blob: 91d2a0309704b643394d1944142de80bf9c05d3e [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",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001477 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001478]
1479
1480ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1481 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001482]
1483
1484MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001485 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001486 "include/xnnpack.h",
1487]
1488
1489MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1490 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001491 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001492 "src/xnnpack/requantization.h",
1493 "include/xnnpack.h",
1494]
1495
1496OPERATOR_TEST_PARAMS_HDRS = [
1497 "src/xnnpack/params.h",
1498 "src/xnnpack/common.h",
1499]
1500
1501WEIGHTS_PACK_HDRS = [
1502 "src/xnnpack/pack.h",
1503 "src/xnnpack/operator.h",
1504 "src/xnnpack/compute.h",
1505]
1506
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001507LOGGING_COPTS = select({
1508 # No logging in optimized mode
1509 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1510 # Full logging in debug mode
1511 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1512 # Error-only logging in default (fastbuild) mode
1513 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1514})
1515
1516LOGGING_HDRS = [
1517 "src/xnnpack/log.h",
1518]
1519
Marat Dukhan08c4a432019-10-03 09:29:21 -07001520xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001521 name = "tables",
1522 srcs = TABLE_SRCS,
1523 hdrs = INTERNAL_HDRS,
1524 copts = xnnpack_std_copts(),
1525)
1526
1527xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528 name = "scalar_ukernels",
1529 srcs = SCALAR_UKERNELS,
1530 hdrs = INTERNAL_HDRS,
1531 aarch32_copts = ["-marm"],
1532 copts = xnnpack_std_copts(),
1533 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001534 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001535 "@FP16",
1536 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001537 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001538 ],
1539)
1540
1541xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001542 name = "wasm_ukernels",
1543 hdrs = INTERNAL_HDRS,
1544 copts = xnnpack_std_copts(),
1545 wasm_srcs = WASM_UKERNELS,
1546 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001547 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001548 "@FP16",
1549 "@FXdiv",
1550 "@pthreadpool",
1551 ],
1552)
1553
1554xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001555 name = "psimd_fastmath_ukernels",
1556 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001557 hdrs = INTERNAL_HDRS,
1558 aarch32_copts = [
1559 "-marm",
1560 "-mfpu=neon",
1561 ],
1562 copts = xnnpack_std_copts(),
1563 optimized_copts = [
1564 "-O3",
1565 "-ffast-math",
1566 ],
1567 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001568 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 "@FP16",
1570 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001571 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001572 ],
1573)
1574
1575xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001576 name = "psimd_accmath_ukernels",
1577 srcs = PSIMD_ACCMATH_UKERNELS,
1578 hdrs = INTERNAL_HDRS,
1579 aarch32_copts = [
1580 "-marm",
1581 "-mfpu=neon",
1582 ],
1583 copts = xnnpack_std_copts(),
1584 optimized_copts = [
1585 "-O3",
1586 ],
1587 deps = [
1588 ":tables",
1589 "@FP16",
1590 "@psimd",
1591 "@pthreadpool",
1592 ],
1593)
1594
1595xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001596 name = "neon_ukernels",
1597 hdrs = INTERNAL_HDRS,
1598 aarch32_copts = [
1599 "-marm",
1600 "-mfpu=neon",
1601 ],
1602 aarch32_srcs = NEON_UKERNELS,
1603 aarch64_srcs = NEON_UKERNELS,
1604 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001605 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001606 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001607 "@FP16",
1608 "@pthreadpool",
1609 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001610)
1611
1612xnnpack_cc_library(
1613 name = "neonfma_ukernels",
1614 hdrs = INTERNAL_HDRS,
1615 aarch32_copts = [
1616 "-marm",
1617 "-mfpu=neon-vfpv4",
1618 ],
1619 aarch32_srcs = NEONFMA_UKERNELS,
1620 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1621 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001622 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001623 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001624 "@FP16",
1625 "@pthreadpool",
1626 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001627)
1628
1629xnnpack_cc_library(
1630 name = "neonfp16arith_ukernels",
1631 hdrs = INTERNAL_HDRS,
1632 aarch64_copts = ["-march=armv8.2-a+fp16"],
1633 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1634 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001635 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001636 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001637 "@FP16",
1638 "@pthreadpool",
1639 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001640)
1641
1642xnnpack_cc_library(
1643 name = "sse2_ukernels",
1644 hdrs = INTERNAL_HDRS,
1645 copts = xnnpack_std_copts(),
1646 x86_copts = ["-msse2"],
1647 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001648 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001649 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001650 "@FP16",
1651 "@pthreadpool",
1652 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653)
1654
1655xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001656 name = "sse41_ukernels",
1657 hdrs = INTERNAL_HDRS,
1658 copts = xnnpack_std_copts(),
1659 x86_copts = ["-msse4.1"],
1660 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001661 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001662 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001663 "@FP16",
1664 "@pthreadpool",
1665 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001666)
1667
1668xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001669 name = "avx_ukernels",
1670 hdrs = INTERNAL_HDRS,
1671 copts = xnnpack_std_copts(),
1672 x86_copts = ["-mavx"],
1673 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001674 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001675 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001676 "@FP16",
1677 "@pthreadpool",
1678 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001679)
1680
1681xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001682 name = "fma3_ukernels",
1683 hdrs = INTERNAL_HDRS,
1684 copts = xnnpack_std_copts(),
1685 x86_copts = [
1686 "-mfma",
1687 ],
1688 x86_srcs = FMA3_UKERNELS,
1689 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001690 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001691 "@FP16",
1692 "@pthreadpool",
1693 ],
1694)
1695
1696xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001697 name = "avx2_ukernels",
1698 hdrs = INTERNAL_HDRS,
1699 copts = xnnpack_std_copts(),
1700 x86_copts = [
1701 "-mfma",
1702 "-mavx2",
1703 ],
1704 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001705 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001706 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001707 "@FP16",
1708 "@pthreadpool",
1709 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001710)
1711
1712xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713 name = "avx512f_ukernels",
1714 hdrs = INTERNAL_HDRS,
1715 copts = xnnpack_std_copts(),
1716 x86_copts = ["-mavx512f"],
1717 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001718 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001719 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001720 "@FP16",
1721 "@pthreadpool",
1722 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723)
1724
1725xnnpack_cc_library(
1726 name = "asm_ukernels",
1727 hdrs = ["src/xnnpack/assembly.h"],
1728 aarch32_srcs = AARCH32_ASM_UKERNELS,
1729 aarch64_srcs = AARCH64_ASM_UKERNELS,
1730)
1731
1732xnnpack_aggregate_library(
1733 name = "ukernels",
1734 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001735 ":psimd_fastmath_ukernels",
1736 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001737 ":neon_ukernels",
1738 ":neonfma_ukernels",
1739 ":asm_ukernels",
1740 ],
1741 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001742 ":psimd_fastmath_ukernels",
1743 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001744 ":neon_ukernels",
1745 ":neonfma_ukernels",
1746 ":neonfp16arith_ukernels",
1747 ":asm_ukernels",
1748 ],
1749 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001750 wasm_deps = [
1751 ":wasm_ukernels",
1752 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001753 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001754 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001755 ":psimd_fastmath_ukernels",
1756 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001757 ],
1758 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001759 ":psimd_fastmath_ukernels",
1760 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001761 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001762 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001763 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001764 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001765 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001766 ":avx512f_ukernels",
1767 ],
1768)
1769
1770xnnpack_cc_library(
1771 name = "im2col",
1772 srcs = ["src/im2col.c"],
1773 hdrs = [
1774 "src/xnnpack/common.h",
1775 "src/xnnpack/im2col.h",
1776 ],
1777 copts = xnnpack_std_copts(),
1778)
1779
1780xnnpack_cc_library(
1781 name = "indirection",
1782 srcs = ["src/indirection.c"],
1783 hdrs = INTERNAL_HDRS,
1784 copts = xnnpack_std_copts(),
1785 deps = [
1786 "@FP16",
1787 "@FXdiv",
1788 "@pthreadpool",
1789 ],
1790)
1791
1792xnnpack_cc_library(
1793 name = "operator_run",
1794 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001795 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1796 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001797 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1798 "-Wno-vla",
1799 ],
1800 deps = [
1801 "@FP16",
1802 "@FXdiv",
1803 "@clog",
1804 "@pthreadpool",
1805 ],
1806)
1807
1808cc_library(
1809 name = "enable_assembly",
1810 defines = select({
1811 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1812 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001813 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 }),
1815)
1816
Marat Dukhancf056b22019-10-07 10:26:29 -07001817xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818 name = "operators",
1819 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001820 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001821 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001822 ],
1823 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001824 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825 "-Isrc",
1826 "-Iinclude",
1827 ] + select({
1828 ":debug_build": [],
1829 "//conditions:default": xnnpack_min_size_copts(),
1830 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001831 wasm_srcs = ["src/wasm-stubs.c"],
1832 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001834 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "@FP16",
1836 "@FXdiv",
1837 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001839 ],
1840)
1841
1842cc_library(
1843 name = "XNNPACK",
1844 srcs = [
1845 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001846 "src/runtime.c",
1847 "src/subgraph.c",
1848 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001849 ],
1850 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1851 "-Isrc",
1852 "-Iinclude",
1853 ] + select({
1854 ":debug_build": [],
1855 "//conditions:default": xnnpack_min_size_copts(),
1856 }),
1857 includes = ["include"],
1858 linkstatic = True,
1859 textual_hdrs = ["include/xnnpack.h"],
1860 visibility = xnnpack_visibility(),
1861 deps = [
1862 ":enable_assembly",
1863 ":ukernels",
1864 ":operator_run",
1865 ":operators",
1866 "@clog",
1867 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001868 ] + select({
1869 ":emscripten": [],
1870 "//conditions:default": ["@cpuinfo"],
1871 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872)
1873
1874cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001875 name = "xnnpack_operators_nhwc_f32",
1876 srcs = [
1877 "src/init.c",
1878 ],
1879 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1880 "-Isrc",
1881 "-Iinclude",
1882 ] + select({
1883 ":debug_build": [],
1884 "//conditions:default": xnnpack_min_size_copts(),
1885 }),
1886 defines = [
1887 "XNN_NO_Q8_OPERATORS",
1888 "XNN_NO_U8_OPERATORS",
1889 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001890 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001891 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001892 includes = ["include"],
1893 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001894 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001895 visibility = xnnpack_visibility(),
1896 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001897 ":enable_assembly",
1898 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001899 ":operator_run",
1900 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001901 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001902 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001903 ] + select({
1904 ":emscripten": [],
1905 "//conditions:default": ["@cpuinfo"],
1906 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001907)
1908
Marat Dukhancf056b22019-10-07 10:26:29 -07001909xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001910 name = "bench_utils",
1911 srcs = ["bench/utils.cc"],
1912 hdrs = ["bench/utils.h"],
1913 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001914 deps = [
1915 "@com_google_benchmark//:benchmark",
1916 "@cpuinfo",
1917 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918)
1919
Frank Barchard7e955972019-10-11 10:34:25 -07001920######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001921
1922xnnpack_benchmark(
1923 name = "q8_gemm_bench",
1924 srcs = [
1925 "bench/gemm.h",
1926 "bench/q8-gemm.cc",
1927 "src/xnnpack/AlignedAllocator.h",
1928 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1929 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1930 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1931)
1932
1933xnnpack_benchmark(
1934 name = "f16_gemm_bench",
1935 srcs = [
1936 "bench/f16-gemm.cc",
1937 "bench/gemm.h",
1938 "src/xnnpack/AlignedAllocator.h",
1939 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1940 copts = ["-Wno-unused-function"],
1941 deps = MICROKERNEL_BENCHMARK_DEPS,
1942)
1943
1944xnnpack_benchmark(
1945 name = "f32_igemm_bench",
1946 srcs = [
1947 "bench/f32-igemm.cc",
1948 "bench/conv.h",
1949 "src/xnnpack/AlignedAllocator.h",
1950 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001951 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001952)
1953
1954xnnpack_benchmark(
1955 name = "f32_conv_hwc_bench",
1956 srcs = [
1957 "bench/f32-conv-hwc.cc",
1958 "bench/dconv.h",
1959 "src/xnnpack/AlignedAllocator.h",
1960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1961 copts = ["-Wno-unused-function"],
1962 deps = MICROKERNEL_BENCHMARK_DEPS,
1963)
1964
1965xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001966 name = "f32_conv_hwc2spchw_bench",
1967 srcs = [
1968 "bench/f32-conv-hwc2spchw.cc",
1969 "bench/dconv.h",
1970 "src/xnnpack/AlignedAllocator.h",
1971 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1972 copts = ["-Wno-unused-function"],
1973 deps = MICROKERNEL_BENCHMARK_DEPS,
1974)
1975
1976xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001977 name = "f32_dwconv_bench",
1978 srcs = [
1979 "bench/f32-dwconv.cc",
1980 "bench/dwconv.h",
1981 "src/xnnpack/AlignedAllocator.h",
1982 ] + MICROKERNEL_BENCHMARK_HDRS,
1983 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1984)
1985
1986xnnpack_benchmark(
1987 name = "f32_dwconv_spchw_bench",
1988 srcs = [
1989 "bench/f32-dwconv-spchw.cc",
1990 "bench/dwconv.h",
1991 "src/xnnpack/AlignedAllocator.h",
1992 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1993 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1994)
1995
1996xnnpack_benchmark(
1997 name = "f32_gemm_bench",
1998 srcs = [
1999 "bench/f32-gemm.cc",
2000 "bench/gemm.h",
2001 "src/xnnpack/AlignedAllocator.h",
2002 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2003 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002004 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002005)
2006
2007xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002008 name = "f32_raddexpminusmax_bench",
2009 srcs = [
2010 "bench/f32-raddexpminusmax.cc",
2011 "src/xnnpack/AlignedAllocator.h",
2012 ] + MICROKERNEL_BENCHMARK_HDRS,
2013 deps = MICROKERNEL_BENCHMARK_DEPS,
2014)
2015
2016xnnpack_benchmark(
2017 name = "f32_raddextexp_bench",
2018 srcs = [
2019 "bench/f32-raddextexp.cc",
2020 "src/xnnpack/AlignedAllocator.h",
2021 ] + MICROKERNEL_BENCHMARK_HDRS,
2022 deps = MICROKERNEL_BENCHMARK_DEPS,
2023)
2024
2025xnnpack_benchmark(
2026 name = "f32_raddstoreexpminusmax_bench",
2027 srcs = [
2028 "bench/f32-raddstoreexpminusmax.cc",
2029 "src/xnnpack/AlignedAllocator.h",
2030 ] + MICROKERNEL_BENCHMARK_HDRS,
2031 deps = MICROKERNEL_BENCHMARK_DEPS,
2032)
2033
2034xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002035 name = "f32_rmax_bench",
2036 srcs = [
2037 "bench/f32-rmax.cc",
2038 "src/xnnpack/AlignedAllocator.h",
2039 ] + MICROKERNEL_BENCHMARK_HDRS,
2040 deps = MICROKERNEL_BENCHMARK_DEPS,
2041)
2042
2043xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002044 name = "f32_sigmoid_bench",
2045 srcs = [
2046 "bench/f32-sigmoid.cc",
2047 "src/xnnpack/AlignedAllocator.h",
2048 ] + MICROKERNEL_BENCHMARK_HDRS,
2049 copts = ["-Wno-unused-function"],
2050 deps = MICROKERNEL_BENCHMARK_DEPS,
2051)
2052
2053xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 name = "f32_spmm_bench",
2055 srcs = [
2056 "bench/f32-spmm.cc",
2057 "bench/gemm.h",
2058 "src/xnnpack/AlignedAllocator.h",
2059 ] + MICROKERNEL_BENCHMARK_HDRS,
2060 copts = ["-Wno-unused-function"],
2061 deps = MICROKERNEL_BENCHMARK_DEPS,
2062)
2063
2064xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002065 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002066 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002067 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002068 ] + MICROKERNEL_BENCHMARK_HDRS,
2069 copts = ["-Wno-unused-function"],
2070 deps = MICROKERNEL_BENCHMARK_DEPS,
2071)
2072
2073xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002074 name = "f32_vscaleexpminusmax_bench",
2075 srcs = [
2076 "bench/f32-vscaleexpminusmax.cc",
2077 "src/xnnpack/AlignedAllocator.h",
2078 ] + MICROKERNEL_BENCHMARK_HDRS,
2079 deps = MICROKERNEL_BENCHMARK_DEPS,
2080)
2081
2082xnnpack_benchmark(
2083 name = "f32_vscaleextexp_bench",
2084 srcs = [
2085 "bench/f32-vscaleextexp.cc",
2086 "src/xnnpack/AlignedAllocator.h",
2087 ] + MICROKERNEL_BENCHMARK_HDRS,
2088 deps = MICROKERNEL_BENCHMARK_DEPS,
2089)
2090
2091xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092 name = "f32_im2col_gemm_bench",
2093 srcs = [
2094 "bench/f32-im2col-gemm.cc",
2095 "bench/conv.h",
2096 "src/xnnpack/AlignedAllocator.h",
2097 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2098 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2099)
2100
2101########################### Benchmarks for operators ###########################
2102
2103xnnpack_benchmark(
2104 name = "add_bench",
2105 srcs = ["bench/add.cc"],
2106 deps = OPERATOR_BENCHMARK_DEPS,
2107)
2108
2109xnnpack_benchmark(
2110 name = "average_pooling_bench",
2111 srcs = ["bench/average-pooling.cc"],
2112 deps = OPERATOR_BENCHMARK_DEPS,
2113)
2114
2115xnnpack_benchmark(
2116 name = "channel_shuffle_bench",
2117 srcs = ["bench/channel-shuffle.cc"],
2118 deps = OPERATOR_BENCHMARK_DEPS,
2119)
2120
2121xnnpack_benchmark(
2122 name = "convolution_bench",
2123 srcs = ["bench/convolution.cc"],
2124 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2125 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2126)
2127
2128xnnpack_benchmark(
2129 name = "deconvolution_bench",
2130 srcs = ["bench/deconvolution.cc"],
2131 copts = xnnpack_optional_tflite_copts(),
2132 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2133)
2134
2135xnnpack_benchmark(
2136 name = "global_average_pooling_bench",
2137 srcs = ["bench/global-average-pooling.cc"],
2138 deps = OPERATOR_BENCHMARK_DEPS,
2139)
2140
2141xnnpack_benchmark(
2142 name = "max_pooling_bench",
2143 srcs = ["bench/max-pooling.cc"],
2144 deps = OPERATOR_BENCHMARK_DEPS,
2145)
2146
2147xnnpack_benchmark(
2148 name = "sigmoid_bench",
2149 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002150 copts = xnnpack_optional_tflite_copts(),
2151 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002152)
2153
2154xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002155 name = "prelu_bench",
2156 srcs = ["bench/prelu.cc"],
2157 copts = xnnpack_optional_tflite_copts(),
2158 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2159)
2160
2161xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002162 name = "softmax_bench",
2163 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002164 copts = xnnpack_optional_tflite_copts(),
2165 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166)
2167
Marat Dukhanc068bb62019-10-04 13:24:39 -07002168############################# End-to-end benchmarks ############################
2169
2170cc_library(
2171 name = "mobilenet_v1",
2172 srcs = ["models/mobilenet-v1.cc"],
2173 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002174 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002175 linkstatic = True,
2176 deps = [
2177 ":XNNPACK",
2178 "@pthreadpool",
2179 ],
2180)
2181
2182cc_library(
2183 name = "mobilenet_v2",
2184 srcs = ["models/mobilenet-v2.cc"],
2185 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002186 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002187 linkstatic = True,
2188 deps = [
2189 ":XNNPACK",
2190 "@pthreadpool",
2191 ],
2192)
2193
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002194cc_library(
2195 name = "mobilenet_v3_large",
2196 srcs = ["models/mobilenet-v3-large.cc"],
2197 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002198 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002199 linkstatic = True,
2200 deps = [
2201 ":XNNPACK",
2202 "@pthreadpool",
2203 ],
2204)
2205
2206cc_library(
2207 name = "mobilenet_v3_small",
2208 srcs = ["models/mobilenet-v3-small.cc"],
2209 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002210 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002211 linkstatic = True,
2212 deps = [
2213 ":XNNPACK",
2214 "@pthreadpool",
2215 ],
2216)
2217
Marat Dukhanc068bb62019-10-04 13:24:39 -07002218xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002219 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002220 srcs = [
2221 "bench/f32-dwconv-e2e.cc",
2222 "bench/end2end.h",
2223 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002224 copts = ["-Wno-unused-function"],
2225 deps = MICROKERNEL_BENCHMARK_DEPS + [
2226 ":XNNPACK",
2227 ":mobilenet_v1",
2228 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002229 ":mobilenet_v3_large",
2230 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002231 ],
2232)
2233
2234xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002235 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002236 srcs = [
2237 "bench/f32-gemm-e2e.cc",
2238 "bench/end2end.h",
2239 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002240 copts = ["-Wno-unused-function"],
2241 deps = MICROKERNEL_BENCHMARK_DEPS + [
2242 ":XNNPACK",
2243 ":mobilenet_v1",
2244 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002245 ":mobilenet_v3_large",
2246 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002247 ],
2248)
2249
2250xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002251 name = "end2end_bench",
2252 srcs = ["bench/end2end.cc"],
2253 deps = [
2254 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002255 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002256 ":mobilenet_v1",
2257 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002258 ":mobilenet_v3_large",
2259 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002260 "@pthreadpool",
2261 ],
2262)
2263
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002264#################### Accuracy evaluation for math functions ####################
2265
2266xnnpack_benchmark(
2267 name = "f32_exp_eval",
2268 srcs = [
2269 "eval/f32-exp.cc",
2270 "src/xnnpack/AlignedAllocator.h",
2271 ] + ACCURACY_EVAL_HDRS,
2272 deps = ACCURACY_EVAL_DEPS,
2273)
2274
Marat Dukhan515c9772019-10-17 18:07:57 -07002275xnnpack_benchmark(
2276 name = "f32_expminus_eval",
2277 srcs = [
2278 "eval/f32-expminus.cc",
2279 "src/xnnpack/AlignedAllocator.h",
2280 ] + ACCURACY_EVAL_HDRS,
2281 deps = ACCURACY_EVAL_DEPS,
2282)
2283
Marat Dukhan98ba4412019-10-23 02:14:28 -07002284xnnpack_benchmark(
2285 name = "f32_extexp_eval",
2286 srcs = [
2287 "eval/f32-extexp.cc",
2288 "src/xnnpack/AlignedAllocator.h",
2289 ] + ACCURACY_EVAL_HDRS,
2290 deps = ACCURACY_EVAL_DEPS,
2291)
2292
Marat Dukhan346a9e52019-11-15 09:06:30 -08002293xnnpack_benchmark(
2294 name = "f32_sigmoid_eval",
2295 srcs = [
2296 "eval/f32-sigmoid.cc",
2297 "src/xnnpack/AlignedAllocator.h",
2298 ] + ACCURACY_EVAL_HDRS,
2299 deps = ACCURACY_EVAL_DEPS,
2300)
2301
Marat Dukhan08c4a432019-10-03 09:29:21 -07002302######################### Unit tests for micro-kernels #########################
2303
2304xnnpack_unit_test(
2305 name = "f16_gemm_test",
2306 srcs = [
2307 "test/f16-gemm.cc",
2308 "test/gemm-microkernel-tester.h",
2309 "src/xnnpack/AlignedAllocator.h",
2310 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2311 deps = MICROKERNEL_TEST_DEPS,
2312)
2313
2314xnnpack_unit_test(
2315 name = "f32_argmaxpool_test",
2316 srcs = [
2317 "test/f32-argmaxpool.cc",
2318 "test/argmaxpool-microkernel-tester.h",
2319 "src/xnnpack/AlignedAllocator.h",
2320 ] + MICROKERNEL_TEST_HDRS,
2321 deps = MICROKERNEL_TEST_DEPS,
2322)
2323
2324xnnpack_unit_test(
2325 name = "f32_avgpool_test",
2326 srcs = [
2327 "test/f32-avgpool.cc",
2328 "test/avgpool-microkernel-tester.h",
2329 "src/xnnpack/AlignedAllocator.h",
2330 ] + MICROKERNEL_TEST_HDRS,
2331 deps = MICROKERNEL_TEST_DEPS,
2332)
2333
2334xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002335 name = "f32_bilinear_test",
2336 srcs = [
2337 "test/f32-bilinear.cc",
2338 "test/bilinear-microkernel-tester.h",
2339 "src/xnnpack/AlignedAllocator.h",
2340 ] + MICROKERNEL_TEST_HDRS,
2341 deps = MICROKERNEL_TEST_DEPS,
2342)
2343
2344xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345 name = "f32_clamp_test",
2346 srcs = [
2347 "test/f32-clamp.cc",
2348 "test/clamp-microkernel-tester.h",
2349 ] + MICROKERNEL_TEST_HDRS,
2350 deps = MICROKERNEL_TEST_DEPS,
2351)
2352
2353xnnpack_unit_test(
2354 name = "f32_igemm_test",
2355 srcs = [
2356 "test/f32-igemm.cc",
2357 "test/gemm-microkernel-tester.h",
2358 "src/xnnpack/AlignedAllocator.h",
2359 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2360 deps = MICROKERNEL_TEST_DEPS,
2361)
2362
2363xnnpack_unit_test(
2364 name = "f32_conv_hwc_test",
2365 srcs = [
2366 "test/f32-conv-hwc.cc",
2367 "test/conv-hwc-microkernel-tester.h",
2368 "src/xnnpack/AlignedAllocator.h",
2369 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2370 deps = MICROKERNEL_TEST_DEPS,
2371)
2372
2373xnnpack_unit_test(
2374 name = "f32_conv_hwc2spchw_test",
2375 srcs = [
2376 "test/f32-conv-hwc2spchw.cc",
2377 "test/conv-hwc2spchw-microkernel-tester.h",
2378 "src/xnnpack/AlignedAllocator.h",
2379 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2380 deps = MICROKERNEL_TEST_DEPS,
2381)
2382
2383xnnpack_unit_test(
2384 name = "f32_dwconv_test",
2385 srcs = [
2386 "test/f32-dwconv.cc",
2387 "test/dwconv-microkernel-tester.h",
2388 "src/xnnpack/AlignedAllocator.h",
2389 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2390 deps = MICROKERNEL_TEST_DEPS,
2391)
2392
2393xnnpack_unit_test(
2394 name = "f32_dwconv_spchw_test",
2395 srcs = [
2396 "test/f32-dwconv-spchw.cc",
2397 "test/dwconv-spchw-microkernel-tester.h",
2398 "src/xnnpack/AlignedAllocator.h",
2399 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2400 deps = MICROKERNEL_TEST_DEPS,
2401)
2402
2403xnnpack_unit_test(
2404 name = "f32_gavgpool_test",
2405 srcs = [
2406 "test/f32-gavgpool.cc",
2407 "test/gavgpool-microkernel-tester.h",
2408 "src/xnnpack/AlignedAllocator.h",
2409 ] + MICROKERNEL_TEST_HDRS,
2410 deps = MICROKERNEL_TEST_DEPS,
2411)
2412
2413xnnpack_unit_test(
2414 name = "f32_gavgpool_spchw_test",
2415 srcs = [
2416 "test/f32-gavgpool-spchw.cc",
2417 "test/gavgpool-spchw-microkernel-tester.h",
2418 "src/xnnpack/AlignedAllocator.h",
2419 ] + MICROKERNEL_TEST_HDRS,
2420 deps = MICROKERNEL_TEST_DEPS,
2421)
2422
2423xnnpack_unit_test(
2424 name = "f32_gemm_test",
2425 srcs = [
2426 "test/f32-gemm.cc",
2427 "test/gemm-microkernel-tester.h",
2428 "src/xnnpack/AlignedAllocator.h",
2429 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2430 deps = MICROKERNEL_TEST_DEPS,
2431)
2432
2433xnnpack_unit_test(
2434 name = "f32_gemminc_test",
2435 srcs = [
2436 "test/f32-gemminc.cc",
2437 "test/gemm-microkernel-tester.h",
2438 "src/xnnpack/AlignedAllocator.h",
2439 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2440 deps = MICROKERNEL_TEST_DEPS,
2441)
2442
2443xnnpack_unit_test(
2444 name = "f32_hswish_test",
2445 srcs = [
2446 "test/f32-hswish.cc",
2447 "test/hswish-microkernel-tester.h",
2448 ] + MICROKERNEL_TEST_HDRS,
2449 deps = MICROKERNEL_TEST_DEPS,
2450)
2451
2452xnnpack_unit_test(
2453 name = "f32_maxpool_test",
2454 srcs = [
2455 "test/f32-maxpool.cc",
2456 "test/maxpool-microkernel-tester.h",
2457 ] + MICROKERNEL_TEST_HDRS,
2458 deps = MICROKERNEL_TEST_DEPS,
2459)
2460
2461xnnpack_unit_test(
2462 name = "f32_pavgpool_test",
2463 srcs = [
2464 "test/f32-pavgpool.cc",
2465 "test/avgpool-microkernel-tester.h",
2466 "src/xnnpack/AlignedAllocator.h",
2467 ] + MICROKERNEL_TEST_HDRS,
2468 deps = MICROKERNEL_TEST_DEPS,
2469)
2470
2471xnnpack_unit_test(
2472 name = "f32_ppmm_test",
2473 srcs = [
2474 "test/f32-ppmm.cc",
2475 "test/gemm-microkernel-tester.h",
2476 "src/xnnpack/AlignedAllocator.h",
2477 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2478 deps = MICROKERNEL_TEST_DEPS,
2479)
2480
2481xnnpack_unit_test(
2482 name = "f32_prelu_test",
2483 srcs = [
2484 "test/f32-prelu.cc",
2485 "test/prelu-microkernel-tester.h",
2486 "src/xnnpack/AlignedAllocator.h",
2487 ] + MICROKERNEL_TEST_HDRS,
2488 deps = MICROKERNEL_TEST_DEPS,
2489)
2490
2491xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002492 name = "f32_raddexpminusmax_test",
2493 srcs = [
2494 "test/f32-raddexpminusmax.cc",
2495 "test/raddexpminusmax-microkernel-tester.h",
2496 ] + MICROKERNEL_TEST_HDRS,
2497 deps = MICROKERNEL_TEST_DEPS,
2498)
2499
2500xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002501 name = "f32_raddextexp_test",
2502 srcs = [
2503 "test/f32-raddextexp.cc",
2504 "test/raddextexp-microkernel-tester.h",
2505 ] + MICROKERNEL_TEST_HDRS,
2506 deps = MICROKERNEL_TEST_DEPS,
2507)
2508
2509xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002510 name = "f32_raddstoreexpminusmax_test",
2511 srcs = [
2512 "test/f32-raddstoreexpminusmax.cc",
2513 "test/raddstoreexpminusmax-microkernel-tester.h",
2514 ] + MICROKERNEL_TEST_HDRS,
2515 deps = MICROKERNEL_TEST_DEPS,
2516)
2517
2518xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002519 name = "f32_rmax_test",
2520 srcs = [
2521 "test/f32-rmax.cc",
2522 "test/rmax-microkernel-tester.h",
2523 ] + MICROKERNEL_TEST_HDRS,
2524 deps = MICROKERNEL_TEST_DEPS,
2525)
2526
2527xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002528 name = "f32_sigmoid_test",
2529 srcs = [
2530 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002531 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002532 ] + MICROKERNEL_TEST_HDRS,
2533 deps = MICROKERNEL_TEST_DEPS,
2534)
2535
2536xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002537 name = "f32_spmm_test",
2538 srcs = [
2539 "test/f32-spmm.cc",
2540 "test/spmm-microkernel-tester.h",
2541 "src/xnnpack/AlignedAllocator.h",
2542 ] + MICROKERNEL_TEST_HDRS,
2543 deps = MICROKERNEL_TEST_DEPS,
2544)
2545
2546xnnpack_unit_test(
2547 name = "f32_vadd_test",
2548 srcs = [
2549 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002550 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002551 ] + MICROKERNEL_TEST_HDRS,
2552 deps = MICROKERNEL_TEST_DEPS,
2553)
2554
2555xnnpack_unit_test(
2556 name = "f32_vaddc_test",
2557 srcs = [
2558 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002559 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560 ] + MICROKERNEL_TEST_HDRS,
2561 deps = MICROKERNEL_TEST_DEPS,
2562)
2563
2564xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002565 name = "f32_vdiv_test",
2566 srcs = [
2567 "test/f32-vdiv.cc",
2568 "test/vbinary-microkernel-tester.h",
2569 ] + MICROKERNEL_TEST_HDRS,
2570 deps = MICROKERNEL_TEST_DEPS,
2571)
2572
2573xnnpack_unit_test(
2574 name = "f32_vdivc_test",
2575 srcs = [
2576 "test/f32-vdivc.cc",
2577 "test/vbinaryc-microkernel-tester.h",
2578 ] + MICROKERNEL_TEST_HDRS,
2579 deps = MICROKERNEL_TEST_DEPS,
2580)
2581
2582xnnpack_unit_test(
2583 name = "f32_vrdivc_test",
2584 srcs = [
2585 "test/f32-vrdivc.cc",
2586 "test/vbinaryc-microkernel-tester.h",
2587 ] + MICROKERNEL_TEST_HDRS,
2588 deps = MICROKERNEL_TEST_DEPS,
2589)
2590
2591xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002592 name = "f32_vmax_test",
2593 srcs = [
2594 "test/f32-vmax.cc",
2595 "test/vbinary-microkernel-tester.h",
2596 ] + MICROKERNEL_TEST_HDRS,
2597 deps = MICROKERNEL_TEST_DEPS,
2598)
2599
2600xnnpack_unit_test(
2601 name = "f32_vmaxc_test",
2602 srcs = [
2603 "test/f32-vmaxc.cc",
2604 "test/vbinaryc-microkernel-tester.h",
2605 ] + MICROKERNEL_TEST_HDRS,
2606 deps = MICROKERNEL_TEST_DEPS,
2607)
2608
2609xnnpack_unit_test(
2610 name = "f32_vmin_test",
2611 srcs = [
2612 "test/f32-vmin.cc",
2613 "test/vbinary-microkernel-tester.h",
2614 ] + MICROKERNEL_TEST_HDRS,
2615 deps = MICROKERNEL_TEST_DEPS,
2616)
2617
2618xnnpack_unit_test(
2619 name = "f32_vminc_test",
2620 srcs = [
2621 "test/f32-vminc.cc",
2622 "test/vbinaryc-microkernel-tester.h",
2623 ] + MICROKERNEL_TEST_HDRS,
2624 deps = MICROKERNEL_TEST_DEPS,
2625)
2626
2627xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002628 name = "f32_vmul_test",
2629 srcs = [
2630 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002631 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002632 ] + MICROKERNEL_TEST_HDRS,
2633 deps = MICROKERNEL_TEST_DEPS,
2634)
2635
2636xnnpack_unit_test(
2637 name = "f32_vmulc_test",
2638 srcs = [
2639 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002640 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002641 ] + MICROKERNEL_TEST_HDRS,
2642 deps = MICROKERNEL_TEST_DEPS,
2643)
2644
2645xnnpack_unit_test(
2646 name = "f32_vmulcaddc_test",
2647 srcs = [
2648 "test/f32-vmulcaddc.cc",
2649 "test/vmulcaddc-microkernel-tester.h",
2650 "src/xnnpack/AlignedAllocator.h",
2651 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2652 deps = MICROKERNEL_TEST_DEPS,
2653)
2654
2655xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002656 name = "f32_vscale_test",
2657 srcs = [
2658 "test/f32-vscale.cc",
2659 "test/vscale-microkernel-tester.h",
2660 ] + MICROKERNEL_TEST_HDRS,
2661 deps = MICROKERNEL_TEST_DEPS,
2662)
2663
2664xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002665 name = "f32_vscaleexpminusmax_test",
2666 srcs = [
2667 "test/f32-vscaleexpminusmax.cc",
2668 "test/vscaleexpminusmax-microkernel-tester.h",
2669 ] + MICROKERNEL_TEST_HDRS,
2670 deps = MICROKERNEL_TEST_DEPS,
2671)
2672
2673xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002674 name = "f32_vscaleextexp_test",
2675 srcs = [
2676 "test/f32-vscaleextexp.cc",
2677 "test/vscaleextexp-microkernel-tester.h",
2678 ] + MICROKERNEL_TEST_HDRS,
2679 deps = MICROKERNEL_TEST_DEPS,
2680)
2681
2682xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002683 name = "f32_vsub_test",
2684 srcs = [
2685 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002686 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002687 ] + MICROKERNEL_TEST_HDRS,
2688 deps = MICROKERNEL_TEST_DEPS,
2689)
2690
2691xnnpack_unit_test(
2692 name = "f32_vsubc_test",
2693 srcs = [
2694 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002695 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002696 ] + MICROKERNEL_TEST_HDRS,
2697 deps = MICROKERNEL_TEST_DEPS,
2698)
2699
2700xnnpack_unit_test(
2701 name = "f32_vrsubc_test",
2702 srcs = [
2703 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002704 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002705 ] + MICROKERNEL_TEST_HDRS,
2706 deps = MICROKERNEL_TEST_DEPS,
2707)
2708
2709xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002710 name = "q8_avgpool_test",
2711 srcs = [
2712 "test/q8-avgpool.cc",
2713 "test/avgpool-microkernel-tester.h",
2714 "src/xnnpack/AlignedAllocator.h",
2715 ] + MICROKERNEL_TEST_HDRS,
2716 deps = MICROKERNEL_TEST_DEPS,
2717)
2718
2719xnnpack_unit_test(
2720 name = "q8_igemm_test",
2721 srcs = [
2722 "test/q8-igemm.cc",
2723 "test/gemm-microkernel-tester.h",
2724 "src/xnnpack/AlignedAllocator.h",
2725 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2726 deps = MICROKERNEL_TEST_DEPS,
2727)
2728
2729xnnpack_unit_test(
2730 name = "q8_dwconv_test",
2731 srcs = [
2732 "test/q8-dwconv.cc",
2733 "test/dwconv-microkernel-tester.h",
2734 "src/xnnpack/AlignedAllocator.h",
2735 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2736 deps = MICROKERNEL_TEST_DEPS,
2737)
2738
2739xnnpack_unit_test(
2740 name = "q8_gavgpool_test",
2741 srcs = [
2742 "test/q8-gavgpool.cc",
2743 "test/gavgpool-microkernel-tester.h",
2744 "src/xnnpack/AlignedAllocator.h",
2745 ] + MICROKERNEL_TEST_HDRS,
2746 deps = MICROKERNEL_TEST_DEPS,
2747)
2748
2749xnnpack_unit_test(
2750 name = "q8_gemm_test",
2751 srcs = [
2752 "test/q8-gemm.cc",
2753 "test/gemm-microkernel-tester.h",
2754 "src/xnnpack/AlignedAllocator.h",
2755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2756 deps = MICROKERNEL_TEST_DEPS,
2757)
2758
2759xnnpack_unit_test(
2760 name = "q8_vadd_test",
2761 srcs = [
2762 "test/q8-vadd.cc",
2763 "test/vadd-microkernel-tester.h",
2764 ] + MICROKERNEL_TEST_HDRS,
2765 deps = MICROKERNEL_TEST_DEPS,
2766)
2767
2768xnnpack_unit_test(
2769 name = "u8_clamp_test",
2770 srcs = [
2771 "test/u8-clamp.cc",
2772 "test/clamp-microkernel-tester.h",
2773 ] + MICROKERNEL_TEST_HDRS,
2774 deps = MICROKERNEL_TEST_DEPS,
2775)
2776
2777xnnpack_unit_test(
2778 name = "u8_lut32norm_test",
2779 srcs = [
2780 "test/u8-lut32norm.cc",
2781 "test/lut-norm-microkernel-tester.h",
2782 ] + MICROKERNEL_TEST_HDRS,
2783 deps = MICROKERNEL_TEST_DEPS,
2784)
2785
2786xnnpack_unit_test(
2787 name = "u8_maxpool_test",
2788 srcs = [
2789 "test/u8-maxpool.cc",
2790 "test/maxpool-microkernel-tester.h",
2791 ] + MICROKERNEL_TEST_HDRS,
2792 deps = MICROKERNEL_TEST_DEPS,
2793)
2794
2795xnnpack_unit_test(
2796 name = "u8_rmax_test",
2797 srcs = [
2798 "test/u8-rmax.cc",
2799 "test/rmax-microkernel-tester.h",
2800 ] + MICROKERNEL_TEST_HDRS,
2801 deps = MICROKERNEL_TEST_DEPS,
2802)
2803
2804xnnpack_unit_test(
2805 name = "x32_packx_test",
2806 srcs = [
2807 "test/x32-packx.cc",
2808 "test/pack-microkernel-tester.h",
2809 "src/xnnpack/AlignedAllocator.h",
2810 ] + MICROKERNEL_TEST_HDRS,
2811 deps = MICROKERNEL_TEST_DEPS,
2812)
2813
2814xnnpack_unit_test(
2815 name = "x32_pad_test",
2816 srcs = [
2817 "test/x32-pad.cc",
2818 "test/pad-microkernel-tester.h",
2819 ] + MICROKERNEL_TEST_HDRS,
2820 deps = MICROKERNEL_TEST_DEPS,
2821)
2822
2823xnnpack_unit_test(
2824 name = "x32_unpool_test",
2825 srcs = [
2826 "test/x32-unpool.cc",
2827 "test/unpool-microkernel-tester.h",
2828 ] + MICROKERNEL_TEST_HDRS,
2829 deps = MICROKERNEL_TEST_DEPS,
2830)
2831
2832xnnpack_unit_test(
2833 name = "x32_zip_test",
2834 srcs = [
2835 "test/x32-zip.cc",
2836 "test/zip-microkernel-tester.h",
2837 ] + MICROKERNEL_TEST_HDRS,
2838 deps = MICROKERNEL_TEST_DEPS,
2839)
2840
2841xnnpack_unit_test(
2842 name = "x8_lut_test",
2843 srcs = [
2844 "test/x8-lut.cc",
2845 "test/lut-microkernel-tester.h",
2846 ] + MICROKERNEL_TEST_HDRS,
2847 deps = MICROKERNEL_TEST_DEPS,
2848)
2849
2850xnnpack_unit_test(
2851 name = "x8_zip_test",
2852 srcs = [
2853 "test/x8-zip.cc",
2854 "test/zip-microkernel-tester.h",
2855 ] + MICROKERNEL_TEST_HDRS,
2856 deps = MICROKERNEL_TEST_DEPS,
2857)
2858
2859########################### Size test for the library ##########################
2860
2861xnnpack_binary(
2862 name = "size_test",
2863 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002864 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865)
2866
2867########################### Unit tests for operators ###########################
2868
2869xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002870 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002872 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002873 "test/add-operator-tester.h",
2874 ],
2875 deps = OPERATOR_TEST_DEPS,
2876)
2877
2878xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002879 name = "add_nd_test",
2880 srcs = [
2881 "test/add-nd.cc",
2882 "test/binary-elementwise-operator-tester.h",
2883 ],
2884 deps = OPERATOR_TEST_DEPS,
2885)
2886
2887xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002888 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002889 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002890 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 "test/argmax-pooling-operator-tester.h",
2892 ] + OPERATOR_TEST_PARAMS_HDRS,
2893 deps = OPERATOR_TEST_DEPS,
2894)
2895
2896xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002897 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002899 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002900 "test/average-pooling-operator-tester.h",
2901 ] + OPERATOR_TEST_PARAMS_HDRS,
2902 deps = OPERATOR_TEST_DEPS,
2903)
2904
2905xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002906 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002907 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002908 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002909 "test/channel-pad-operator-tester.h",
2910 ] + OPERATOR_TEST_PARAMS_HDRS,
2911 deps = OPERATOR_TEST_DEPS,
2912)
2913
2914xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002915 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002916 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002917 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002918 "test/channel-shuffle-operator-tester.h",
2919 ],
2920 deps = OPERATOR_TEST_DEPS,
2921)
2922
2923xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002924 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002925 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002926 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002927 "test/clamp-operator-tester.h",
2928 ],
2929 deps = OPERATOR_TEST_DEPS,
2930)
2931
2932xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002933 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002935 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 "test/convolution-operator-tester.h",
2937 ],
2938 deps = OPERATOR_TEST_DEPS,
2939)
2940
2941xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002942 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002944 "test/convolution-nchw.cc",
2945 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 ],
2947 deps = OPERATOR_TEST_DEPS,
2948)
2949
2950xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002951 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002952 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002953 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002954 "test/deconvolution-operator-tester.h",
2955 ] + OPERATOR_TEST_PARAMS_HDRS,
2956 deps = OPERATOR_TEST_DEPS,
2957)
2958
2959xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002960 name = "divide_nd_test",
2961 srcs = [
2962 "test/binary-elementwise-operator-tester.h",
2963 "test/divide-nd.cc",
2964 ],
2965 deps = OPERATOR_TEST_DEPS,
2966)
2967
2968xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002969 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002971 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002972 "test/fully-connected-operator-tester.h",
2973 ],
2974 deps = OPERATOR_TEST_DEPS,
2975)
2976
2977xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002978 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002980 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002981 "test/global-average-pooling-operator-tester.h",
2982 ] + OPERATOR_TEST_PARAMS_HDRS,
2983 deps = OPERATOR_TEST_DEPS,
2984)
2985
2986xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002987 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002989 "test/global-average-pooling-ncw.cc",
2990 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 ],
2992 deps = OPERATOR_TEST_DEPS,
2993)
2994
2995xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002996 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002997 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002998 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 "test/hardswish-operator-tester.h",
3000 ],
3001 deps = OPERATOR_TEST_DEPS,
3002)
3003
3004xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003005 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003007 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 "test/leaky-relu-operator-tester.h",
3009 ],
3010 deps = OPERATOR_TEST_DEPS,
3011)
3012
3013xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003014 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003016 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003017 "test/max-pooling-operator-tester.h",
3018 ] + OPERATOR_TEST_PARAMS_HDRS,
3019 deps = OPERATOR_TEST_DEPS,
3020)
3021
3022xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003023 name = "maximum_nd_test",
3024 srcs = [
3025 "test/binary-elementwise-operator-tester.h",
3026 "test/maximum-nd.cc",
3027 ],
3028 deps = OPERATOR_TEST_DEPS,
3029)
3030
3031xnnpack_unit_test(
3032 name = "minimum_nd_test",
3033 srcs = [
3034 "test/binary-elementwise-operator-tester.h",
3035 "test/minimum-nd.cc",
3036 ],
3037 deps = OPERATOR_TEST_DEPS,
3038)
3039
3040xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003041 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003042 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003043 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003044 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003045 ],
3046 deps = OPERATOR_TEST_DEPS,
3047)
3048
3049xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003050 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003051 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003052 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003053 "test/prelu-operator-tester.h",
3054 ] + OPERATOR_TEST_PARAMS_HDRS,
3055 deps = OPERATOR_TEST_DEPS,
3056)
3057
3058xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003059 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003060 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003061 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003062 "test/resize-bilinear-operator-tester.h",
3063 ] + OPERATOR_TEST_PARAMS_HDRS,
3064 deps = OPERATOR_TEST_DEPS,
3065)
3066
3067xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003068 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003070 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003071 "test/sigmoid-operator-tester.h",
3072 ],
3073 deps = OPERATOR_TEST_DEPS,
3074)
3075
3076xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003077 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003079 "test/softmax-nc.cc",
3080 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003081 ],
3082 deps = OPERATOR_TEST_DEPS,
3083)
3084
3085xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003086 name = "subtract_nd_test",
3087 srcs = [
3088 "test/binary-elementwise-operator-tester.h",
3089 "test/subtract-nd.cc",
3090 ],
3091 deps = OPERATOR_TEST_DEPS,
3092)
3093
3094xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003095 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003097 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003098 "test/unpooling-operator-tester.h",
3099 ],
3100 deps = OPERATOR_TEST_DEPS,
3101)
3102
3103############################# Build configurations #############################
3104
Marat Dukhanb8642352019-10-30 15:43:02 -07003105# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003107 name = "xnn_enable_assembly_explicit_true",
3108 define_values = {"xnn_enable_assembly": "true"},
3109)
3110
3111# Disables usage of assembly kernels.
3112config_setting(
3113 name = "xnn_enable_assembly_explicit_false",
3114 define_values = {"xnn_enable_assembly": "false"},
3115)
3116
3117# Builds with -c dbg
3118config_setting(
3119 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003121 "compilation_mode": "dbg",
3122 },
3123)
3124
3125# Builds with -c opt
3126config_setting(
3127 name = "optimized_build",
3128 values = {
3129 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 },
3131)
3132
3133config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003134 name = "linux_k8",
3135 values = {"cpu": "k8"},
3136)
3137
3138config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003139 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003140 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003141)
3142
3143config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 name = "android",
3145 values = {"crosstool_top": "//external:android/crosstool"},
3146)
3147
3148config_setting(
3149 name = "android_armv7",
3150 values = {
3151 "crosstool_top": "//external:android/crosstool",
3152 "cpu": "armeabi-v7a",
3153 },
3154)
3155
3156config_setting(
3157 name = "android_arm64",
3158 values = {
3159 "crosstool_top": "//external:android/crosstool",
3160 "cpu": "arm64-v8a",
3161 },
3162)
3163
3164config_setting(
3165 name = "android_x86",
3166 values = {
3167 "crosstool_top": "//external:android/crosstool",
3168 "cpu": "x86",
3169 },
3170)
3171
3172config_setting(
3173 name = "android_x86_64",
3174 values = {
3175 "crosstool_top": "//external:android/crosstool",
3176 "cpu": "x86_64",
3177 },
3178)
3179
3180config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003181 name = "macos_x86_64",
3182 values = {
3183 "apple_platform_type": "macos",
3184 "cpu": "darwin",
3185 },
3186)
3187
3188config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003190 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191)
3192
3193config_setting(
3194 name = "emscripten_wasm",
3195 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003196 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197 "cpu": "wasm",
3198 },
3199)
3200
3201config_setting(
3202 name = "emscripten_wasmsimd",
3203 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003204 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003205 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003206 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207 },
3208)
3209
3210config_setting(
3211 name = "emscripten_asmjs",
3212 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003213 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 "cpu": "asmjs",
3215 },
3216)