blob: 619a16745c004e943a783390cb0745e98c2b047b [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 Dukhan08c4a432019-10-03 09:29:21 -0700846]
847
848SSE_UKERNELS = [
849 "src/f32-avgpool/mp9p8q-sse.c",
850 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800851 "src/f32-bilinear/gen/sse-c4.c",
852 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700853 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800854 "src/f32-dwconv-spchw/3x3p1-sse.c",
855 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800856 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
857 "src/f32-dwconv/gen/up4x25-sse.c",
858 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
859 "src/f32-dwconv/gen/up4x4-sse.c",
860 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
861 "src/f32-dwconv/gen/up4x9-sse.c",
862 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
863 "src/f32-dwconv/gen/up8x25-sse.c",
864 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
865 "src/f32-dwconv/gen/up8x4-sse.c",
866 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
867 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700868 "src/f32-gavgpool-spchw/sse-x4.c",
869 "src/f32-gavgpool/mp7p7q-sse.c",
870 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800871 "src/f32-gemm/gen/1x8-sse-dup.c",
872 "src/f32-gemm/gen/1x8-sse-load1.c",
873 "src/f32-gemm/gen/1x8s4-sse.c",
874 "src/f32-gemm/gen/4x8-sse-dup.c",
875 "src/f32-gemm/gen/4x8-sse-load1.c",
876 "src/f32-gemm/gen/4x8s4-sse.c",
877 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
878 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
879 "src/f32-gemm/gen-inc/1x8s4-sse.c",
880 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
881 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
882 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800883 "src/f32-hswish/gen/sse-x4.c",
884 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800885 "src/f32-igemm/gen/1x8-sse-dup.c",
886 "src/f32-igemm/gen/1x8-sse-load1.c",
887 "src/f32-igemm/gen/1x8s4-sse.c",
888 "src/f32-igemm/gen/4x2c4-sse.c",
889 "src/f32-igemm/gen/4x8-sse-dup.c",
890 "src/f32-igemm/gen/4x8-sse-load1.c",
891 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800892 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700893 "src/f32-pavgpool/mp9p8q-sse.c",
894 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800895 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700896 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800897 "src/f32-spmm/gen/4x1-sse.c",
898 "src/f32-spmm/gen/8x1-sse.c",
899 "src/f32-vbinary/gen/vadd-sse-x4.c",
900 "src/f32-vbinary/gen/vadd-sse-x8.c",
901 "src/f32-vbinary/gen/vaddc-sse-x4.c",
902 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800903 "src/f32-vbinary/gen/vdiv-sse-x4.c",
904 "src/f32-vbinary/gen/vdiv-sse-x8.c",
905 "src/f32-vbinary/gen/vdivc-sse-x4.c",
906 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800907 "src/f32-vbinary/gen/vmax-sse-x4.c",
908 "src/f32-vbinary/gen/vmax-sse-x8.c",
909 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
910 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
911 "src/f32-vbinary/gen/vmin-sse-x4.c",
912 "src/f32-vbinary/gen/vmin-sse-x8.c",
913 "src/f32-vbinary/gen/vminc-sse-x4.c",
914 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800915 "src/f32-vbinary/gen/vmul-sse-x4.c",
916 "src/f32-vbinary/gen/vmul-sse-x8.c",
917 "src/f32-vbinary/gen/vmulc-sse-x4.c",
918 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800919 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
920 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800921 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
922 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
923 "src/f32-vbinary/gen/vsub-sse-x4.c",
924 "src/f32-vbinary/gen/vsub-sse-x8.c",
925 "src/f32-vbinary/gen/vsubc-sse-x4.c",
926 "src/f32-vbinary/gen/vsubc-sse-x8.c",
927 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
928 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700929 "src/x32-packx/x4-sse.c",
930]
931
932SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800933 "src/f32-argmaxpool/9p8x-sse2-c4.c",
934 "src/f32-argmaxpool/4x-sse2-c4.c",
935 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800936 "src/f32-prelu/gen/sse2-2x4.c",
937 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800938 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
939 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
940 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
941 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
942 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
943 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
944 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
945 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
946 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
947 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
948 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
949 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800950 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800951 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800952 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800953 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800954 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
955 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700956 "src/q8-avgpool/mp9p8q-sse2.c",
957 "src/q8-avgpool/up9-sse2.c",
958 "src/q8-igemm/4x4c2-sse2.c",
959 "src/q8-dwconv/up8x9-sse2.c",
960 "src/q8-gavgpool/mp7p7q-sse2.c",
961 "src/q8-gavgpool/up7-sse2.c",
962 "src/q8-gemm/2x4c8-sse2.c",
963 "src/q8-gemm/4x4c2-sse2.c",
964 "src/q8-vadd/sse2.c",
965 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800966 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700967 "src/u8-rmax/sse2.c",
968 "src/x32-pad/x2-sse2.c",
969 "src/x32-zip/x2-sse2.c",
970 "src/x32-zip/x3-sse2.c",
971 "src/x32-zip/x4-sse2.c",
972 "src/x32-zip/xm-sse2.c",
973 "src/x8-zip/x2-sse2.c",
974 "src/x8-zip/x3-sse2.c",
975 "src/x8-zip/x4-sse2.c",
976 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800977 "src/math/exp-sse2-p5.c",
978 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800979 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700980]
981
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800982SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800983 "src/f32-prelu/gen/sse41-2x4.c",
984 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800985 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800986 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800987 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800988 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800989 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
990 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800991]
992
Marat Dukhan08c4a432019-10-03 09:29:21 -0700993AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800994 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800995 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
996 "src/f32-dwconv/gen/up16x4-avx.c",
997 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
998 "src/f32-dwconv/gen/up8x4-avx.c",
999 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
1000 "src/f32-dwconv/gen/up16x9-avx.c",
1001 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
1002 "src/f32-dwconv/gen/up8x9-avx.c",
1003 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
1004 "src/f32-dwconv/gen/up16x25-avx.c",
1005 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
1006 "src/f32-dwconv/gen/up8x25-avx.c",
1007 "src/f32-gemm/gen/1x8-avx-broadcast.c",
1008 "src/f32-gemm/gen/4x8-avx-broadcast.c",
1009 "src/f32-gemm/gen/5x8-avx-broadcast.c",
1010 "src/f32-gemm/gen/6x8-avx-broadcast.c",
1011 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001012 "src/f32-gemm/gen/1x16-avx-broadcast.c",
1013 "src/f32-gemm/gen/3x16-avx-broadcast.c",
1014 "src/f32-gemm/gen/4x16-avx-broadcast.c",
1015 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001016 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
1017 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
1018 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
1019 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
1020 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001021 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
1022 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
1023 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
1024 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001025 "src/f32-hswish/gen/avx-x8.c",
1026 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001027 "src/f32-igemm/gen/1x8-avx-broadcast.c",
1028 "src/f32-igemm/gen/4x8-avx-broadcast.c",
1029 "src/f32-igemm/gen/5x8-avx-broadcast.c",
1030 "src/f32-igemm/gen/6x8-avx-broadcast.c",
1031 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001032 "src/f32-igemm/gen/1x16-avx-broadcast.c",
1033 "src/f32-igemm/gen/3x16-avx-broadcast.c",
1034 "src/f32-igemm/gen/4x16-avx-broadcast.c",
1035 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001036 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001037 "src/f32-vbinary/gen/vadd-avx-x8.c",
1038 "src/f32-vbinary/gen/vadd-avx-x16.c",
1039 "src/f32-vbinary/gen/vaddc-avx-x8.c",
1040 "src/f32-vbinary/gen/vaddc-avx-x16.c",
1041 "src/f32-vbinary/gen/vdiv-avx-x8.c",
1042 "src/f32-vbinary/gen/vdiv-avx-x16.c",
1043 "src/f32-vbinary/gen/vdivc-avx-x8.c",
1044 "src/f32-vbinary/gen/vdivc-avx-x16.c",
1045 "src/f32-vbinary/gen/vmax-avx-x8.c",
1046 "src/f32-vbinary/gen/vmax-avx-x16.c",
1047 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1048 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1049 "src/f32-vbinary/gen/vmin-avx-x8.c",
1050 "src/f32-vbinary/gen/vmin-avx-x16.c",
1051 "src/f32-vbinary/gen/vminc-avx-x8.c",
1052 "src/f32-vbinary/gen/vminc-avx-x16.c",
1053 "src/f32-vbinary/gen/vmul-avx-x8.c",
1054 "src/f32-vbinary/gen/vmul-avx-x16.c",
1055 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1056 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1057 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1058 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1059 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1060 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1061 "src/f32-vbinary/gen/vsub-avx-x8.c",
1062 "src/f32-vbinary/gen/vsub-avx-x16.c",
1063 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1064 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001065 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001066]
1067
Marat Dukhanfda12b82019-11-21 12:27:59 -08001068FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001069 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
1070 "src/f32-dwconv/gen/up16x4-fma3.c",
1071 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
1072 "src/f32-dwconv/gen/up8x4-fma3.c",
1073 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
1074 "src/f32-dwconv/gen/up16x9-fma3.c",
1075 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
1076 "src/f32-dwconv/gen/up8x9-fma3.c",
1077 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
1078 "src/f32-dwconv/gen/up16x25-fma3.c",
1079 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
1080 "src/f32-dwconv/gen/up8x25-fma3.c",
1081 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
1082 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
1083 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
1084 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
1085 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
1086 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001087 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
1088 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
1089 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
1090 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001091 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
1092 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
1093 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
1094 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001095 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
1096 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
1097 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
1098 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
1099 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
1100 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001101 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
1102 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
1103 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
1104 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001105 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
1106 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
1107 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
1108 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001109 "src/f32-hswish/gen/fma3-x8.c",
1110 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001111 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
1112 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
1113 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
1114 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
1115 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
1116 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001117 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
1118 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
1119 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
1120 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001121 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
1122 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
1123 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
1124 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001125]
1126
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001127AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001128 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1129 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1130 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1131 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1132 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1133 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1134 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1135 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1136 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1137 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1138 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1139 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1140 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1141 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1142 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1143 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1144 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1145 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1146 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1147 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1148 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1149 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1150 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1151 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1152 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1153 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1154 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1155 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1156 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1157 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1158 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1159 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1160 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1161 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1162 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1163 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001164 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1165 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1166 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1167 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1168 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1169 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1170 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1171 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1172 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1173 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1174 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1175 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1176 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1177 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1178 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1179 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1180 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1181 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1182 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1185 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1186 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1187 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1188 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1189 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1190 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1191 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1192 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1193 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001194 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1195 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1196 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1197 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1198 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1199 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1200 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1201 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1202 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1203 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1204 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1205 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1206 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1207 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1208 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1209 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1210 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1211 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1212 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1213 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1214 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1215 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1216 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1217 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001218 "src/math/exp-avx2-p5.c",
1219 "src/math/exp-avx2-perm-p3.c",
1220 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001221 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001222 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001223 "src/math/sigmoid-avx2-rr2-p5-div.c",
1224 "src/math/sigmoid-avx2-rr1-p5-div.c",
1225 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1226 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1227 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1228 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001229]
1230
Marat Dukhan08c4a432019-10-03 09:29:21 -07001231AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001232 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -08001233 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
1234 "src/f32-dwconv/gen/up32x4-avx512f.c",
1235 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
1236 "src/f32-dwconv/gen/up16x4-avx512f.c",
1237 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
1238 "src/f32-dwconv/gen/up32x9-avx512f.c",
1239 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
1240 "src/f32-dwconv/gen/up16x9-avx512f.c",
1241 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
1242 "src/f32-dwconv/gen/up32x25-avx512f.c",
1243 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
1244 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001245 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
1246 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
1247 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
1248 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
1249 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
1250 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
1251 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
1252 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
1253 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
1254 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
1255 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
1256 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001257 "src/f32-hswish/gen/avx512f-x16.c",
1258 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001259 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
1260 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
1261 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
1262 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
1263 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
1264 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001265 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1266 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1267 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1268 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1269 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1270 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1271 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1272 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1273 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1274 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1275 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1276 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1277 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1278 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1279 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1280 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1281 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1282 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1283 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1284 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1285 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1286 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1287 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1288 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1289 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1290 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1291 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1292 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1293 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1294 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1295 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1296 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1297 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1298 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1299 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1300 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001301 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001302 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1303 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1304 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1305 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1306 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1307 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1308 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1309 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1310 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1311 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1312 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1313 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1314 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1315 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1316 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1317 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1318 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1319 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1320 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1321 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1322 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1323 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1324 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1325 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1326 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1327 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1328 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1329 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001330 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001331 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1332 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1333 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1334 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1335 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1336 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1337 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1338 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1339 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1340 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1341 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1342 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1343 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1344 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1345 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1346 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1347 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1348 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1349 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1350 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1351 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1352 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1353 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1354 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001355 "src/math/exp-avx512f-p5-scalef.c",
1356 "src/math/exp-avx512f-p5.c",
1357 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001358 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001359 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360]
1361
1362AARCH32_ASM_UKERNELS = [
1363 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001364 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001365 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1366 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001367 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368]
1369
1370AARCH64_ASM_UKERNELS = [
1371 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1372 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001373 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1374 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1375 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1376 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1377 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1378 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1379 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1380 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1381 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1382 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001383 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001384 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1385 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001386 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001387 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001388 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1389 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1390 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1391 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1392 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1393 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1394 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1395 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1396 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1397 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1398 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1399 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1400 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001401 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001402 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1403 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001404 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001405 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001406 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1407 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1408 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001409 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001410 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001411 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1412 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001413 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001414 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001415 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1416 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1417 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1418 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001419 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001420 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001421 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1422 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001423]
1424
1425INTERNAL_MICROKERNEL_HDRS = [
1426 "src/xnnpack/argmaxpool.h",
1427 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001428 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001429 "src/xnnpack/clamp.h",
1430 "src/xnnpack/common.h",
1431 "src/xnnpack/conv.h",
1432 "src/xnnpack/dwconv.h",
1433 "src/xnnpack/gavgpool.h",
1434 "src/xnnpack/gemm.h",
1435 "src/xnnpack/hswish.h",
1436 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001437 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001438 "src/xnnpack/lut.h",
1439 "src/xnnpack/math.h",
1440 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001441 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001442 "src/xnnpack/packx.h",
1443 "src/xnnpack/pad.h",
1444 "src/xnnpack/params.h",
1445 "src/xnnpack/pavgpool.h",
1446 "src/xnnpack/ppmm.h",
1447 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001448 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001449 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001450 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001451 "src/xnnpack/rmax.h",
1452 "src/xnnpack/scalar-utils.h",
1453 "src/xnnpack/spmm.h",
1454 "src/xnnpack/unpool.h",
1455 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001456 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001457 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001458 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001459 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001460 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001461 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462 "src/xnnpack/zip.h",
1463]
1464
1465INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1466 "include/xnnpack.h",
1467 "src/xnnpack/allocator.h",
1468 "src/xnnpack/compute.h",
1469 "src/xnnpack/im2col.h",
1470 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001471 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001472 "src/xnnpack/operator.h",
1473 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001474 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001475 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001476 "src/xnnpack/requantization.h",
1477]
1478
1479ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1480 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001481]
1482
1483MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001484 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001485 "include/xnnpack.h",
1486]
1487
1488MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1489 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001490 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001491 "src/xnnpack/requantization.h",
1492 "include/xnnpack.h",
1493]
1494
1495OPERATOR_TEST_PARAMS_HDRS = [
1496 "src/xnnpack/params.h",
1497 "src/xnnpack/common.h",
1498]
1499
1500WEIGHTS_PACK_HDRS = [
1501 "src/xnnpack/pack.h",
1502 "src/xnnpack/operator.h",
1503 "src/xnnpack/compute.h",
1504]
1505
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001506LOGGING_COPTS = select({
1507 # No logging in optimized mode
1508 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1509 # Full logging in debug mode
1510 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1511 # Error-only logging in default (fastbuild) mode
1512 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1513})
1514
1515LOGGING_HDRS = [
1516 "src/xnnpack/log.h",
1517]
1518
Marat Dukhan08c4a432019-10-03 09:29:21 -07001519xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001520 name = "tables",
1521 srcs = TABLE_SRCS,
1522 hdrs = INTERNAL_HDRS,
1523 copts = xnnpack_std_copts(),
1524)
1525
1526xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001527 name = "scalar_ukernels",
1528 srcs = SCALAR_UKERNELS,
1529 hdrs = INTERNAL_HDRS,
1530 aarch32_copts = ["-marm"],
1531 copts = xnnpack_std_copts(),
1532 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001533 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001534 "@FP16",
1535 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001536 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001537 ],
1538)
1539
1540xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001541 name = "wasm_ukernels",
1542 hdrs = INTERNAL_HDRS,
1543 copts = xnnpack_std_copts(),
1544 wasm_srcs = WASM_UKERNELS,
1545 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001546 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001547 "@FP16",
1548 "@FXdiv",
1549 "@pthreadpool",
1550 ],
1551)
1552
1553xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001554 name = "psimd_fastmath_ukernels",
1555 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556 hdrs = INTERNAL_HDRS,
1557 aarch32_copts = [
1558 "-marm",
1559 "-mfpu=neon",
1560 ],
1561 copts = xnnpack_std_copts(),
1562 optimized_copts = [
1563 "-O3",
1564 "-ffast-math",
1565 ],
1566 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001567 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001568 "@FP16",
1569 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001570 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001571 ],
1572)
1573
1574xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001575 name = "psimd_accmath_ukernels",
1576 srcs = PSIMD_ACCMATH_UKERNELS,
1577 hdrs = INTERNAL_HDRS,
1578 aarch32_copts = [
1579 "-marm",
1580 "-mfpu=neon",
1581 ],
1582 copts = xnnpack_std_copts(),
1583 optimized_copts = [
1584 "-O3",
1585 ],
1586 deps = [
1587 ":tables",
1588 "@FP16",
1589 "@psimd",
1590 "@pthreadpool",
1591 ],
1592)
1593
1594xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595 name = "neon_ukernels",
1596 hdrs = INTERNAL_HDRS,
1597 aarch32_copts = [
1598 "-marm",
1599 "-mfpu=neon",
1600 ],
1601 aarch32_srcs = NEON_UKERNELS,
1602 aarch64_srcs = NEON_UKERNELS,
1603 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001604 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001605 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001606 "@FP16",
1607 "@pthreadpool",
1608 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001609)
1610
1611xnnpack_cc_library(
1612 name = "neonfma_ukernels",
1613 hdrs = INTERNAL_HDRS,
1614 aarch32_copts = [
1615 "-marm",
1616 "-mfpu=neon-vfpv4",
1617 ],
1618 aarch32_srcs = NEONFMA_UKERNELS,
1619 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1620 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001621 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001622 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001623 "@FP16",
1624 "@pthreadpool",
1625 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001626)
1627
1628xnnpack_cc_library(
1629 name = "neonfp16arith_ukernels",
1630 hdrs = INTERNAL_HDRS,
1631 aarch64_copts = ["-march=armv8.2-a+fp16"],
1632 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1633 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001634 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001635 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001636 "@FP16",
1637 "@pthreadpool",
1638 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001639)
1640
1641xnnpack_cc_library(
1642 name = "sse2_ukernels",
1643 hdrs = INTERNAL_HDRS,
1644 copts = xnnpack_std_copts(),
1645 x86_copts = ["-msse2"],
1646 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001647 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001648 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001649 "@FP16",
1650 "@pthreadpool",
1651 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001652)
1653
1654xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001655 name = "sse41_ukernels",
1656 hdrs = INTERNAL_HDRS,
1657 copts = xnnpack_std_copts(),
1658 x86_copts = ["-msse4.1"],
1659 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001660 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001661 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001662 "@FP16",
1663 "@pthreadpool",
1664 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001665)
1666
1667xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001668 name = "avx_ukernels",
1669 hdrs = INTERNAL_HDRS,
1670 copts = xnnpack_std_copts(),
1671 x86_copts = ["-mavx"],
1672 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001673 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001674 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001675 "@FP16",
1676 "@pthreadpool",
1677 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001678)
1679
1680xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001681 name = "fma3_ukernels",
1682 hdrs = INTERNAL_HDRS,
1683 copts = xnnpack_std_copts(),
1684 x86_copts = [
1685 "-mfma",
1686 ],
1687 x86_srcs = FMA3_UKERNELS,
1688 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001689 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001690 "@FP16",
1691 "@pthreadpool",
1692 ],
1693)
1694
1695xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001696 name = "avx2_ukernels",
1697 hdrs = INTERNAL_HDRS,
1698 copts = xnnpack_std_copts(),
1699 x86_copts = [
1700 "-mfma",
1701 "-mavx2",
1702 ],
1703 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001704 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001705 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001706 "@FP16",
1707 "@pthreadpool",
1708 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001709)
1710
1711xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 name = "avx512f_ukernels",
1713 hdrs = INTERNAL_HDRS,
1714 copts = xnnpack_std_copts(),
1715 x86_copts = ["-mavx512f"],
1716 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001717 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001718 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001719 "@FP16",
1720 "@pthreadpool",
1721 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001722)
1723
1724xnnpack_cc_library(
1725 name = "asm_ukernels",
1726 hdrs = ["src/xnnpack/assembly.h"],
1727 aarch32_srcs = AARCH32_ASM_UKERNELS,
1728 aarch64_srcs = AARCH64_ASM_UKERNELS,
1729)
1730
1731xnnpack_aggregate_library(
1732 name = "ukernels",
1733 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001734 ":psimd_fastmath_ukernels",
1735 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001736 ":neon_ukernels",
1737 ":neonfma_ukernels",
1738 ":asm_ukernels",
1739 ],
1740 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001741 ":psimd_fastmath_ukernels",
1742 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001743 ":neon_ukernels",
1744 ":neonfma_ukernels",
1745 ":neonfp16arith_ukernels",
1746 ":asm_ukernels",
1747 ],
1748 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001749 wasm_deps = [
1750 ":wasm_ukernels",
1751 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001752 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001753 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001754 ":psimd_fastmath_ukernels",
1755 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001756 ],
1757 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001758 ":psimd_fastmath_ukernels",
1759 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001760 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001761 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001763 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001764 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001765 ":avx512f_ukernels",
1766 ],
1767)
1768
1769xnnpack_cc_library(
1770 name = "im2col",
1771 srcs = ["src/im2col.c"],
1772 hdrs = [
1773 "src/xnnpack/common.h",
1774 "src/xnnpack/im2col.h",
1775 ],
1776 copts = xnnpack_std_copts(),
1777)
1778
1779xnnpack_cc_library(
1780 name = "indirection",
1781 srcs = ["src/indirection.c"],
1782 hdrs = INTERNAL_HDRS,
1783 copts = xnnpack_std_copts(),
1784 deps = [
1785 "@FP16",
1786 "@FXdiv",
1787 "@pthreadpool",
1788 ],
1789)
1790
1791xnnpack_cc_library(
1792 name = "operator_run",
1793 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001794 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1795 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001796 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1797 "-Wno-vla",
1798 ],
1799 deps = [
1800 "@FP16",
1801 "@FXdiv",
1802 "@clog",
1803 "@pthreadpool",
1804 ],
1805)
1806
1807cc_library(
1808 name = "enable_assembly",
1809 defines = select({
1810 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1811 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001812 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001813 }),
1814)
1815
Marat Dukhancf056b22019-10-07 10:26:29 -07001816xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001817 name = "operators",
1818 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001819 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001821 ],
1822 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001823 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001824 "-Isrc",
1825 "-Iinclude",
1826 ] + select({
1827 ":debug_build": [],
1828 "//conditions:default": xnnpack_min_size_copts(),
1829 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001830 wasm_srcs = ["src/wasm-stubs.c"],
1831 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001832 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001834 "@FP16",
1835 "@FXdiv",
1836 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001837 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001838 ],
1839)
1840
1841cc_library(
1842 name = "XNNPACK",
1843 srcs = [
1844 "src/init.c",
1845 ],
1846 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1847 "-Isrc",
1848 "-Iinclude",
1849 ] + select({
1850 ":debug_build": [],
1851 "//conditions:default": xnnpack_min_size_copts(),
1852 }),
1853 includes = ["include"],
1854 linkstatic = True,
1855 textual_hdrs = ["include/xnnpack.h"],
1856 visibility = xnnpack_visibility(),
1857 deps = [
1858 ":enable_assembly",
1859 ":ukernels",
1860 ":operator_run",
1861 ":operators",
1862 "@clog",
1863 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001864 ] + select({
1865 ":emscripten": [],
1866 "//conditions:default": ["@cpuinfo"],
1867 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001868)
1869
1870cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001871 name = "xnnpack_operators_nhwc_f32",
1872 srcs = [
1873 "src/init.c",
1874 ],
1875 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1876 "-Isrc",
1877 "-Iinclude",
1878 ] + select({
1879 ":debug_build": [],
1880 "//conditions:default": xnnpack_min_size_copts(),
1881 }),
1882 defines = [
1883 "XNN_NO_Q8_OPERATORS",
1884 "XNN_NO_U8_OPERATORS",
1885 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001886 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001887 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001888 includes = ["include"],
1889 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001890 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001891 visibility = xnnpack_visibility(),
1892 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001893 ":enable_assembly",
1894 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001895 ":operator_run",
1896 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001897 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001898 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001899 ] + select({
1900 ":emscripten": [],
1901 "//conditions:default": ["@cpuinfo"],
1902 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001903)
1904
Marat Dukhancf056b22019-10-07 10:26:29 -07001905xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906 name = "bench_utils",
1907 srcs = ["bench/utils.cc"],
1908 hdrs = ["bench/utils.h"],
1909 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001910 deps = [
1911 "@com_google_benchmark//:benchmark",
1912 "@cpuinfo",
1913 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001914)
1915
Frank Barchard7e955972019-10-11 10:34:25 -07001916######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001917
1918xnnpack_benchmark(
1919 name = "q8_gemm_bench",
1920 srcs = [
1921 "bench/gemm.h",
1922 "bench/q8-gemm.cc",
1923 "src/xnnpack/AlignedAllocator.h",
1924 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1925 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1926 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1927)
1928
1929xnnpack_benchmark(
1930 name = "f16_gemm_bench",
1931 srcs = [
1932 "bench/f16-gemm.cc",
1933 "bench/gemm.h",
1934 "src/xnnpack/AlignedAllocator.h",
1935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1936 copts = ["-Wno-unused-function"],
1937 deps = MICROKERNEL_BENCHMARK_DEPS,
1938)
1939
1940xnnpack_benchmark(
1941 name = "f32_igemm_bench",
1942 srcs = [
1943 "bench/f32-igemm.cc",
1944 "bench/conv.h",
1945 "src/xnnpack/AlignedAllocator.h",
1946 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001947 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001948)
1949
1950xnnpack_benchmark(
1951 name = "f32_conv_hwc_bench",
1952 srcs = [
1953 "bench/f32-conv-hwc.cc",
1954 "bench/dconv.h",
1955 "src/xnnpack/AlignedAllocator.h",
1956 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1957 copts = ["-Wno-unused-function"],
1958 deps = MICROKERNEL_BENCHMARK_DEPS,
1959)
1960
1961xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001962 name = "f32_conv_hwc2spchw_bench",
1963 srcs = [
1964 "bench/f32-conv-hwc2spchw.cc",
1965 "bench/dconv.h",
1966 "src/xnnpack/AlignedAllocator.h",
1967 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1968 copts = ["-Wno-unused-function"],
1969 deps = MICROKERNEL_BENCHMARK_DEPS,
1970)
1971
1972xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973 name = "f32_dwconv_bench",
1974 srcs = [
1975 "bench/f32-dwconv.cc",
1976 "bench/dwconv.h",
1977 "src/xnnpack/AlignedAllocator.h",
1978 ] + MICROKERNEL_BENCHMARK_HDRS,
1979 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1980)
1981
1982xnnpack_benchmark(
1983 name = "f32_dwconv_spchw_bench",
1984 srcs = [
1985 "bench/f32-dwconv-spchw.cc",
1986 "bench/dwconv.h",
1987 "src/xnnpack/AlignedAllocator.h",
1988 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1989 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1990)
1991
1992xnnpack_benchmark(
1993 name = "f32_gemm_bench",
1994 srcs = [
1995 "bench/f32-gemm.cc",
1996 "bench/gemm.h",
1997 "src/xnnpack/AlignedAllocator.h",
1998 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1999 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002000 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001)
2002
2003xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002004 name = "f32_raddexpminusmax_bench",
2005 srcs = [
2006 "bench/f32-raddexpminusmax.cc",
2007 "src/xnnpack/AlignedAllocator.h",
2008 ] + MICROKERNEL_BENCHMARK_HDRS,
2009 deps = MICROKERNEL_BENCHMARK_DEPS,
2010)
2011
2012xnnpack_benchmark(
2013 name = "f32_raddextexp_bench",
2014 srcs = [
2015 "bench/f32-raddextexp.cc",
2016 "src/xnnpack/AlignedAllocator.h",
2017 ] + MICROKERNEL_BENCHMARK_HDRS,
2018 deps = MICROKERNEL_BENCHMARK_DEPS,
2019)
2020
2021xnnpack_benchmark(
2022 name = "f32_raddstoreexpminusmax_bench",
2023 srcs = [
2024 "bench/f32-raddstoreexpminusmax.cc",
2025 "src/xnnpack/AlignedAllocator.h",
2026 ] + MICROKERNEL_BENCHMARK_HDRS,
2027 deps = MICROKERNEL_BENCHMARK_DEPS,
2028)
2029
2030xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031 name = "f32_rmax_bench",
2032 srcs = [
2033 "bench/f32-rmax.cc",
2034 "src/xnnpack/AlignedAllocator.h",
2035 ] + MICROKERNEL_BENCHMARK_HDRS,
2036 deps = MICROKERNEL_BENCHMARK_DEPS,
2037)
2038
2039xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002040 name = "f32_sigmoid_bench",
2041 srcs = [
2042 "bench/f32-sigmoid.cc",
2043 "src/xnnpack/AlignedAllocator.h",
2044 ] + MICROKERNEL_BENCHMARK_HDRS,
2045 copts = ["-Wno-unused-function"],
2046 deps = MICROKERNEL_BENCHMARK_DEPS,
2047)
2048
2049xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002050 name = "f32_spmm_bench",
2051 srcs = [
2052 "bench/f32-spmm.cc",
2053 "bench/gemm.h",
2054 "src/xnnpack/AlignedAllocator.h",
2055 ] + MICROKERNEL_BENCHMARK_HDRS,
2056 copts = ["-Wno-unused-function"],
2057 deps = MICROKERNEL_BENCHMARK_DEPS,
2058)
2059
2060xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002061 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002062 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002063 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002064 ] + MICROKERNEL_BENCHMARK_HDRS,
2065 copts = ["-Wno-unused-function"],
2066 deps = MICROKERNEL_BENCHMARK_DEPS,
2067)
2068
2069xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002070 name = "f32_vscaleexpminusmax_bench",
2071 srcs = [
2072 "bench/f32-vscaleexpminusmax.cc",
2073 "src/xnnpack/AlignedAllocator.h",
2074 ] + MICROKERNEL_BENCHMARK_HDRS,
2075 deps = MICROKERNEL_BENCHMARK_DEPS,
2076)
2077
2078xnnpack_benchmark(
2079 name = "f32_vscaleextexp_bench",
2080 srcs = [
2081 "bench/f32-vscaleextexp.cc",
2082 "src/xnnpack/AlignedAllocator.h",
2083 ] + MICROKERNEL_BENCHMARK_HDRS,
2084 deps = MICROKERNEL_BENCHMARK_DEPS,
2085)
2086
2087xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088 name = "f32_im2col_gemm_bench",
2089 srcs = [
2090 "bench/f32-im2col-gemm.cc",
2091 "bench/conv.h",
2092 "src/xnnpack/AlignedAllocator.h",
2093 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2094 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2095)
2096
2097########################### Benchmarks for operators ###########################
2098
2099xnnpack_benchmark(
2100 name = "add_bench",
2101 srcs = ["bench/add.cc"],
2102 deps = OPERATOR_BENCHMARK_DEPS,
2103)
2104
2105xnnpack_benchmark(
2106 name = "average_pooling_bench",
2107 srcs = ["bench/average-pooling.cc"],
2108 deps = OPERATOR_BENCHMARK_DEPS,
2109)
2110
2111xnnpack_benchmark(
2112 name = "channel_shuffle_bench",
2113 srcs = ["bench/channel-shuffle.cc"],
2114 deps = OPERATOR_BENCHMARK_DEPS,
2115)
2116
2117xnnpack_benchmark(
2118 name = "convolution_bench",
2119 srcs = ["bench/convolution.cc"],
2120 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2121 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2122)
2123
2124xnnpack_benchmark(
2125 name = "deconvolution_bench",
2126 srcs = ["bench/deconvolution.cc"],
2127 copts = xnnpack_optional_tflite_copts(),
2128 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2129)
2130
2131xnnpack_benchmark(
2132 name = "global_average_pooling_bench",
2133 srcs = ["bench/global-average-pooling.cc"],
2134 deps = OPERATOR_BENCHMARK_DEPS,
2135)
2136
2137xnnpack_benchmark(
2138 name = "max_pooling_bench",
2139 srcs = ["bench/max-pooling.cc"],
2140 deps = OPERATOR_BENCHMARK_DEPS,
2141)
2142
2143xnnpack_benchmark(
2144 name = "sigmoid_bench",
2145 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002146 copts = xnnpack_optional_tflite_copts(),
2147 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002148)
2149
2150xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002151 name = "prelu_bench",
2152 srcs = ["bench/prelu.cc"],
2153 copts = xnnpack_optional_tflite_copts(),
2154 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2155)
2156
2157xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002158 name = "softmax_bench",
2159 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002160 copts = xnnpack_optional_tflite_copts(),
2161 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002162)
2163
Marat Dukhanc068bb62019-10-04 13:24:39 -07002164############################# End-to-end benchmarks ############################
2165
2166cc_library(
2167 name = "mobilenet_v1",
2168 srcs = ["models/mobilenet-v1.cc"],
2169 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002170 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002171 linkstatic = True,
2172 deps = [
2173 ":XNNPACK",
2174 "@pthreadpool",
2175 ],
2176)
2177
2178cc_library(
2179 name = "mobilenet_v2",
2180 srcs = ["models/mobilenet-v2.cc"],
2181 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002182 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002183 linkstatic = True,
2184 deps = [
2185 ":XNNPACK",
2186 "@pthreadpool",
2187 ],
2188)
2189
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002190cc_library(
2191 name = "mobilenet_v3_large",
2192 srcs = ["models/mobilenet-v3-large.cc"],
2193 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002194 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002195 linkstatic = True,
2196 deps = [
2197 ":XNNPACK",
2198 "@pthreadpool",
2199 ],
2200)
2201
2202cc_library(
2203 name = "mobilenet_v3_small",
2204 srcs = ["models/mobilenet-v3-small.cc"],
2205 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002206 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002207 linkstatic = True,
2208 deps = [
2209 ":XNNPACK",
2210 "@pthreadpool",
2211 ],
2212)
2213
Marat Dukhanc068bb62019-10-04 13:24:39 -07002214xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002215 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002216 srcs = [
2217 "bench/f32-dwconv-e2e.cc",
2218 "bench/end2end.h",
2219 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002220 copts = ["-Wno-unused-function"],
2221 deps = MICROKERNEL_BENCHMARK_DEPS + [
2222 ":XNNPACK",
2223 ":mobilenet_v1",
2224 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002225 ":mobilenet_v3_large",
2226 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002227 ],
2228)
2229
2230xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002231 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002232 srcs = [
2233 "bench/f32-gemm-e2e.cc",
2234 "bench/end2end.h",
2235 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002236 copts = ["-Wno-unused-function"],
2237 deps = MICROKERNEL_BENCHMARK_DEPS + [
2238 ":XNNPACK",
2239 ":mobilenet_v1",
2240 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002241 ":mobilenet_v3_large",
2242 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002243 ],
2244)
2245
2246xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002247 name = "end2end_bench",
2248 srcs = ["bench/end2end.cc"],
2249 deps = [
2250 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002251 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002252 ":mobilenet_v1",
2253 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002254 ":mobilenet_v3_large",
2255 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002256 "@pthreadpool",
2257 ],
2258)
2259
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002260#################### Accuracy evaluation for math functions ####################
2261
2262xnnpack_benchmark(
2263 name = "f32_exp_eval",
2264 srcs = [
2265 "eval/f32-exp.cc",
2266 "src/xnnpack/AlignedAllocator.h",
2267 ] + ACCURACY_EVAL_HDRS,
2268 deps = ACCURACY_EVAL_DEPS,
2269)
2270
Marat Dukhan515c9772019-10-17 18:07:57 -07002271xnnpack_benchmark(
2272 name = "f32_expminus_eval",
2273 srcs = [
2274 "eval/f32-expminus.cc",
2275 "src/xnnpack/AlignedAllocator.h",
2276 ] + ACCURACY_EVAL_HDRS,
2277 deps = ACCURACY_EVAL_DEPS,
2278)
2279
Marat Dukhan98ba4412019-10-23 02:14:28 -07002280xnnpack_benchmark(
2281 name = "f32_extexp_eval",
2282 srcs = [
2283 "eval/f32-extexp.cc",
2284 "src/xnnpack/AlignedAllocator.h",
2285 ] + ACCURACY_EVAL_HDRS,
2286 deps = ACCURACY_EVAL_DEPS,
2287)
2288
Marat Dukhan346a9e52019-11-15 09:06:30 -08002289xnnpack_benchmark(
2290 name = "f32_sigmoid_eval",
2291 srcs = [
2292 "eval/f32-sigmoid.cc",
2293 "src/xnnpack/AlignedAllocator.h",
2294 ] + ACCURACY_EVAL_HDRS,
2295 deps = ACCURACY_EVAL_DEPS,
2296)
2297
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298######################### Unit tests for micro-kernels #########################
2299
2300xnnpack_unit_test(
2301 name = "f16_gemm_test",
2302 srcs = [
2303 "test/f16-gemm.cc",
2304 "test/gemm-microkernel-tester.h",
2305 "src/xnnpack/AlignedAllocator.h",
2306 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2307 deps = MICROKERNEL_TEST_DEPS,
2308)
2309
2310xnnpack_unit_test(
2311 name = "f32_argmaxpool_test",
2312 srcs = [
2313 "test/f32-argmaxpool.cc",
2314 "test/argmaxpool-microkernel-tester.h",
2315 "src/xnnpack/AlignedAllocator.h",
2316 ] + MICROKERNEL_TEST_HDRS,
2317 deps = MICROKERNEL_TEST_DEPS,
2318)
2319
2320xnnpack_unit_test(
2321 name = "f32_avgpool_test",
2322 srcs = [
2323 "test/f32-avgpool.cc",
2324 "test/avgpool-microkernel-tester.h",
2325 "src/xnnpack/AlignedAllocator.h",
2326 ] + MICROKERNEL_TEST_HDRS,
2327 deps = MICROKERNEL_TEST_DEPS,
2328)
2329
2330xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002331 name = "f32_bilinear_test",
2332 srcs = [
2333 "test/f32-bilinear.cc",
2334 "test/bilinear-microkernel-tester.h",
2335 "src/xnnpack/AlignedAllocator.h",
2336 ] + MICROKERNEL_TEST_HDRS,
2337 deps = MICROKERNEL_TEST_DEPS,
2338)
2339
2340xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 name = "f32_clamp_test",
2342 srcs = [
2343 "test/f32-clamp.cc",
2344 "test/clamp-microkernel-tester.h",
2345 ] + MICROKERNEL_TEST_HDRS,
2346 deps = MICROKERNEL_TEST_DEPS,
2347)
2348
2349xnnpack_unit_test(
2350 name = "f32_igemm_test",
2351 srcs = [
2352 "test/f32-igemm.cc",
2353 "test/gemm-microkernel-tester.h",
2354 "src/xnnpack/AlignedAllocator.h",
2355 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2356 deps = MICROKERNEL_TEST_DEPS,
2357)
2358
2359xnnpack_unit_test(
2360 name = "f32_conv_hwc_test",
2361 srcs = [
2362 "test/f32-conv-hwc.cc",
2363 "test/conv-hwc-microkernel-tester.h",
2364 "src/xnnpack/AlignedAllocator.h",
2365 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2366 deps = MICROKERNEL_TEST_DEPS,
2367)
2368
2369xnnpack_unit_test(
2370 name = "f32_conv_hwc2spchw_test",
2371 srcs = [
2372 "test/f32-conv-hwc2spchw.cc",
2373 "test/conv-hwc2spchw-microkernel-tester.h",
2374 "src/xnnpack/AlignedAllocator.h",
2375 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2376 deps = MICROKERNEL_TEST_DEPS,
2377)
2378
2379xnnpack_unit_test(
2380 name = "f32_dwconv_test",
2381 srcs = [
2382 "test/f32-dwconv.cc",
2383 "test/dwconv-microkernel-tester.h",
2384 "src/xnnpack/AlignedAllocator.h",
2385 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2386 deps = MICROKERNEL_TEST_DEPS,
2387)
2388
2389xnnpack_unit_test(
2390 name = "f32_dwconv_spchw_test",
2391 srcs = [
2392 "test/f32-dwconv-spchw.cc",
2393 "test/dwconv-spchw-microkernel-tester.h",
2394 "src/xnnpack/AlignedAllocator.h",
2395 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2396 deps = MICROKERNEL_TEST_DEPS,
2397)
2398
2399xnnpack_unit_test(
2400 name = "f32_gavgpool_test",
2401 srcs = [
2402 "test/f32-gavgpool.cc",
2403 "test/gavgpool-microkernel-tester.h",
2404 "src/xnnpack/AlignedAllocator.h",
2405 ] + MICROKERNEL_TEST_HDRS,
2406 deps = MICROKERNEL_TEST_DEPS,
2407)
2408
2409xnnpack_unit_test(
2410 name = "f32_gavgpool_spchw_test",
2411 srcs = [
2412 "test/f32-gavgpool-spchw.cc",
2413 "test/gavgpool-spchw-microkernel-tester.h",
2414 "src/xnnpack/AlignedAllocator.h",
2415 ] + MICROKERNEL_TEST_HDRS,
2416 deps = MICROKERNEL_TEST_DEPS,
2417)
2418
2419xnnpack_unit_test(
2420 name = "f32_gemm_test",
2421 srcs = [
2422 "test/f32-gemm.cc",
2423 "test/gemm-microkernel-tester.h",
2424 "src/xnnpack/AlignedAllocator.h",
2425 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2426 deps = MICROKERNEL_TEST_DEPS,
2427)
2428
2429xnnpack_unit_test(
2430 name = "f32_gemminc_test",
2431 srcs = [
2432 "test/f32-gemminc.cc",
2433 "test/gemm-microkernel-tester.h",
2434 "src/xnnpack/AlignedAllocator.h",
2435 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2436 deps = MICROKERNEL_TEST_DEPS,
2437)
2438
2439xnnpack_unit_test(
2440 name = "f32_hswish_test",
2441 srcs = [
2442 "test/f32-hswish.cc",
2443 "test/hswish-microkernel-tester.h",
2444 ] + MICROKERNEL_TEST_HDRS,
2445 deps = MICROKERNEL_TEST_DEPS,
2446)
2447
2448xnnpack_unit_test(
2449 name = "f32_maxpool_test",
2450 srcs = [
2451 "test/f32-maxpool.cc",
2452 "test/maxpool-microkernel-tester.h",
2453 ] + MICROKERNEL_TEST_HDRS,
2454 deps = MICROKERNEL_TEST_DEPS,
2455)
2456
2457xnnpack_unit_test(
2458 name = "f32_pavgpool_test",
2459 srcs = [
2460 "test/f32-pavgpool.cc",
2461 "test/avgpool-microkernel-tester.h",
2462 "src/xnnpack/AlignedAllocator.h",
2463 ] + MICROKERNEL_TEST_HDRS,
2464 deps = MICROKERNEL_TEST_DEPS,
2465)
2466
2467xnnpack_unit_test(
2468 name = "f32_ppmm_test",
2469 srcs = [
2470 "test/f32-ppmm.cc",
2471 "test/gemm-microkernel-tester.h",
2472 "src/xnnpack/AlignedAllocator.h",
2473 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2474 deps = MICROKERNEL_TEST_DEPS,
2475)
2476
2477xnnpack_unit_test(
2478 name = "f32_prelu_test",
2479 srcs = [
2480 "test/f32-prelu.cc",
2481 "test/prelu-microkernel-tester.h",
2482 "src/xnnpack/AlignedAllocator.h",
2483 ] + MICROKERNEL_TEST_HDRS,
2484 deps = MICROKERNEL_TEST_DEPS,
2485)
2486
2487xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002488 name = "f32_raddexpminusmax_test",
2489 srcs = [
2490 "test/f32-raddexpminusmax.cc",
2491 "test/raddexpminusmax-microkernel-tester.h",
2492 ] + MICROKERNEL_TEST_HDRS,
2493 deps = MICROKERNEL_TEST_DEPS,
2494)
2495
2496xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002497 name = "f32_raddextexp_test",
2498 srcs = [
2499 "test/f32-raddextexp.cc",
2500 "test/raddextexp-microkernel-tester.h",
2501 ] + MICROKERNEL_TEST_HDRS,
2502 deps = MICROKERNEL_TEST_DEPS,
2503)
2504
2505xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002506 name = "f32_raddstoreexpminusmax_test",
2507 srcs = [
2508 "test/f32-raddstoreexpminusmax.cc",
2509 "test/raddstoreexpminusmax-microkernel-tester.h",
2510 ] + MICROKERNEL_TEST_HDRS,
2511 deps = MICROKERNEL_TEST_DEPS,
2512)
2513
2514xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002515 name = "f32_rmax_test",
2516 srcs = [
2517 "test/f32-rmax.cc",
2518 "test/rmax-microkernel-tester.h",
2519 ] + MICROKERNEL_TEST_HDRS,
2520 deps = MICROKERNEL_TEST_DEPS,
2521)
2522
2523xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002524 name = "f32_sigmoid_test",
2525 srcs = [
2526 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002527 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002528 ] + MICROKERNEL_TEST_HDRS,
2529 deps = MICROKERNEL_TEST_DEPS,
2530)
2531
2532xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002533 name = "f32_spmm_test",
2534 srcs = [
2535 "test/f32-spmm.cc",
2536 "test/spmm-microkernel-tester.h",
2537 "src/xnnpack/AlignedAllocator.h",
2538 ] + MICROKERNEL_TEST_HDRS,
2539 deps = MICROKERNEL_TEST_DEPS,
2540)
2541
2542xnnpack_unit_test(
2543 name = "f32_vadd_test",
2544 srcs = [
2545 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002546 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002547 ] + MICROKERNEL_TEST_HDRS,
2548 deps = MICROKERNEL_TEST_DEPS,
2549)
2550
2551xnnpack_unit_test(
2552 name = "f32_vaddc_test",
2553 srcs = [
2554 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002555 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002556 ] + MICROKERNEL_TEST_HDRS,
2557 deps = MICROKERNEL_TEST_DEPS,
2558)
2559
2560xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002561 name = "f32_vdiv_test",
2562 srcs = [
2563 "test/f32-vdiv.cc",
2564 "test/vbinary-microkernel-tester.h",
2565 ] + MICROKERNEL_TEST_HDRS,
2566 deps = MICROKERNEL_TEST_DEPS,
2567)
2568
2569xnnpack_unit_test(
2570 name = "f32_vdivc_test",
2571 srcs = [
2572 "test/f32-vdivc.cc",
2573 "test/vbinaryc-microkernel-tester.h",
2574 ] + MICROKERNEL_TEST_HDRS,
2575 deps = MICROKERNEL_TEST_DEPS,
2576)
2577
2578xnnpack_unit_test(
2579 name = "f32_vrdivc_test",
2580 srcs = [
2581 "test/f32-vrdivc.cc",
2582 "test/vbinaryc-microkernel-tester.h",
2583 ] + MICROKERNEL_TEST_HDRS,
2584 deps = MICROKERNEL_TEST_DEPS,
2585)
2586
2587xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002588 name = "f32_vmax_test",
2589 srcs = [
2590 "test/f32-vmax.cc",
2591 "test/vbinary-microkernel-tester.h",
2592 ] + MICROKERNEL_TEST_HDRS,
2593 deps = MICROKERNEL_TEST_DEPS,
2594)
2595
2596xnnpack_unit_test(
2597 name = "f32_vmaxc_test",
2598 srcs = [
2599 "test/f32-vmaxc.cc",
2600 "test/vbinaryc-microkernel-tester.h",
2601 ] + MICROKERNEL_TEST_HDRS,
2602 deps = MICROKERNEL_TEST_DEPS,
2603)
2604
2605xnnpack_unit_test(
2606 name = "f32_vmin_test",
2607 srcs = [
2608 "test/f32-vmin.cc",
2609 "test/vbinary-microkernel-tester.h",
2610 ] + MICROKERNEL_TEST_HDRS,
2611 deps = MICROKERNEL_TEST_DEPS,
2612)
2613
2614xnnpack_unit_test(
2615 name = "f32_vminc_test",
2616 srcs = [
2617 "test/f32-vminc.cc",
2618 "test/vbinaryc-microkernel-tester.h",
2619 ] + MICROKERNEL_TEST_HDRS,
2620 deps = MICROKERNEL_TEST_DEPS,
2621)
2622
2623xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002624 name = "f32_vmul_test",
2625 srcs = [
2626 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002627 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002628 ] + MICROKERNEL_TEST_HDRS,
2629 deps = MICROKERNEL_TEST_DEPS,
2630)
2631
2632xnnpack_unit_test(
2633 name = "f32_vmulc_test",
2634 srcs = [
2635 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002636 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002637 ] + MICROKERNEL_TEST_HDRS,
2638 deps = MICROKERNEL_TEST_DEPS,
2639)
2640
2641xnnpack_unit_test(
2642 name = "f32_vmulcaddc_test",
2643 srcs = [
2644 "test/f32-vmulcaddc.cc",
2645 "test/vmulcaddc-microkernel-tester.h",
2646 "src/xnnpack/AlignedAllocator.h",
2647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2648 deps = MICROKERNEL_TEST_DEPS,
2649)
2650
2651xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002652 name = "f32_vscale_test",
2653 srcs = [
2654 "test/f32-vscale.cc",
2655 "test/vscale-microkernel-tester.h",
2656 ] + MICROKERNEL_TEST_HDRS,
2657 deps = MICROKERNEL_TEST_DEPS,
2658)
2659
2660xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002661 name = "f32_vscaleexpminusmax_test",
2662 srcs = [
2663 "test/f32-vscaleexpminusmax.cc",
2664 "test/vscaleexpminusmax-microkernel-tester.h",
2665 ] + MICROKERNEL_TEST_HDRS,
2666 deps = MICROKERNEL_TEST_DEPS,
2667)
2668
2669xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002670 name = "f32_vscaleextexp_test",
2671 srcs = [
2672 "test/f32-vscaleextexp.cc",
2673 "test/vscaleextexp-microkernel-tester.h",
2674 ] + MICROKERNEL_TEST_HDRS,
2675 deps = MICROKERNEL_TEST_DEPS,
2676)
2677
2678xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002679 name = "f32_vsub_test",
2680 srcs = [
2681 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002682 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002683 ] + MICROKERNEL_TEST_HDRS,
2684 deps = MICROKERNEL_TEST_DEPS,
2685)
2686
2687xnnpack_unit_test(
2688 name = "f32_vsubc_test",
2689 srcs = [
2690 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002691 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002692 ] + MICROKERNEL_TEST_HDRS,
2693 deps = MICROKERNEL_TEST_DEPS,
2694)
2695
2696xnnpack_unit_test(
2697 name = "f32_vrsubc_test",
2698 srcs = [
2699 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002700 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002701 ] + MICROKERNEL_TEST_HDRS,
2702 deps = MICROKERNEL_TEST_DEPS,
2703)
2704
2705xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706 name = "q8_avgpool_test",
2707 srcs = [
2708 "test/q8-avgpool.cc",
2709 "test/avgpool-microkernel-tester.h",
2710 "src/xnnpack/AlignedAllocator.h",
2711 ] + MICROKERNEL_TEST_HDRS,
2712 deps = MICROKERNEL_TEST_DEPS,
2713)
2714
2715xnnpack_unit_test(
2716 name = "q8_igemm_test",
2717 srcs = [
2718 "test/q8-igemm.cc",
2719 "test/gemm-microkernel-tester.h",
2720 "src/xnnpack/AlignedAllocator.h",
2721 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2722 deps = MICROKERNEL_TEST_DEPS,
2723)
2724
2725xnnpack_unit_test(
2726 name = "q8_dwconv_test",
2727 srcs = [
2728 "test/q8-dwconv.cc",
2729 "test/dwconv-microkernel-tester.h",
2730 "src/xnnpack/AlignedAllocator.h",
2731 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2732 deps = MICROKERNEL_TEST_DEPS,
2733)
2734
2735xnnpack_unit_test(
2736 name = "q8_gavgpool_test",
2737 srcs = [
2738 "test/q8-gavgpool.cc",
2739 "test/gavgpool-microkernel-tester.h",
2740 "src/xnnpack/AlignedAllocator.h",
2741 ] + MICROKERNEL_TEST_HDRS,
2742 deps = MICROKERNEL_TEST_DEPS,
2743)
2744
2745xnnpack_unit_test(
2746 name = "q8_gemm_test",
2747 srcs = [
2748 "test/q8-gemm.cc",
2749 "test/gemm-microkernel-tester.h",
2750 "src/xnnpack/AlignedAllocator.h",
2751 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2752 deps = MICROKERNEL_TEST_DEPS,
2753)
2754
2755xnnpack_unit_test(
2756 name = "q8_vadd_test",
2757 srcs = [
2758 "test/q8-vadd.cc",
2759 "test/vadd-microkernel-tester.h",
2760 ] + MICROKERNEL_TEST_HDRS,
2761 deps = MICROKERNEL_TEST_DEPS,
2762)
2763
2764xnnpack_unit_test(
2765 name = "u8_clamp_test",
2766 srcs = [
2767 "test/u8-clamp.cc",
2768 "test/clamp-microkernel-tester.h",
2769 ] + MICROKERNEL_TEST_HDRS,
2770 deps = MICROKERNEL_TEST_DEPS,
2771)
2772
2773xnnpack_unit_test(
2774 name = "u8_lut32norm_test",
2775 srcs = [
2776 "test/u8-lut32norm.cc",
2777 "test/lut-norm-microkernel-tester.h",
2778 ] + MICROKERNEL_TEST_HDRS,
2779 deps = MICROKERNEL_TEST_DEPS,
2780)
2781
2782xnnpack_unit_test(
2783 name = "u8_maxpool_test",
2784 srcs = [
2785 "test/u8-maxpool.cc",
2786 "test/maxpool-microkernel-tester.h",
2787 ] + MICROKERNEL_TEST_HDRS,
2788 deps = MICROKERNEL_TEST_DEPS,
2789)
2790
2791xnnpack_unit_test(
2792 name = "u8_rmax_test",
2793 srcs = [
2794 "test/u8-rmax.cc",
2795 "test/rmax-microkernel-tester.h",
2796 ] + MICROKERNEL_TEST_HDRS,
2797 deps = MICROKERNEL_TEST_DEPS,
2798)
2799
2800xnnpack_unit_test(
2801 name = "x32_packx_test",
2802 srcs = [
2803 "test/x32-packx.cc",
2804 "test/pack-microkernel-tester.h",
2805 "src/xnnpack/AlignedAllocator.h",
2806 ] + MICROKERNEL_TEST_HDRS,
2807 deps = MICROKERNEL_TEST_DEPS,
2808)
2809
2810xnnpack_unit_test(
2811 name = "x32_pad_test",
2812 srcs = [
2813 "test/x32-pad.cc",
2814 "test/pad-microkernel-tester.h",
2815 ] + MICROKERNEL_TEST_HDRS,
2816 deps = MICROKERNEL_TEST_DEPS,
2817)
2818
2819xnnpack_unit_test(
2820 name = "x32_unpool_test",
2821 srcs = [
2822 "test/x32-unpool.cc",
2823 "test/unpool-microkernel-tester.h",
2824 ] + MICROKERNEL_TEST_HDRS,
2825 deps = MICROKERNEL_TEST_DEPS,
2826)
2827
2828xnnpack_unit_test(
2829 name = "x32_zip_test",
2830 srcs = [
2831 "test/x32-zip.cc",
2832 "test/zip-microkernel-tester.h",
2833 ] + MICROKERNEL_TEST_HDRS,
2834 deps = MICROKERNEL_TEST_DEPS,
2835)
2836
2837xnnpack_unit_test(
2838 name = "x8_lut_test",
2839 srcs = [
2840 "test/x8-lut.cc",
2841 "test/lut-microkernel-tester.h",
2842 ] + MICROKERNEL_TEST_HDRS,
2843 deps = MICROKERNEL_TEST_DEPS,
2844)
2845
2846xnnpack_unit_test(
2847 name = "x8_zip_test",
2848 srcs = [
2849 "test/x8-zip.cc",
2850 "test/zip-microkernel-tester.h",
2851 ] + MICROKERNEL_TEST_HDRS,
2852 deps = MICROKERNEL_TEST_DEPS,
2853)
2854
2855########################### Size test for the library ##########################
2856
2857xnnpack_binary(
2858 name = "size_test",
2859 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002860 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002861)
2862
2863########################### Unit tests for operators ###########################
2864
2865xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002866 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002868 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002869 "test/add-operator-tester.h",
2870 ],
2871 deps = OPERATOR_TEST_DEPS,
2872)
2873
2874xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002875 name = "add_nd_test",
2876 srcs = [
2877 "test/add-nd.cc",
2878 "test/binary-elementwise-operator-tester.h",
2879 ],
2880 deps = OPERATOR_TEST_DEPS,
2881)
2882
2883xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002884 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002886 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887 "test/argmax-pooling-operator-tester.h",
2888 ] + OPERATOR_TEST_PARAMS_HDRS,
2889 deps = OPERATOR_TEST_DEPS,
2890)
2891
2892xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002893 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002894 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002895 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002896 "test/average-pooling-operator-tester.h",
2897 ] + OPERATOR_TEST_PARAMS_HDRS,
2898 deps = OPERATOR_TEST_DEPS,
2899)
2900
2901xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002902 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002904 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905 "test/channel-pad-operator-tester.h",
2906 ] + OPERATOR_TEST_PARAMS_HDRS,
2907 deps = OPERATOR_TEST_DEPS,
2908)
2909
2910xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002911 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002912 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002913 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002914 "test/channel-shuffle-operator-tester.h",
2915 ],
2916 deps = OPERATOR_TEST_DEPS,
2917)
2918
2919xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002920 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002922 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002923 "test/clamp-operator-tester.h",
2924 ],
2925 deps = OPERATOR_TEST_DEPS,
2926)
2927
2928xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002929 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002931 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002932 "test/convolution-operator-tester.h",
2933 ],
2934 deps = OPERATOR_TEST_DEPS,
2935)
2936
2937xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002938 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002939 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002940 "test/convolution-nchw.cc",
2941 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 ],
2943 deps = OPERATOR_TEST_DEPS,
2944)
2945
2946xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002947 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002949 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002950 "test/deconvolution-operator-tester.h",
2951 ] + OPERATOR_TEST_PARAMS_HDRS,
2952 deps = OPERATOR_TEST_DEPS,
2953)
2954
2955xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002956 name = "divide_nd_test",
2957 srcs = [
2958 "test/binary-elementwise-operator-tester.h",
2959 "test/divide-nd.cc",
2960 ],
2961 deps = OPERATOR_TEST_DEPS,
2962)
2963
2964xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002965 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002967 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002968 "test/fully-connected-operator-tester.h",
2969 ],
2970 deps = OPERATOR_TEST_DEPS,
2971)
2972
2973xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002974 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002975 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002976 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 "test/global-average-pooling-operator-tester.h",
2978 ] + OPERATOR_TEST_PARAMS_HDRS,
2979 deps = OPERATOR_TEST_DEPS,
2980)
2981
2982xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002983 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002984 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002985 "test/global-average-pooling-ncw.cc",
2986 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 ],
2988 deps = OPERATOR_TEST_DEPS,
2989)
2990
2991xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002992 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002994 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995 "test/hardswish-operator-tester.h",
2996 ],
2997 deps = OPERATOR_TEST_DEPS,
2998)
2999
3000xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003001 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003003 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004 "test/leaky-relu-operator-tester.h",
3005 ],
3006 deps = OPERATOR_TEST_DEPS,
3007)
3008
3009xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003010 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003012 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003013 "test/max-pooling-operator-tester.h",
3014 ] + OPERATOR_TEST_PARAMS_HDRS,
3015 deps = OPERATOR_TEST_DEPS,
3016)
3017
3018xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003019 name = "maximum_nd_test",
3020 srcs = [
3021 "test/binary-elementwise-operator-tester.h",
3022 "test/maximum-nd.cc",
3023 ],
3024 deps = OPERATOR_TEST_DEPS,
3025)
3026
3027xnnpack_unit_test(
3028 name = "minimum_nd_test",
3029 srcs = [
3030 "test/binary-elementwise-operator-tester.h",
3031 "test/minimum-nd.cc",
3032 ],
3033 deps = OPERATOR_TEST_DEPS,
3034)
3035
3036xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003037 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003038 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003039 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003040 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003041 ],
3042 deps = OPERATOR_TEST_DEPS,
3043)
3044
3045xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003046 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003047 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003048 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049 "test/prelu-operator-tester.h",
3050 ] + OPERATOR_TEST_PARAMS_HDRS,
3051 deps = OPERATOR_TEST_DEPS,
3052)
3053
3054xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003055 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003056 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003057 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003058 "test/resize-bilinear-operator-tester.h",
3059 ] + OPERATOR_TEST_PARAMS_HDRS,
3060 deps = OPERATOR_TEST_DEPS,
3061)
3062
3063xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003064 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003065 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003066 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067 "test/sigmoid-operator-tester.h",
3068 ],
3069 deps = OPERATOR_TEST_DEPS,
3070)
3071
3072xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003073 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003075 "test/softmax-nc.cc",
3076 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077 ],
3078 deps = OPERATOR_TEST_DEPS,
3079)
3080
3081xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003082 name = "subtract_nd_test",
3083 srcs = [
3084 "test/binary-elementwise-operator-tester.h",
3085 "test/subtract-nd.cc",
3086 ],
3087 deps = OPERATOR_TEST_DEPS,
3088)
3089
3090xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003091 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003092 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003093 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 "test/unpooling-operator-tester.h",
3095 ],
3096 deps = OPERATOR_TEST_DEPS,
3097)
3098
3099############################# Build configurations #############################
3100
Marat Dukhanb8642352019-10-30 15:43:02 -07003101# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003102config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003103 name = "xnn_enable_assembly_explicit_true",
3104 define_values = {"xnn_enable_assembly": "true"},
3105)
3106
3107# Disables usage of assembly kernels.
3108config_setting(
3109 name = "xnn_enable_assembly_explicit_false",
3110 define_values = {"xnn_enable_assembly": "false"},
3111)
3112
3113# Builds with -c dbg
3114config_setting(
3115 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003116 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003117 "compilation_mode": "dbg",
3118 },
3119)
3120
3121# Builds with -c opt
3122config_setting(
3123 name = "optimized_build",
3124 values = {
3125 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 },
3127)
3128
3129config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003130 name = "linux_k8",
3131 values = {"cpu": "k8"},
3132)
3133
3134config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003135 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003136 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003137)
3138
3139config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 name = "android",
3141 values = {"crosstool_top": "//external:android/crosstool"},
3142)
3143
3144config_setting(
3145 name = "android_armv7",
3146 values = {
3147 "crosstool_top": "//external:android/crosstool",
3148 "cpu": "armeabi-v7a",
3149 },
3150)
3151
3152config_setting(
3153 name = "android_arm64",
3154 values = {
3155 "crosstool_top": "//external:android/crosstool",
3156 "cpu": "arm64-v8a",
3157 },
3158)
3159
3160config_setting(
3161 name = "android_x86",
3162 values = {
3163 "crosstool_top": "//external:android/crosstool",
3164 "cpu": "x86",
3165 },
3166)
3167
3168config_setting(
3169 name = "android_x86_64",
3170 values = {
3171 "crosstool_top": "//external:android/crosstool",
3172 "cpu": "x86_64",
3173 },
3174)
3175
3176config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003177 name = "macos_x86_64",
3178 values = {
3179 "apple_platform_type": "macos",
3180 "cpu": "darwin",
3181 },
3182)
3183
3184config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003185 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003186 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187)
3188
3189config_setting(
3190 name = "emscripten_wasm",
3191 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003192 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 "cpu": "wasm",
3194 },
3195)
3196
3197config_setting(
3198 name = "emscripten_wasmsimd",
3199 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003200 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003202 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 },
3204)
3205
3206config_setting(
3207 name = "emscripten_asmjs",
3208 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003209 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 "cpu": "asmjs",
3211 },
3212)