spmm/gen/8x1-sse.c",
907 "src/f32-vbinary/gen/vadd-sse-x4.c",
908 "src/f32-vbinary/gen/vadd-sse-x8.c",
909 "src/f32-vbinary/gen/vaddc-sse-x4.c",
910 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800911 "src/f32-vbinary/gen/vdiv-sse-x4.c",
912 "src/f32-vbinary/gen/vdiv-sse-x8.c",
913 "src/f32-vbinary/gen/vdivc-sse-x4.c",
914 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800915 "src/f32-vbinary/gen/vmax-sse-x4.c",
916 "src/f32-vbinary/gen/vmax-sse-x8.c",
917 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
918 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
919 "src/f32-vbinary/gen/vmin-sse-x4.c",
920 "src/f32-vbinary/gen/vmin-sse-x8.c",
921 "src/f32-vbinary/gen/vminc-sse-x4.c",
922 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800923 "src/f32-vbinary/gen/vmul-sse-x4.c",
924 "src/f32-vbinary/gen/vmul-sse-x8.c",
925 "src/f32-vbinary/gen/vmulc-sse-x4.c",
926 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800927 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
928 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800929 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
930 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
931 "src/f32-vbinary/gen/vsub-sse-x4.c",
932 "src/f32-vbinary/gen/vsub-sse-x8.c",
933 "src/f32-vbinary/gen/vsubc-sse-x4.c",
934 "src/f32-vbinary/gen/vsubc-sse-x8.c",
935 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
936 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937 "src/x32-packx/x4-sse.c",
938]
939
940SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800941 "src/f32-argmaxpool/9p8x-sse2-c4.c",
942 "src/f32-argmaxpool/4x-sse2-c4.c",
943 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800944 "src/f32-prelu/gen/sse2-2x4.c",
945 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800946 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
947 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
948 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
949 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
950 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
951 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
952 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
953 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
954 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
955 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
956 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
957 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800958 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800959 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800960 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800961 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800962 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
963 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700964 "src/q8-avgpool/mp9p8q-sse2.c",
965 "src/q8-avgpool/up9-sse2.c",
966 "src/q8-igemm/4x4c2-sse2.c",
967 "src/q8-dwconv/up8x9-sse2.c",
968 "src/q8-gavgpool/mp7p7q-sse2.c",
969 "src/q8-gavgpool/up7-sse2.c",
970 "src/q8-gemm/2x4c8-sse2.c",
971 "src/q8-gemm/4x4c2-sse2.c",
972 "src/q8-vadd/sse2.c",
973 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800974 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700975 "src/u8-rmax/sse2.c",
976 "src/x32-pad/x2-sse2.c",
977 "src/x32-zip/x2-sse2.c",
978 "src/x32-zip/x3-sse2.c",
979 "src/x32-zip/x4-sse2.c",
980 "src/x32-zip/xm-sse2.c",
981 "src/x8-zip/x2-sse2.c",
982 "src/x8-zip/x3-sse2.c",
983 "src/x8-zip/x4-sse2.c",
984 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800985 "src/math/exp-sse2-p5.c",
986 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800987 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988]
989
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800990SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800991 "src/f32-prelu/gen/sse41-2x4.c",
992 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800993 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800994 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800995 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800996 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800997 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
998 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800999]
1000
Marat Dukhan08c4a432019-10-03 09:29:21 -07001001AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001002 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001003 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
1004 "src/f32-dwconv/gen/up16x4-avx.c",
1005 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
1006 "src/f32-dwconv/gen/up8x4-avx.c",
1007 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
1008 "src/f32-dwconv/gen/up16x9-avx.c",
1009 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
1010 "src/f32-dwconv/gen/up8x9-avx.c",
1011 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
1012 "src/f32-dwconv/gen/up16x25-avx.c",
1013 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
1014 "src/f32-dwconv/gen/up8x25-avx.c",
1015 "src/f32-gemm/gen/1x8-avx-broadcast.c",
1016 "src/f32-gemm/gen/4x8-avx-broadcast.c",
1017 "src/f32-gemm/gen/5x8-avx-broadcast.c",
1018 "src/f32-gemm/gen/6x8-avx-broadcast.c",
1019 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001020 "src/f32-gemm/gen/1x16-avx-broadcast.c",
1021 "src/f32-gemm/gen/3x16-avx-broadcast.c",
1022 "src/f32-gemm/gen/4x16-avx-broadcast.c",
1023 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001024 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
1025 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
1026 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
1027 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
1028 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001029 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
1030 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
1031 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
1032 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001033 "src/f32-hswish/gen/avx-x8.c",
1034 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001035 "src/f32-igemm/gen/1x8-avx-broadcast.c",
1036 "src/f32-igemm/gen/4x8-avx-broadcast.c",
1037 "src/f32-igemm/gen/5x8-avx-broadcast.c",
1038 "src/f32-igemm/gen/6x8-avx-broadcast.c",
1039 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001040 "src/f32-igemm/gen/1x16-avx-broadcast.c",
1041 "src/f32-igemm/gen/3x16-avx-broadcast.c",
1042 "src/f32-igemm/gen/4x16-avx-broadcast.c",
1043 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001044 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001045 "src/f32-vbinary/gen/vadd-avx-x8.c",
1046 "src/f32-vbinary/gen/vadd-avx-x16.c",
1047 "src/f32-vbinary/gen/vaddc-avx-x8.c",
1048 "src/f32-vbinary/gen/vaddc-avx-x16.c",
1049 "src/f32-vbinary/gen/vdiv-avx-x8.c",
1050 "src/f32-vbinary/gen/vdiv-avx-x16.c",
1051 "src/f32-vbinary/gen/vdivc-avx-x8.c",
1052 "src/f32-vbinary/gen/vdivc-avx-x16.c",
1053 "src/f32-vbinary/gen/vmax-avx-x8.c",
1054 "src/f32-vbinary/gen/vmax-avx-x16.c",
1055 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1056 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1057 "src/f32-vbinary/gen/vmin-avx-x8.c",
1058 "src/f32-vbinary/gen/vmin-avx-x16.c",
1059 "src/f32-vbinary/gen/vminc-avx-x8.c",
1060 "src/f32-vbinary/gen/vminc-avx-x16.c",
1061 "src/f32-vbinary/gen/vmul-avx-x8.c",
1062 "src/f32-vbinary/gen/vmul-avx-x16.c",
1063 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1064 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1065 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1066 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1067 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1068 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1069 "src/f32-vbinary/gen/vsub-avx-x8.c",
1070 "src/f32-vbinary/gen/vsub-avx-x16.c",
1071 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1072 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001073 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001074]
1075
Marat Dukhanfda12b82019-11-21 12:27:59 -08001076FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001077 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
1078 "src/f32-dwconv/gen/up16x4-fma3.c",
1079 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
1080 "src/f32-dwconv/gen/up8x4-fma3.c",
1081 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
1082 "src/f32-dwconv/gen/up16x9-fma3.c",
1083 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
1084 "src/f32-dwconv/gen/up8x9-fma3.c",
1085 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
1086 "src/f32-dwconv/gen/up16x25-fma3.c",
1087 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
1088 "src/f32-dwconv/gen/up8x25-fma3.c",
1089 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
1090 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
1091 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
1092 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
1093 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
1094 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001095 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
1096 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
1097 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
1098 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001099 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
1100 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
1101 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
1102 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001103 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
1104 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
1105 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
1106 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
1107 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
1108 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001109 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
1110 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
1111 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
1112 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001113 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
1114 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
1115 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
1116 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001117 "src/f32-hswish/gen/fma3-x8.c",
1118 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001119 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
1120 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
1121 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
1122 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
1123 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
1124 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001125 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
1126 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
1127 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
1128 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001129 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
1130 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
1131 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
1132 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001133]
1134
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001135AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001136 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1137 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1138 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1139 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1140 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1141 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1142 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1143 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1144 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1145 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1146 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1147 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1148 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1149 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1150 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1151 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1152 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1153 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1154 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1155 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1156 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1157 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1158 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1159 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1160 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1161 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1162 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1163 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1164 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1165 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1166 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1167 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1168 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1169 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1170 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1171 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001172 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1173 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1174 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1175 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1176 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1177 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1178 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1179 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1180 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1181 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1182 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1185 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1186 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1187 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1188 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1189 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1190 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1191 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1192 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1193 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1194 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1195 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1196 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1197 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1198 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1199 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1200 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1201 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001202 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1203 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1204 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1205 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1206 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1207 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1208 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1209 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1210 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1211 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1212 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1213 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1214 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1215 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1216 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1217 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1218 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1219 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1220 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1221 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1222 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1223 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1224 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1225 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001226 "src/math/exp-avx2-p5.c",
1227 "src/math/exp-avx2-perm-p3.c",
1228 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001229 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001230 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001231 "src/math/sigmoid-avx2-rr2-p5-div.c",
1232 "src/math/sigmoid-avx2-rr1-p5-div.c",
1233 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1234 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1235 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1236 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001237]
1238
Marat Dukhan08c4a432019-10-03 09:29:21 -07001239AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001240 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -08001241 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
1242 "src/f32-dwconv/gen/up32x4-avx512f.c",
1243 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
1244 "src/f32-dwconv/gen/up16x4-avx512f.c",
1245 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
1246 "src/f32-dwconv/gen/up32x9-avx512f.c",
1247 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
1248 "src/f32-dwconv/gen/up16x9-avx512f.c",
1249 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
1250 "src/f32-dwconv/gen/up32x25-avx512f.c",
1251 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
1252 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001253 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
1254 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
1255 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
1256 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
1257 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
1258 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
1259 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
1260 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
1261 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
1262 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
1263 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
1264 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001265 "src/f32-hswish/gen/avx512f-x16.c",
1266 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001267 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
1268 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
1269 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
1270 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
1271 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
1272 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001273 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1274 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1275 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1276 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1277 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1278 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1279 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1280 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1281 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1282 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1283 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1284 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1285 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1286 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1287 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1288 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1289 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1290 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1291 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1292 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1293 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1294 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1295 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1296 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1297 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1298 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1299 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1300 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1301 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1302 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1303 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1304 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1305 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1306 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1307 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1308 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001309 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001310 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1311 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1312 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1313 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1314 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1315 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1316 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1317 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1318 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1319 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1320 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1321 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1322 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1323 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1324 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1325 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1326 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1327 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1328 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1329 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1330 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1331 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1332 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1333 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1334 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1335 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1336 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1337 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001338 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001339 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1340 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1341 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1342 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1343 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1344 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1345 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1346 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1347 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1348 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1349 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1350 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1351 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1352 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1353 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1354 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1355 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1356 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1357 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1358 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1359 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1360 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1361 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1362 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001363 "src/math/exp-avx512f-p5-scalef.c",
1364 "src/math/exp-avx512f-p5.c",
1365 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001366 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001367 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368]
1369
1370AARCH32_ASM_UKERNELS = [
1371 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001372 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001373 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1374 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001375 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001376]
1377
1378AARCH64_ASM_UKERNELS = [
1379 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1380 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001381 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1382 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1383 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1384 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1385 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1386 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1387 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1388 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1389 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1390 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001391 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001392 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1393 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001394 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001395 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001396 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1397 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1398 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1399 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1400 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1401 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1402 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1403 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1404 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1405 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1406 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1407 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1408 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001409 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001410 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1411 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001412 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001413 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001414 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1415 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1416 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001417 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001418 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001419 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1420 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001421 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001422 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001423 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1424 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1425 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1426 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001427 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001428 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001429 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1430 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001431]
1432
1433INTERNAL_MICROKERNEL_HDRS = [
1434 "src/xnnpack/argmaxpool.h",
1435 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001436 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001437 "src/xnnpack/clamp.h",
1438 "src/xnnpack/common.h",
1439 "src/xnnpack/conv.h",
1440 "src/xnnpack/dwconv.h",
1441 "src/xnnpack/gavgpool.h",
1442 "src/xnnpack/gemm.h",
1443 "src/xnnpack/hswish.h",
1444 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001445 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001446 "src/xnnpack/lut.h",
1447 "src/xnnpack/math.h",
1448 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001449 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001450 "src/xnnpack/packx.h",
1451 "src/xnnpack/pad.h",
1452 "src/xnnpack/params.h",
1453 "src/xnnpack/pavgpool.h",
1454 "src/xnnpack/ppmm.h",
1455 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001456 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001457 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001458 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001459 "src/xnnpack/rmax.h",
1460 "src/xnnpack/scalar-utils.h",
1461 "src/xnnpack/spmm.h",
1462 "src/xnnpack/unpool.h",
1463 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001464 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001465 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001466 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001467 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001468 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001469 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001470 "src/xnnpack/zip.h",
1471]
1472
1473INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1474 "include/xnnpack.h",
1475 "src/xnnpack/allocator.h",
1476 "src/xnnpack/compute.h",
1477 "src/xnnpack/im2col.h",
1478 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001479 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001480 "src/xnnpack/operator.h",
1481 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001482 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001483 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001484 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001485 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001486]
1487
1488ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1489 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001490]
1491
1492MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001493 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001494 "include/xnnpack.h",
1495]
1496
1497MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1498 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001499 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001500 "src/xnnpack/requantization.h",
1501 "include/xnnpack.h",
1502]
1503
1504OPERATOR_TEST_PARAMS_HDRS = [
1505 "src/xnnpack/params.h",
1506 "src/xnnpack/common.h",
1507]
1508
1509WEIGHTS_PACK_HDRS = [
1510 "src/xnnpack/pack.h",
1511 "src/xnnpack/operator.h",
1512 "src/xnnpack/compute.h",
1513]
1514
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001515LOGGING_COPTS = select({
1516 # No logging in optimized mode
1517 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1518 # Full logging in debug mode
1519 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1520 # Error-only logging in default (fastbuild) mode
1521 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1522})
1523
1524LOGGING_HDRS = [
1525 "src/xnnpack/log.h",
1526]
1527
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001529 name = "tables",
1530 srcs = TABLE_SRCS,
1531 hdrs = INTERNAL_HDRS,
1532 copts = xnnpack_std_copts(),
1533)
1534
1535xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001536 name = "scalar_ukernels",
1537 srcs = SCALAR_UKERNELS,
1538 hdrs = INTERNAL_HDRS,
1539 aarch32_copts = ["-marm"],
1540 copts = xnnpack_std_copts(),
1541 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001542 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001543 "@FP16",
1544 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001545 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001546 ],
1547)
1548
1549xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001550 name = "wasm_ukernels",
1551 hdrs = INTERNAL_HDRS,
1552 copts = xnnpack_std_copts(),
1553 wasm_srcs = WASM_UKERNELS,
1554 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001555 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001556 "@FP16",
1557 "@FXdiv",
1558 "@pthreadpool",
1559 ],
1560)
1561
1562xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001563 name = "psimd_fastmath_ukernels",
1564 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001565 hdrs = INTERNAL_HDRS,
1566 aarch32_copts = [
1567 "-marm",
1568 "-mfpu=neon",
1569 ],
1570 copts = xnnpack_std_copts(),
1571 optimized_copts = [
1572 "-O3",
1573 "-ffast-math",
1574 ],
1575 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001576 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001577 "@FP16",
1578 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001579 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580 ],
1581)
1582
1583xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001584 name = "psimd_accmath_ukernels",
1585 srcs = PSIMD_ACCMATH_UKERNELS,
1586 hdrs = INTERNAL_HDRS,
1587 aarch32_copts = [
1588 "-marm",
1589 "-mfpu=neon",
1590 ],
1591 copts = xnnpack_std_copts(),
1592 optimized_copts = [
1593 "-O3",
1594 ],
1595 deps = [
1596 ":tables",
1597 "@FP16",
1598 "@psimd",
1599 "@pthreadpool",
1600 ],
1601)
1602
1603xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604 name = "neon_ukernels",
1605 hdrs = INTERNAL_HDRS,
1606 aarch32_copts = [
1607 "-marm",
1608 "-mfpu=neon",
1609 ],
1610 aarch32_srcs = NEON_UKERNELS,
1611 aarch64_srcs = NEON_UKERNELS,
1612 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001613 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001614 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001615 "@FP16",
1616 "@pthreadpool",
1617 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618)
1619
1620xnnpack_cc_library(
1621 name = "neonfma_ukernels",
1622 hdrs = INTERNAL_HDRS,
1623 aarch32_copts = [
1624 "-marm",
1625 "-mfpu=neon-vfpv4",
1626 ],
1627 aarch32_srcs = NEONFMA_UKERNELS,
1628 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1629 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001630 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001631 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001632 "@FP16",
1633 "@pthreadpool",
1634 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001635)
1636
1637xnnpack_cc_library(
1638 name = "neonfp16arith_ukernels",
1639 hdrs = INTERNAL_HDRS,
1640 aarch64_copts = ["-march=armv8.2-a+fp16"],
1641 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1642 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001643 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001644 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001645 "@FP16",
1646 "@pthreadpool",
1647 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001648)
1649
1650xnnpack_cc_library(
1651 name = "sse2_ukernels",
1652 hdrs = INTERNAL_HDRS,
1653 copts = xnnpack_std_copts(),
1654 x86_copts = ["-msse2"],
1655 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001656 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001657 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001658 "@FP16",
1659 "@pthreadpool",
1660 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001661)
1662
1663xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001664 name = "sse41_ukernels",
1665 hdrs = INTERNAL_HDRS,
1666 copts = xnnpack_std_copts(),
1667 x86_copts = ["-msse4.1"],
1668 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001669 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001670 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001671 "@FP16",
1672 "@pthreadpool",
1673 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001674)
1675
1676xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001677 name = "avx_ukernels",
1678 hdrs = INTERNAL_HDRS,
1679 copts = xnnpack_std_copts(),
1680 x86_copts = ["-mavx"],
1681 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001682 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001683 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001684 "@FP16",
1685 "@pthreadpool",
1686 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001687)
1688
1689xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001690 name = "fma3_ukernels",
1691 hdrs = INTERNAL_HDRS,
1692 copts = xnnpack_std_copts(),
1693 x86_copts = [
1694 "-mfma",
1695 ],
1696 x86_srcs = FMA3_UKERNELS,
1697 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001698 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001699 "@FP16",
1700 "@pthreadpool",
1701 ],
1702)
1703
1704xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001705 name = "avx2_ukernels",
1706 hdrs = INTERNAL_HDRS,
1707 copts = xnnpack_std_copts(),
1708 x86_copts = [
1709 "-mfma",
1710 "-mavx2",
1711 ],
1712 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001713 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001714 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001715 "@FP16",
1716 "@pthreadpool",
1717 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001718)
1719
1720xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001721 name = "avx512f_ukernels",
1722 hdrs = INTERNAL_HDRS,
1723 copts = xnnpack_std_copts(),
1724 x86_copts = ["-mavx512f"],
1725 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001726 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001727 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001728 "@FP16",
1729 "@pthreadpool",
1730 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731)
1732
1733xnnpack_cc_library(
1734 name = "asm_ukernels",
1735 hdrs = ["src/xnnpack/assembly.h"],
1736 aarch32_srcs = AARCH32_ASM_UKERNELS,
1737 aarch64_srcs = AARCH64_ASM_UKERNELS,
1738)
1739
1740xnnpack_aggregate_library(
1741 name = "ukernels",
1742 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001743 ":psimd_fastmath_ukernels",
1744 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001745 ":neon_ukernels",
1746 ":neonfma_ukernels",
1747 ":asm_ukernels",
1748 ],
1749 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001750 ":psimd_fastmath_ukernels",
1751 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001752 ":neon_ukernels",
1753 ":neonfma_ukernels",
1754 ":neonfp16arith_ukernels",
1755 ":asm_ukernels",
1756 ],
1757 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001758 wasm_deps = [
1759 ":wasm_ukernels",
1760 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001761 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001762 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001763 ":psimd_fastmath_ukernels",
1764 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001765 ],
1766 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001767 ":psimd_fastmath_ukernels",
1768 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001770 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001771 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001772 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001773 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001774 ":avx512f_ukernels",
1775 ],
1776)
1777
1778xnnpack_cc_library(
1779 name = "im2col",
1780 srcs = ["src/im2col.c"],
1781 hdrs = [
1782 "src/xnnpack/common.h",
1783 "src/xnnpack/im2col.h",
1784 ],
1785 copts = xnnpack_std_copts(),
1786)
1787
1788xnnpack_cc_library(
1789 name = "indirection",
1790 srcs = ["src/indirection.c"],
1791 hdrs = INTERNAL_HDRS,
1792 copts = xnnpack_std_copts(),
1793 deps = [
1794 "@FP16",
1795 "@FXdiv",
1796 "@pthreadpool",
1797 ],
1798)
1799
1800xnnpack_cc_library(
1801 name = "operator_run",
1802 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001803 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1804 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001805 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1806 "-Wno-vla",
1807 ],
1808 deps = [
1809 "@FP16",
1810 "@FXdiv",
1811 "@clog",
1812 "@pthreadpool",
1813 ],
1814)
1815
1816cc_library(
1817 name = "enable_assembly",
1818 defines = select({
1819 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1820 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001821 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822 }),
1823)
1824
Marat Dukhancf056b22019-10-07 10:26:29 -07001825xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001826 name = "operators",
1827 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001828 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001830 ],
1831 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001832 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833 "-Isrc",
1834 "-Iinclude",
1835 ] + select({
1836 ":debug_build": [],
1837 "//conditions:default": xnnpack_min_size_copts(),
1838 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001839 wasm_srcs = ["src/wasm-stubs.c"],
1840 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001842 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001843 "@FP16",
1844 "@FXdiv",
1845 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001846 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001847 ],
1848)
1849
1850cc_library(
1851 name = "XNNPACK",
1852 srcs = [
1853 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001854 "src/runtime.c",
1855 "src/subgraph.c",
1856 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001857 ],
1858 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1859 "-Isrc",
1860 "-Iinclude",
1861 ] + select({
1862 ":debug_build": [],
1863 "//conditions:default": xnnpack_min_size_copts(),
1864 }),
1865 includes = ["include"],
1866 linkstatic = True,
1867 textual_hdrs = ["include/xnnpack.h"],
1868 visibility = xnnpack_visibility(),
1869 deps = [
1870 ":enable_assembly",
1871 ":ukernels",
1872 ":operator_run",
1873 ":operators",
1874 "@clog",
1875 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001876 ] + select({
1877 ":emscripten": [],
1878 "//conditions:default": ["@cpuinfo"],
1879 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001880)
1881
1882cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001883 name = "xnnpack_operators_nhwc_f32",
1884 srcs = [
1885 "src/init.c",
1886 ],
1887 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1888 "-Isrc",
1889 "-Iinclude",
1890 ] + select({
1891 ":debug_build": [],
1892 "//conditions:default": xnnpack_min_size_copts(),
1893 }),
1894 defines = [
1895 "XNN_NO_Q8_OPERATORS",
1896 "XNN_NO_U8_OPERATORS",
1897 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001898 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001899 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900 includes = ["include"],
1901 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001902 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001903 visibility = xnnpack_visibility(),
1904 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001905 ":enable_assembly",
1906 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001907 ":operator_run",
1908 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001909 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001910 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001911 ] + select({
1912 ":emscripten": [],
1913 "//conditions:default": ["@cpuinfo"],
1914 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001915)
1916
Marat Dukhancf056b22019-10-07 10:26:29 -07001917xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918 name = "bench_utils",
1919 srcs = ["bench/utils.cc"],
1920 hdrs = ["bench/utils.h"],
1921 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001922 deps = [
1923 "@com_google_benchmark//:benchmark",
1924 "@cpuinfo",
1925 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001926)
1927
Frank Barchard7e955972019-10-11 10:34:25 -07001928######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001929
1930xnnpack_benchmark(
1931 name = "q8_gemm_bench",
1932 srcs = [
1933 "bench/gemm.h",
1934 "bench/q8-gemm.cc",
1935 "src/xnnpack/AlignedAllocator.h",
1936 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1937 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1938 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1939)
1940
1941xnnpack_benchmark(
1942 name = "f16_gemm_bench",
1943 srcs = [
1944 "bench/f16-gemm.cc",
1945 "bench/gemm.h",
1946 "src/xnnpack/AlignedAllocator.h",
1947 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1948 copts = ["-Wno-unused-function"],
1949 deps = MICROKERNEL_BENCHMARK_DEPS,
1950)
1951
1952xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08001953 name = "f16_spmm_bench",
1954 srcs = [
1955 "bench/f16-spmm.cc",
1956 "bench/gemm.h",
1957 "bench/google/gemm.h",
1958 "src/xnnpack/AlignedAllocator.h",
1959 ] + MICROKERNEL_BENCHMARK_HDRS,
1960 copts = ["-Wno-unused-function"],
1961 deps = MICROKERNEL_BENCHMARK_DEPS,
1962)
1963
1964xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001965 name = "f32_igemm_bench",
1966 srcs = [
1967 "bench/f32-igemm.cc",
1968 "bench/conv.h",
1969 "src/xnnpack/AlignedAllocator.h",
1970 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001971 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001972)
1973
1974xnnpack_benchmark(
1975 name = "f32_conv_hwc_bench",
1976 srcs = [
1977 "bench/f32-conv-hwc.cc",
1978 "bench/dconv.h",
1979 "src/xnnpack/AlignedAllocator.h",
1980 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1981 copts = ["-Wno-unused-function"],
1982 deps = MICROKERNEL_BENCHMARK_DEPS,
1983)
1984
1985xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001986 name = "f32_conv_hwc2spchw_bench",
1987 srcs = [
1988 "bench/f32-conv-hwc2spchw.cc",
1989 "bench/dconv.h",
1990 "src/xnnpack/AlignedAllocator.h",
1991 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1992 copts = ["-Wno-unused-function"],
1993 deps = MICROKERNEL_BENCHMARK_DEPS,
1994)
1995
1996xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 name = "f32_dwconv_bench",
1998 srcs = [
1999 "bench/f32-dwconv.cc",
2000 "bench/dwconv.h",
2001 "src/xnnpack/AlignedAllocator.h",
2002 ] + MICROKERNEL_BENCHMARK_HDRS,
2003 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2004)
2005
2006xnnpack_benchmark(
2007 name = "f32_dwconv_spchw_bench",
2008 srcs = [
2009 "bench/f32-dwconv-spchw.cc",
2010 "bench/dwconv.h",
2011 "src/xnnpack/AlignedAllocator.h",
2012 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2013 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2014)
2015
2016xnnpack_benchmark(
2017 name = "f32_gemm_bench",
2018 srcs = [
2019 "bench/f32-gemm.cc",
2020 "bench/gemm.h",
2021 "src/xnnpack/AlignedAllocator.h",
2022 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2023 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002024 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025)
2026
2027xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002028 name = "f32_raddexpminusmax_bench",
2029 srcs = [
2030 "bench/f32-raddexpminusmax.cc",
2031 "src/xnnpack/AlignedAllocator.h",
2032 ] + MICROKERNEL_BENCHMARK_HDRS,
2033 deps = MICROKERNEL_BENCHMARK_DEPS,
2034)
2035
2036xnnpack_benchmark(
2037 name = "f32_raddextexp_bench",
2038 srcs = [
2039 "bench/f32-raddextexp.cc",
2040 "src/xnnpack/AlignedAllocator.h",
2041 ] + MICROKERNEL_BENCHMARK_HDRS,
2042 deps = MICROKERNEL_BENCHMARK_DEPS,
2043)
2044
2045xnnpack_benchmark(
2046 name = "f32_raddstoreexpminusmax_bench",
2047 srcs = [
2048 "bench/f32-raddstoreexpminusmax.cc",
2049 "src/xnnpack/AlignedAllocator.h",
2050 ] + MICROKERNEL_BENCHMARK_HDRS,
2051 deps = MICROKERNEL_BENCHMARK_DEPS,
2052)
2053
2054xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002055 name = "f32_rmax_bench",
2056 srcs = [
2057 "bench/f32-rmax.cc",
2058 "src/xnnpack/AlignedAllocator.h",
2059 ] + MICROKERNEL_BENCHMARK_HDRS,
2060 deps = MICROKERNEL_BENCHMARK_DEPS,
2061)
2062
2063xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002064 name = "f32_sigmoid_bench",
2065 srcs = [
2066 "bench/f32-sigmoid.cc",
2067 "src/xnnpack/AlignedAllocator.h",
2068 ] + MICROKERNEL_BENCHMARK_HDRS,
2069 copts = ["-Wno-unused-function"],
2070 deps = MICROKERNEL_BENCHMARK_DEPS,
2071)
2072
2073xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002074 name = "f32_spmm_bench",
2075 srcs = [
2076 "bench/f32-spmm.cc",
2077 "bench/gemm.h",
2078 "src/xnnpack/AlignedAllocator.h",
2079 ] + MICROKERNEL_BENCHMARK_HDRS,
2080 copts = ["-Wno-unused-function"],
2081 deps = MICROKERNEL_BENCHMARK_DEPS,
2082)
2083
2084xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002085 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002086 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002087 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002088 ] + MICROKERNEL_BENCHMARK_HDRS,
2089 copts = ["-Wno-unused-function"],
2090 deps = MICROKERNEL_BENCHMARK_DEPS,
2091)
2092
2093xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002094 name = "f32_vscaleexpminusmax_bench",
2095 srcs = [
2096 "bench/f32-vscaleexpminusmax.cc",
2097 "src/xnnpack/AlignedAllocator.h",
2098 ] + MICROKERNEL_BENCHMARK_HDRS,
2099 deps = MICROKERNEL_BENCHMARK_DEPS,
2100)
2101
2102xnnpack_benchmark(
2103 name = "f32_vscaleextexp_bench",
2104 srcs = [
2105 "bench/f32-vscaleextexp.cc",
2106 "src/xnnpack/AlignedAllocator.h",
2107 ] + MICROKERNEL_BENCHMARK_HDRS,
2108 deps = MICROKERNEL_BENCHMARK_DEPS,
2109)
2110
2111xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112 name = "f32_im2col_gemm_bench",
2113 srcs = [
2114 "bench/f32-im2col-gemm.cc",
2115 "bench/conv.h",
2116 "src/xnnpack/AlignedAllocator.h",
2117 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2118 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2119)
2120
2121########################### Benchmarks for operators ###########################
2122
2123xnnpack_benchmark(
2124 name = "add_bench",
2125 srcs = ["bench/add.cc"],
2126 deps = OPERATOR_BENCHMARK_DEPS,
2127)
2128
2129xnnpack_benchmark(
2130 name = "average_pooling_bench",
2131 srcs = ["bench/average-pooling.cc"],
2132 deps = OPERATOR_BENCHMARK_DEPS,
2133)
2134
2135xnnpack_benchmark(
2136 name = "channel_shuffle_bench",
2137 srcs = ["bench/channel-shuffle.cc"],
2138 deps = OPERATOR_BENCHMARK_DEPS,
2139)
2140
2141xnnpack_benchmark(
2142 name = "convolution_bench",
2143 srcs = ["bench/convolution.cc"],
2144 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2145 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2146)
2147
2148xnnpack_benchmark(
2149 name = "deconvolution_bench",
2150 srcs = ["bench/deconvolution.cc"],
2151 copts = xnnpack_optional_tflite_copts(),
2152 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2153)
2154
2155xnnpack_benchmark(
2156 name = "global_average_pooling_bench",
2157 srcs = ["bench/global-average-pooling.cc"],
2158 deps = OPERATOR_BENCHMARK_DEPS,
2159)
2160
2161xnnpack_benchmark(
2162 name = "max_pooling_bench",
2163 srcs = ["bench/max-pooling.cc"],
2164 deps = OPERATOR_BENCHMARK_DEPS,
2165)
2166
2167xnnpack_benchmark(
2168 name = "sigmoid_bench",
2169 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002170 copts = xnnpack_optional_tflite_copts(),
2171 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002172)
2173
2174xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002175 name = "prelu_bench",
2176 srcs = ["bench/prelu.cc"],
2177 copts = xnnpack_optional_tflite_copts(),
2178 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2179)
2180
2181xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002182 name = "softmax_bench",
2183 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002184 copts = xnnpack_optional_tflite_copts(),
2185 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002186)
2187
Marat Dukhanc068bb62019-10-04 13:24:39 -07002188############################# End-to-end benchmarks ############################
2189
2190cc_library(
2191 name = "mobilenet_v1",
2192 srcs = ["models/mobilenet-v1.cc"],
2193 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002194 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002195 linkstatic = True,
2196 deps = [
2197 ":XNNPACK",
2198 "@pthreadpool",
2199 ],
2200)
2201
2202cc_library(
2203 name = "mobilenet_v2",
2204 srcs = ["models/mobilenet-v2.cc"],
2205 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002206 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002207 linkstatic = True,
2208 deps = [
2209 ":XNNPACK",
2210 "@pthreadpool",
2211 ],
2212)
2213
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002214cc_library(
2215 name = "mobilenet_v3_large",
2216 srcs = ["models/mobilenet-v3-large.cc"],
2217 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002218 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002219 linkstatic = True,
2220 deps = [
2221 ":XNNPACK",
2222 "@pthreadpool",
2223 ],
2224)
2225
2226cc_library(
2227 name = "mobilenet_v3_small",
2228 srcs = ["models/mobilenet-v3-small.cc"],
2229 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002230 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002231 linkstatic = True,
2232 deps = [
2233 ":XNNPACK",
2234 "@pthreadpool",
2235 ],
2236)
2237
Marat Dukhanc068bb62019-10-04 13:24:39 -07002238xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002239 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002240 srcs = [
2241 "bench/f32-dwconv-e2e.cc",
2242 "bench/end2end.h",
2243 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002244 copts = ["-Wno-unused-function"],
2245 deps = MICROKERNEL_BENCHMARK_DEPS + [
2246 ":XNNPACK",
2247 ":mobilenet_v1",
2248 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002249 ":mobilenet_v3_large",
2250 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002251 ],
2252)
2253
2254xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002255 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002256 srcs = [
2257 "bench/f32-gemm-e2e.cc",
2258 "bench/end2end.h",
2259 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002260 copts = ["-Wno-unused-function"],
2261 deps = MICROKERNEL_BENCHMARK_DEPS + [
2262 ":XNNPACK",
2263 ":mobilenet_v1",
2264 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002265 ":mobilenet_v3_large",
2266 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002267 ],
2268)
2269
2270xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002271 name = "end2end_bench",
2272 srcs = ["bench/end2end.cc"],
2273 deps = [
2274 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002275 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002276 ":mobilenet_v1",
2277 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002278 ":mobilenet_v3_large",
2279 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002280 "@pthreadpool",
2281 ],
2282)
2283
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002284#################### Accuracy evaluation for math functions ####################
2285
2286xnnpack_benchmark(
2287 name = "f32_exp_eval",
2288 srcs = [
2289 "eval/f32-exp.cc",
2290 "src/xnnpack/AlignedAllocator.h",
2291 ] + ACCURACY_EVAL_HDRS,
2292 deps = ACCURACY_EVAL_DEPS,
2293)
2294
Marat Dukhan515c9772019-10-17 18:07:57 -07002295xnnpack_benchmark(
2296 name = "f32_expminus_eval",
2297 srcs = [
2298 "eval/f32-expminus.cc",
2299 "src/xnnpack/AlignedAllocator.h",
2300 ] + ACCURACY_EVAL_HDRS,
2301 deps = ACCURACY_EVAL_DEPS,
2302)
2303
Marat Dukhan98ba4412019-10-23 02:14:28 -07002304xnnpack_benchmark(
2305 name = "f32_extexp_eval",
2306 srcs = [
2307 "eval/f32-extexp.cc",
2308 "src/xnnpack/AlignedAllocator.h",
2309 ] + ACCURACY_EVAL_HDRS,
2310 deps = ACCURACY_EVAL_DEPS,
2311)
2312
Marat Dukhan346a9e52019-11-15 09:06:30 -08002313xnnpack_benchmark(
2314 name = "f32_sigmoid_eval",
2315 srcs = [
2316 "eval/f32-sigmoid.cc",
2317 "src/xnnpack/AlignedAllocator.h",
2318 ] + ACCURACY_EVAL_HDRS,
2319 deps = ACCURACY_EVAL_DEPS,
2320)
2321
Marat Dukhan08c4a432019-10-03 09:29:21 -07002322######################### Unit tests for micro-kernels #########################
2323
2324xnnpack_unit_test(
2325 name = "f16_gemm_test",
2326 srcs = [
2327 "test/f16-gemm.cc",
2328 "test/gemm-microkernel-tester.h",
2329 "src/xnnpack/AlignedAllocator.h",
2330 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2331 deps = MICROKERNEL_TEST_DEPS,
2332)
2333
2334xnnpack_unit_test(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002335 name = "f16_spmm_test",
2336 srcs = [
2337 "test/f16-spmm.cc",
2338 "test/spmm-microkernel-tester.h",
2339 "src/xnnpack/AlignedAllocator.h",
2340 ] + MICROKERNEL_TEST_HDRS,
2341 deps = MICROKERNEL_TEST_DEPS,
2342)
2343
2344xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345 name = "f32_argmaxpool_test",
2346 srcs = [
2347 "test/f32-argmaxpool.cc",
2348 "test/argmaxpool-microkernel-tester.h",
2349 "src/xnnpack/AlignedAllocator.h",
2350 ] + MICROKERNEL_TEST_HDRS,
2351 deps = MICROKERNEL_TEST_DEPS,
2352)
2353
2354xnnpack_unit_test(
2355 name = "f32_avgpool_test",
2356 srcs = [
2357 "test/f32-avgpool.cc",
2358 "test/avgpool-microkernel-tester.h",
2359 "src/xnnpack/AlignedAllocator.h",
2360 ] + MICROKERNEL_TEST_HDRS,
2361 deps = MICROKERNEL_TEST_DEPS,
2362)
2363
2364xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002365 name = "f32_bilinear_test",
2366 srcs = [
2367 "test/f32-bilinear.cc",
2368 "test/bilinear-microkernel-tester.h",
2369 "src/xnnpack/AlignedAllocator.h",
2370 ] + MICROKERNEL_TEST_HDRS,
2371 deps = MICROKERNEL_TEST_DEPS,
2372)
2373
2374xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002375 name = "f32_clamp_test",
2376 srcs = [
2377 "test/f32-clamp.cc",
2378 "test/clamp-microkernel-tester.h",
2379 ] + MICROKERNEL_TEST_HDRS,
2380 deps = MICROKERNEL_TEST_DEPS,
2381)
2382
2383xnnpack_unit_test(
2384 name = "f32_igemm_test",
2385 srcs = [
2386 "test/f32-igemm.cc",
2387 "test/gemm-microkernel-tester.h",
2388 "src/xnnpack/AlignedAllocator.h",
2389 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2390 deps = MICROKERNEL_TEST_DEPS,
2391)
2392
2393xnnpack_unit_test(
2394 name = "f32_conv_hwc_test",
2395 srcs = [
2396 "test/f32-conv-hwc.cc",
2397 "test/conv-hwc-microkernel-tester.h",
2398 "src/xnnpack/AlignedAllocator.h",
2399 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2400 deps = MICROKERNEL_TEST_DEPS,
2401)
2402
2403xnnpack_unit_test(
2404 name = "f32_conv_hwc2spchw_test",
2405 srcs = [
2406 "test/f32-conv-hwc2spchw.cc",
2407 "test/conv-hwc2spchw-microkernel-tester.h",
2408 "src/xnnpack/AlignedAllocator.h",
2409 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2410 deps = MICROKERNEL_TEST_DEPS,
2411)
2412
2413xnnpack_unit_test(
2414 name = "f32_dwconv_test",
2415 srcs = [
2416 "test/f32-dwconv.cc",
2417 "test/dwconv-microkernel-tester.h",
2418 "src/xnnpack/AlignedAllocator.h",
2419 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2420 deps = MICROKERNEL_TEST_DEPS,
2421)
2422
2423xnnpack_unit_test(
2424 name = "f32_dwconv_spchw_test",
2425 srcs = [
2426 "test/f32-dwconv-spchw.cc",
2427 "test/dwconv-spchw-microkernel-tester.h",
2428 "src/xnnpack/AlignedAllocator.h",
2429 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2430 deps = MICROKERNEL_TEST_DEPS,
2431)
2432
2433xnnpack_unit_test(
2434 name = "f32_gavgpool_test",
2435 srcs = [
2436 "test/f32-gavgpool.cc",
2437 "test/gavgpool-microkernel-tester.h",
2438 "src/xnnpack/AlignedAllocator.h",
2439 ] + MICROKERNEL_TEST_HDRS,
2440 deps = MICROKERNEL_TEST_DEPS,
2441)
2442
2443xnnpack_unit_test(
2444 name = "f32_gavgpool_spchw_test",
2445 srcs = [
2446 "test/f32-gavgpool-spchw.cc",
2447 "test/gavgpool-spchw-microkernel-tester.h",
2448 "src/xnnpack/AlignedAllocator.h",
2449 ] + MICROKERNEL_TEST_HDRS,
2450 deps = MICROKERNEL_TEST_DEPS,
2451)
2452
2453xnnpack_unit_test(
2454 name = "f32_gemm_test",
2455 srcs = [
2456 "test/f32-gemm.cc",
2457 "test/gemm-microkernel-tester.h",
2458 "src/xnnpack/AlignedAllocator.h",
2459 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2460 deps = MICROKERNEL_TEST_DEPS,
2461)
2462
2463xnnpack_unit_test(
2464 name = "f32_gemminc_test",
2465 srcs = [
2466 "test/f32-gemminc.cc",
2467 "test/gemm-microkernel-tester.h",
2468 "src/xnnpack/AlignedAllocator.h",
2469 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2470 deps = MICROKERNEL_TEST_DEPS,
2471)
2472
2473xnnpack_unit_test(
2474 name = "f32_hswish_test",
2475 srcs = [
2476 "test/f32-hswish.cc",
2477 "test/hswish-microkernel-tester.h",
2478 ] + MICROKERNEL_TEST_HDRS,
2479 deps = MICROKERNEL_TEST_DEPS,
2480)
2481
2482xnnpack_unit_test(
2483 name = "f32_maxpool_test",
2484 srcs = [
2485 "test/f32-maxpool.cc",
2486 "test/maxpool-microkernel-tester.h",
2487 ] + MICROKERNEL_TEST_HDRS,
2488 deps = MICROKERNEL_TEST_DEPS,
2489)
2490
2491xnnpack_unit_test(
2492 name = "f32_pavgpool_test",
2493 srcs = [
2494 "test/f32-pavgpool.cc",
2495 "test/avgpool-microkernel-tester.h",
2496 "src/xnnpack/AlignedAllocator.h",
2497 ] + MICROKERNEL_TEST_HDRS,
2498 deps = MICROKERNEL_TEST_DEPS,
2499)
2500
2501xnnpack_unit_test(
2502 name = "f32_ppmm_test",
2503 srcs = [
2504 "test/f32-ppmm.cc",
2505 "test/gemm-microkernel-tester.h",
2506 "src/xnnpack/AlignedAllocator.h",
2507 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2508 deps = MICROKERNEL_TEST_DEPS,
2509)
2510
2511xnnpack_unit_test(
2512 name = "f32_prelu_test",
2513 srcs = [
2514 "test/f32-prelu.cc",
2515 "test/prelu-microkernel-tester.h",
2516 "src/xnnpack/AlignedAllocator.h",
2517 ] + MICROKERNEL_TEST_HDRS,
2518 deps = MICROKERNEL_TEST_DEPS,
2519)
2520
2521xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002522 name = "f32_raddexpminusmax_test",
2523 srcs = [
2524 "test/f32-raddexpminusmax.cc",
2525 "test/raddexpminusmax-microkernel-tester.h",
2526 ] + MICROKERNEL_TEST_HDRS,
2527 deps = MICROKERNEL_TEST_DEPS,
2528)
2529
2530xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002531 name = "f32_raddextexp_test",
2532 srcs = [
2533 "test/f32-raddextexp.cc",
2534 "test/raddextexp-microkernel-tester.h",
2535 ] + MICROKERNEL_TEST_HDRS,
2536 deps = MICROKERNEL_TEST_DEPS,
2537)
2538
2539xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002540 name = "f32_raddstoreexpminusmax_test",
2541 srcs = [
2542 "test/f32-raddstoreexpminusmax.cc",
2543 "test/raddstoreexpminusmax-microkernel-tester.h",
2544 ] + MICROKERNEL_TEST_HDRS,
2545 deps = MICROKERNEL_TEST_DEPS,
2546)
2547
2548xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002549 name = "f32_rmax_test",
2550 srcs = [
2551 "test/f32-rmax.cc",
2552 "test/rmax-microkernel-tester.h",
2553 ] + MICROKERNEL_TEST_HDRS,
2554 deps = MICROKERNEL_TEST_DEPS,
2555)
2556
2557xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002558 name = "f32_sigmoid_test",
2559 srcs = [
2560 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002561 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002562 ] + MICROKERNEL_TEST_HDRS,
2563 deps = MICROKERNEL_TEST_DEPS,
2564)
2565
2566xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002567 name = "f32_spmm_test",
2568 srcs = [
2569 "test/f32-spmm.cc",
2570 "test/spmm-microkernel-tester.h",
2571 "src/xnnpack/AlignedAllocator.h",
2572 ] + MICROKERNEL_TEST_HDRS,
2573 deps = MICROKERNEL_TEST_DEPS,
2574)
2575
2576xnnpack_unit_test(
2577 name = "f32_vadd_test",
2578 srcs = [
2579 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002580 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002581 ] + MICROKERNEL_TEST_HDRS,
2582 deps = MICROKERNEL_TEST_DEPS,
2583)
2584
2585xnnpack_unit_test(
2586 name = "f32_vaddc_test",
2587 srcs = [
2588 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002589 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002590 ] + MICROKERNEL_TEST_HDRS,
2591 deps = MICROKERNEL_TEST_DEPS,
2592)
2593
2594xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002595 name = "f32_vdiv_test",
2596 srcs = [
2597 "test/f32-vdiv.cc",
2598 "test/vbinary-microkernel-tester.h",
2599 ] + MICROKERNEL_TEST_HDRS,
2600 deps = MICROKERNEL_TEST_DEPS,
2601)
2602
2603xnnpack_unit_test(
2604 name = "f32_vdivc_test",
2605 srcs = [
2606 "test/f32-vdivc.cc",
2607 "test/vbinaryc-microkernel-tester.h",
2608 ] + MICROKERNEL_TEST_HDRS,
2609 deps = MICROKERNEL_TEST_DEPS,
2610)
2611
2612xnnpack_unit_test(
2613 name = "f32_vrdivc_test",
2614 srcs = [
2615 "test/f32-vrdivc.cc",
2616 "test/vbinaryc-microkernel-tester.h",
2617 ] + MICROKERNEL_TEST_HDRS,
2618 deps = MICROKERNEL_TEST_DEPS,
2619)
2620
2621xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002622 name = "f32_vmax_test",
2623 srcs = [
2624 "test/f32-vmax.cc",
2625 "test/vbinary-microkernel-tester.h",
2626 ] + MICROKERNEL_TEST_HDRS,
2627 deps = MICROKERNEL_TEST_DEPS,
2628)
2629
2630xnnpack_unit_test(
2631 name = "f32_vmaxc_test",
2632 srcs = [
2633 "test/f32-vmaxc.cc",
2634 "test/vbinaryc-microkernel-tester.h",
2635 ] + MICROKERNEL_TEST_HDRS,
2636 deps = MICROKERNEL_TEST_DEPS,
2637)
2638
2639xnnpack_unit_test(
2640 name = "f32_vmin_test",
2641 srcs = [
2642 "test/f32-vmin.cc",
2643 "test/vbinary-microkernel-tester.h",
2644 ] + MICROKERNEL_TEST_HDRS,
2645 deps = MICROKERNEL_TEST_DEPS,
2646)
2647
2648xnnpack_unit_test(
2649 name = "f32_vminc_test",
2650 srcs = [
2651 "test/f32-vminc.cc",
2652 "test/vbinaryc-microkernel-tester.h",
2653 ] + MICROKERNEL_TEST_HDRS,
2654 deps = MICROKERNEL_TEST_DEPS,
2655)
2656
2657xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002658 name = "f32_vmul_test",
2659 srcs = [
2660 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002661 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002662 ] + MICROKERNEL_TEST_HDRS,
2663 deps = MICROKERNEL_TEST_DEPS,
2664)
2665
2666xnnpack_unit_test(
2667 name = "f32_vmulc_test",
2668 srcs = [
2669 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002670 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002671 ] + MICROKERNEL_TEST_HDRS,
2672 deps = MICROKERNEL_TEST_DEPS,
2673)
2674
2675xnnpack_unit_test(
2676 name = "f32_vmulcaddc_test",
2677 srcs = [
2678 "test/f32-vmulcaddc.cc",
2679 "test/vmulcaddc-microkernel-tester.h",
2680 "src/xnnpack/AlignedAllocator.h",
2681 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2682 deps = MICROKERNEL_TEST_DEPS,
2683)
2684
2685xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002686 name = "f32_vscale_test",
2687 srcs = [
2688 "test/f32-vscale.cc",
2689 "test/vscale-microkernel-tester.h",
2690 ] + MICROKERNEL_TEST_HDRS,
2691 deps = MICROKERNEL_TEST_DEPS,
2692)
2693
2694xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002695 name = "f32_vscaleexpminusmax_test",
2696 srcs = [
2697 "test/f32-vscaleexpminusmax.cc",
2698 "test/vscaleexpminusmax-microkernel-tester.h",
2699 ] + MICROKERNEL_TEST_HDRS,
2700 deps = MICROKERNEL_TEST_DEPS,
2701)
2702
2703xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002704 name = "f32_vscaleextexp_test",
2705 srcs = [
2706 "test/f32-vscaleextexp.cc",
2707 "test/vscaleextexp-microkernel-tester.h",
2708 ] + MICROKERNEL_TEST_HDRS,
2709 deps = MICROKERNEL_TEST_DEPS,
2710)
2711
2712xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002713 name = "f32_vsub_test",
2714 srcs = [
2715 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002716 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002717 ] + MICROKERNEL_TEST_HDRS,
2718 deps = MICROKERNEL_TEST_DEPS,
2719)
2720
2721xnnpack_unit_test(
2722 name = "f32_vsubc_test",
2723 srcs = [
2724 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002725 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002726 ] + MICROKERNEL_TEST_HDRS,
2727 deps = MICROKERNEL_TEST_DEPS,
2728)
2729
2730xnnpack_unit_test(
2731 name = "f32_vrsubc_test",
2732 srcs = [
2733 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002734 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002735 ] + MICROKERNEL_TEST_HDRS,
2736 deps = MICROKERNEL_TEST_DEPS,
2737)
2738
2739xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 name = "q8_avgpool_test",
2741 srcs = [
2742 "test/q8-avgpool.cc",
2743 "test/avgpool-microkernel-tester.h",
2744 "src/xnnpack/AlignedAllocator.h",
2745 ] + MICROKERNEL_TEST_HDRS,
2746 deps = MICROKERNEL_TEST_DEPS,
2747)
2748
2749xnnpack_unit_test(
2750 name = "q8_igemm_test",
2751 srcs = [
2752 "test/q8-igemm.cc",
2753 "test/gemm-microkernel-tester.h",
2754 "src/xnnpack/AlignedAllocator.h",
2755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2756 deps = MICROKERNEL_TEST_DEPS,
2757)
2758
2759xnnpack_unit_test(
2760 name = "q8_dwconv_test",
2761 srcs = [
2762 "test/q8-dwconv.cc",
2763 "test/dwconv-microkernel-tester.h",
2764 "src/xnnpack/AlignedAllocator.h",
2765 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2766 deps = MICROKERNEL_TEST_DEPS,
2767)
2768
2769xnnpack_unit_test(
2770 name = "q8_gavgpool_test",
2771 srcs = [
2772 "test/q8-gavgpool.cc",
2773 "test/gavgpool-microkernel-tester.h",
2774 "src/xnnpack/AlignedAllocator.h",
2775 ] + MICROKERNEL_TEST_HDRS,
2776 deps = MICROKERNEL_TEST_DEPS,
2777)
2778
2779xnnpack_unit_test(
2780 name = "q8_gemm_test",
2781 srcs = [
2782 "test/q8-gemm.cc",
2783 "test/gemm-microkernel-tester.h",
2784 "src/xnnpack/AlignedAllocator.h",
2785 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2786 deps = MICROKERNEL_TEST_DEPS,
2787)
2788
2789xnnpack_unit_test(
2790 name = "q8_vadd_test",
2791 srcs = [
2792 "test/q8-vadd.cc",
2793 "test/vadd-microkernel-tester.h",
2794 ] + MICROKERNEL_TEST_HDRS,
2795 deps = MICROKERNEL_TEST_DEPS,
2796)
2797
2798xnnpack_unit_test(
2799 name = "u8_clamp_test",
2800 srcs = [
2801 "test/u8-clamp.cc",
2802 "test/clamp-microkernel-tester.h",
2803 ] + MICROKERNEL_TEST_HDRS,
2804 deps = MICROKERNEL_TEST_DEPS,
2805)
2806
2807xnnpack_unit_test(
2808 name = "u8_lut32norm_test",
2809 srcs = [
2810 "test/u8-lut32norm.cc",
2811 "test/lut-norm-microkernel-tester.h",
2812 ] + MICROKERNEL_TEST_HDRS,
2813 deps = MICROKERNEL_TEST_DEPS,
2814)
2815
2816xnnpack_unit_test(
2817 name = "u8_maxpool_test",
2818 srcs = [
2819 "test/u8-maxpool.cc",
2820 "test/maxpool-microkernel-tester.h",
2821 ] + MICROKERNEL_TEST_HDRS,
2822 deps = MICROKERNEL_TEST_DEPS,
2823)
2824
2825xnnpack_unit_test(
2826 name = "u8_rmax_test",
2827 srcs = [
2828 "test/u8-rmax.cc",
2829 "test/rmax-microkernel-tester.h",
2830 ] + MICROKERNEL_TEST_HDRS,
2831 deps = MICROKERNEL_TEST_DEPS,
2832)
2833
2834xnnpack_unit_test(
2835 name = "x32_packx_test",
2836 srcs = [
2837 "test/x32-packx.cc",
2838 "test/pack-microkernel-tester.h",
2839 "src/xnnpack/AlignedAllocator.h",
2840 ] + MICROKERNEL_TEST_HDRS,
2841 deps = MICROKERNEL_TEST_DEPS,
2842)
2843
2844xnnpack_unit_test(
2845 name = "x32_pad_test",
2846 srcs = [
2847 "test/x32-pad.cc",
2848 "test/pad-microkernel-tester.h",
2849 ] + MICROKERNEL_TEST_HDRS,
2850 deps = MICROKERNEL_TEST_DEPS,
2851)
2852
2853xnnpack_unit_test(
2854 name = "x32_unpool_test",
2855 srcs = [
2856 "test/x32-unpool.cc",
2857 "test/unpool-microkernel-tester.h",
2858 ] + MICROKERNEL_TEST_HDRS,
2859 deps = MICROKERNEL_TEST_DEPS,
2860)
2861
2862xnnpack_unit_test(
2863 name = "x32_zip_test",
2864 srcs = [
2865 "test/x32-zip.cc",
2866 "test/zip-microkernel-tester.h",
2867 ] + MICROKERNEL_TEST_HDRS,
2868 deps = MICROKERNEL_TEST_DEPS,
2869)
2870
2871xnnpack_unit_test(
2872 name = "x8_lut_test",
2873 srcs = [
2874 "test/x8-lut.cc",
2875 "test/lut-microkernel-tester.h",
2876 ] + MICROKERNEL_TEST_HDRS,
2877 deps = MICROKERNEL_TEST_DEPS,
2878)
2879
2880xnnpack_unit_test(
2881 name = "x8_zip_test",
2882 srcs = [
2883 "test/x8-zip.cc",
2884 "test/zip-microkernel-tester.h",
2885 ] + MICROKERNEL_TEST_HDRS,
2886 deps = MICROKERNEL_TEST_DEPS,
2887)
2888
2889########################### Size test for the library ##########################
2890
2891xnnpack_binary(
2892 name = "size_test",
2893 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002894 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002895)
2896
2897########################### Unit tests for operators ###########################
2898
2899xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002900 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002902 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 "test/add-operator-tester.h",
2904 ],
2905 deps = OPERATOR_TEST_DEPS,
2906)
2907
2908xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002909 name = "add_nd_test",
2910 srcs = [
2911 "test/add-nd.cc",
2912 "test/binary-elementwise-operator-tester.h",
2913 ],
2914 deps = OPERATOR_TEST_DEPS,
2915)
2916
2917xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002918 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002920 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 "test/argmax-pooling-operator-tester.h",
2922 ] + OPERATOR_TEST_PARAMS_HDRS,
2923 deps = OPERATOR_TEST_DEPS,
2924)
2925
2926xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002927 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002929 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 "test/average-pooling-operator-tester.h",
2931 ] + OPERATOR_TEST_PARAMS_HDRS,
2932 deps = OPERATOR_TEST_DEPS,
2933)
2934
2935xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002936 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002938 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002939 "test/channel-pad-operator-tester.h",
2940 ] + OPERATOR_TEST_PARAMS_HDRS,
2941 deps = OPERATOR_TEST_DEPS,
2942)
2943
2944xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002945 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002947 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948 "test/channel-shuffle-operator-tester.h",
2949 ],
2950 deps = OPERATOR_TEST_DEPS,
2951)
2952
2953xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002954 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002955 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002956 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002957 "test/clamp-operator-tester.h",
2958 ],
2959 deps = OPERATOR_TEST_DEPS,
2960)
2961
2962xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002963 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002965 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 "test/convolution-operator-tester.h",
2967 ],
2968 deps = OPERATOR_TEST_DEPS,
2969)
2970
2971xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002972 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002973 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002974 "test/convolution-nchw.cc",
2975 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976 ],
2977 deps = OPERATOR_TEST_DEPS,
2978)
2979
2980xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002981 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002983 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002984 "test/deconvolution-operator-tester.h",
2985 ] + OPERATOR_TEST_PARAMS_HDRS,
2986 deps = OPERATOR_TEST_DEPS,
2987)
2988
2989xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002990 name = "divide_nd_test",
2991 srcs = [
2992 "test/binary-elementwise-operator-tester.h",
2993 "test/divide-nd.cc",
2994 ],
2995 deps = OPERATOR_TEST_DEPS,
2996)
2997
2998xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002999 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003001 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 "test/fully-connected-operator-tester.h",
3003 ],
3004 deps = OPERATOR_TEST_DEPS,
3005)
3006
3007xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003008 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003010 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 "test/global-average-pooling-operator-tester.h",
3012 ] + OPERATOR_TEST_PARAMS_HDRS,
3013 deps = OPERATOR_TEST_DEPS,
3014)
3015
3016xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003017 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003019 "test/global-average-pooling-ncw.cc",
3020 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 ],
3022 deps = OPERATOR_TEST_DEPS,
3023)
3024
3025xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003026 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003028 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029 "test/hardswish-operator-tester.h",
3030 ],
3031 deps = OPERATOR_TEST_DEPS,
3032)
3033
3034xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003035 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003036 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003037 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003038 "test/leaky-relu-operator-tester.h",
3039 ],
3040 deps = OPERATOR_TEST_DEPS,
3041)
3042
3043xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003044 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003046 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003047 "test/max-pooling-operator-tester.h",
3048 ] + OPERATOR_TEST_PARAMS_HDRS,
3049 deps = OPERATOR_TEST_DEPS,
3050)
3051
3052xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003053 name = "maximum_nd_test",
3054 srcs = [
3055 "test/binary-elementwise-operator-tester.h",
3056 "test/maximum-nd.cc",
3057 ],
3058 deps = OPERATOR_TEST_DEPS,
3059)
3060
3061xnnpack_unit_test(
3062 name = "minimum_nd_test",
3063 srcs = [
3064 "test/binary-elementwise-operator-tester.h",
3065 "test/minimum-nd.cc",
3066 ],
3067 deps = OPERATOR_TEST_DEPS,
3068)
3069
3070xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003071 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003072 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003073 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003074 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003075 ],
3076 deps = OPERATOR_TEST_DEPS,
3077)
3078
3079xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003080 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003081 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003082 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083 "test/prelu-operator-tester.h",
3084 ] + OPERATOR_TEST_PARAMS_HDRS,
3085 deps = OPERATOR_TEST_DEPS,
3086)
3087
3088xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003089 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003090 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003091 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003092 "test/resize-bilinear-operator-tester.h",
3093 ] + OPERATOR_TEST_PARAMS_HDRS,
3094 deps = OPERATOR_TEST_DEPS,
3095)
3096
3097xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003098 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003099 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003100 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003101 "test/sigmoid-operator-tester.h",
3102 ],
3103 deps = OPERATOR_TEST_DEPS,
3104)
3105
3106xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003107 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003109 "test/softmax-nc.cc",
3110 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111 ],
3112 deps = OPERATOR_TEST_DEPS,
3113)
3114
3115xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003116 name = "subtract_nd_test",
3117 srcs = [
3118 "test/binary-elementwise-operator-tester.h",
3119 "test/subtract-nd.cc",
3120 ],
3121 deps = OPERATOR_TEST_DEPS,
3122)
3123
3124xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003125 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003127 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128 "test/unpooling-operator-tester.h",
3129 ],
3130 deps = OPERATOR_TEST_DEPS,
3131)
3132
3133############################# Build configurations #############################
3134
Marat Dukhanb8642352019-10-30 15:43:02 -07003135# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003137 name = "xnn_enable_assembly_explicit_true",
3138 define_values = {"xnn_enable_assembly": "true"},
3139)
3140
3141# Disables usage of assembly kernels.
3142config_setting(
3143 name = "xnn_enable_assembly_explicit_false",
3144 define_values = {"xnn_enable_assembly": "false"},
3145)
3146
3147# Builds with -c dbg
3148config_setting(
3149 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003151 "compilation_mode": "dbg",
3152 },
3153)
3154
3155# Builds with -c opt
3156config_setting(
3157 name = "optimized_build",
3158 values = {
3159 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003160 },
3161)
3162
3163config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003164 name = "linux_k8",
3165 values = {"cpu": "k8"},
3166)
3167
3168config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003169 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003170 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003171)
3172
3173config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003174 name = "android",
3175 values = {"crosstool_top": "//external:android/crosstool"},
3176)
3177
3178config_setting(
3179 name = "android_armv7",
3180 values = {
3181 "crosstool_top": "//external:android/crosstool",
3182 "cpu": "armeabi-v7a",
3183 },
3184)
3185
3186config_setting(
3187 name = "android_arm64",
3188 values = {
3189 "crosstool_top": "//external:android/crosstool",
3190 "cpu": "arm64-v8a",
3191 },
3192)
3193
3194config_setting(
3195 name = "android_x86",
3196 values = {
3197 "crosstool_top": "//external:android/crosstool",
3198 "cpu": "x86",
3199 },
3200)
3201
3202config_setting(
3203 name = "android_x86_64",
3204 values = {
3205 "crosstool_top": "//external:android/crosstool",
3206 "cpu": "x86_64",
3207 },
3208)
3209
3210config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003211 name = "macos_x86_64",
3212 values = {
3213 "apple_platform_type": "macos",
3214 "cpu": "darwin",
3215 },
3216)
3217
3218config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003220 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003221)
3222
3223config_setting(
3224 name = "emscripten_wasm",
3225 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003226 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003227 "cpu": "wasm",
3228 },
3229)
3230
3231config_setting(
3232 name = "emscripten_wasmsimd",
3233 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003234 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003236 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237 },
3238)
3239
3240config_setting(
3241 name = "emscripten_asmjs",
3242 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003243 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003244 "cpu": "asmjs",
3245 },
3246)