blob: a6e6c1f485107ff3ac7ffbb91942a7abe966695d [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhana84e40b2019-12-11 15:38:03 -08009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
Marat Dukhanfd8e6892020-01-27 15:25:25 -080072 "src/softmax-nc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080073 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan3a77ea72019-12-23 12:10:24 -080076TABLE_SRCS = [
77 "src/tables/exp2-k-over-64.c",
78 "src/tables/exp2-k-over-2048.c",
79]
80
Marat Dukhan08c4a432019-10-03 09:29:21 -070081SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080082 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080083 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080084 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-avgpool/mp9p8q-scalar.c",
86 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080087 "src/f32-bilinear/gen/scalar-c1.c",
88 "src/f32-bilinear/gen/scalar-c2.c",
89 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070090 "src/f32-clamp/scalar.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080091 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080092 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070093 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070094 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070095 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080096 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080097 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080098 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
99 "src/f32-dwconv/gen/up1x25-scalar.c",
100 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
101 "src/f32-dwconv/gen/up1x4-scalar.c",
102 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
103 "src/f32-dwconv/gen/up1x9-scalar.c",
104 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
105 "src/f32-dwconv/gen/up2x25-scalar.c",
106 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
107 "src/f32-dwconv/gen/up2x4-scalar.c",
108 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
109 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700110 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700111 "src/f32-gavgpool/mp7p7q-scalar.c",
112 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800113 "src/f32-gemm/gen-inc/1x4-scalar.c",
114 "src/f32-gemm/gen-inc/2x4-scalar.c",
115 "src/f32-gemm/gen-inc/4x4-scalar.c",
116 "src/f32-gemm/gen/1x4-scalar.c",
117 "src/f32-gemm/gen/2x4-scalar.c",
118 "src/f32-gemm/gen/4x2-scalar.c",
119 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800120 "src/f32-hswish/gen/scalar-x1.c",
121 "src/f32-hswish/gen/scalar-x2.c",
122 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800123 "src/f32-igemm/gen/1x4-scalar.c",
124 "src/f32-igemm/gen/2x4-scalar.c",
125 "src/f32-igemm/gen/4x2-scalar.c",
126 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800127 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700128 "src/f32-pavgpool/mp9p8q-scalar.c",
129 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800130 "src/f32-ppmm/gen/2x4-scalar.c",
131 "src/f32-ppmm/gen/3x3-scalar.c",
132 "src/f32-ppmm/gen/4x2-scalar.c",
133 "src/f32-ppmm/gen/4x4-scalar.c",
134 "src/f32-prelu/gen/scalar-2x1.c",
135 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800136 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
137 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
138 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
139 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
140 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
141 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
142 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
143 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
144 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
145 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
146 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
147 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700148 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800149 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
150 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
151 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
152 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
153 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
154 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
155 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
156 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
157 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800158 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
159 "src/f32-spmm/gen/1x1-scalar.c",
160 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
161 "src/f32-spmm/gen/2x1-scalar.c",
162 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
163 "src/f32-spmm/gen/4x1-scalar.c",
164 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
165 "src/f32-spmm/gen/8x1-scalar.c",
166 "src/f32-spmm/gen/8x2-scalar.c",
167 "src/f32-spmm/gen/8x4-scalar.c",
168 "src/f32-vbinary/gen/vadd-scalar-x1.c",
169 "src/f32-vbinary/gen/vadd-scalar-x2.c",
170 "src/f32-vbinary/gen/vadd-scalar-x4.c",
171 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
172 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
173 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800174 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
175 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
176 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
177 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
178 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
179 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800180 "src/f32-vbinary/gen/vmax-scalar-x1.c",
181 "src/f32-vbinary/gen/vmax-scalar-x2.c",
182 "src/f32-vbinary/gen/vmax-scalar-x4.c",
183 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
184 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
185 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
186 "src/f32-vbinary/gen/vmin-scalar-x1.c",
187 "src/f32-vbinary/gen/vmin-scalar-x2.c",
188 "src/f32-vbinary/gen/vmin-scalar-x4.c",
189 "src/f32-vbinary/gen/vminc-scalar-x1.c",
190 "src/f32-vbinary/gen/vminc-scalar-x2.c",
191 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800192 "src/f32-vbinary/gen/vmul-scalar-x1.c",
193 "src/f32-vbinary/gen/vmul-scalar-x2.c",
194 "src/f32-vbinary/gen/vmul-scalar-x4.c",
195 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
196 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
197 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800198 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
199 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
200 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
202 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
203 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
204 "src/f32-vbinary/gen/vsub-scalar-x1.c",
205 "src/f32-vbinary/gen/vsub-scalar-x2.c",
206 "src/f32-vbinary/gen/vsub-scalar-x4.c",
207 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
208 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
209 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
210 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
211 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
212 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800213 "src/math/expminus-scalar-lut2048-p1.c",
214 "src/math/expminus-scalar-lut64-p2.c",
215 "src/math/expminus-scalar-p5.c",
216 "src/math/sigmoid-scalar-lut2048-p1-div.c",
217 "src/math/sigmoid-scalar-lut64-p2-div.c",
218 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700219 "src/q8-avgpool/mp9p8q-scalar.c",
220 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700221 "src/q8-dwconv/up1x9-scalar.c",
222 "src/q8-gavgpool/mp7p7q-scalar.c",
223 "src/q8-gavgpool/up7-scalar.c",
224 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800225 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700226 "src/q8-vadd/scalar.c",
227 "src/u8-clamp/scalar.c",
228 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800229 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/u8-rmax/scalar.c",
231 "src/x32-packx/x2-scalar.c",
232 "src/x32-packx/x3-scalar.c",
233 "src/x32-packx/x4-scalar.c",
234 "src/x32-pad/x2-scalar.c",
235 "src/x32-unpool/scalar.c",
236 "src/x32-zip/x2-scalar.c",
237 "src/x32-zip/x3-scalar.c",
238 "src/x32-zip/x4-scalar.c",
239 "src/x32-zip/xm-scalar.c",
240 "src/x8-lut/scalar.c",
241 "src/x8-zip/x2-scalar.c",
242 "src/x8-zip/x3-scalar.c",
243 "src/x8-zip/x4-scalar.c",
244 "src/x8-zip/xm-scalar.c",
245]
246
Marat Dukhan436ebe62019-12-04 15:10:12 -0800247WASM_UKERNELS = [
248 "src/f32-avgpool/mp9p8q-wasm.c",
249 "src/f32-avgpool/up9-wasm.c",
250 "src/f32-clamp/wasm.c",
251 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
252 "src/f32-dwconv/gen/up1x25-wasm.c",
253 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
254 "src/f32-dwconv/gen/up1x4-wasm.c",
255 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
256 "src/f32-dwconv/gen/up1x9-wasm.c",
257 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
258 "src/f32-dwconv/gen/up2x25-wasm.c",
259 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
260 "src/f32-dwconv/gen/up2x4-wasm.c",
261 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
262 "src/f32-dwconv/gen/up2x9-wasm.c",
263 "src/f32-gavgpool/mp7p7q-wasm.c",
264 "src/f32-gavgpool/up7-wasm.c",
265 "src/f32-gemm/gen-inc/1x4-wasm.c",
266 "src/f32-gemm/gen-inc/2x4-wasm.c",
267 "src/f32-gemm/gen-inc/4x4-wasm.c",
268 "src/f32-gemm/gen/1x4-wasm.c",
269 "src/f32-gemm/gen/2x4-wasm.c",
270 "src/f32-gemm/gen/4x2-wasm.c",
271 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800272 "src/f32-hswish/gen/wasm-x1.c",
273 "src/f32-hswish/gen/wasm-x2.c",
274 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800275 "src/f32-igemm/gen/1x4-wasm.c",
276 "src/f32-igemm/gen/2x4-wasm.c",
277 "src/f32-igemm/gen/4x2-wasm.c",
278 "src/f32-igemm/gen/4x4-wasm.c",
279 "src/f32-maxpool/9p8x-wasm-c1.c",
280 "src/f32-pavgpool/mp9p8q-wasm.c",
281 "src/f32-pavgpool/up9-wasm.c",
282 "src/f32-prelu/gen/wasm-2x1.c",
283 "src/f32-prelu/gen/wasm-2x4.c",
284 "src/f32-vbinary/gen/vadd-wasm-x1.c",
285 "src/f32-vbinary/gen/vadd-wasm-x2.c",
286 "src/f32-vbinary/gen/vadd-wasm-x4.c",
287 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
288 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
289 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800290 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
291 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
292 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
293 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
294 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
295 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800296 "src/f32-vbinary/gen/vmax-wasm-x1.c",
297 "src/f32-vbinary/gen/vmax-wasm-x2.c",
298 "src/f32-vbinary/gen/vmax-wasm-x4.c",
299 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
300 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
301 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
302 "src/f32-vbinary/gen/vmin-wasm-x1.c",
303 "src/f32-vbinary/gen/vmin-wasm-x2.c",
304 "src/f32-vbinary/gen/vmin-wasm-x4.c",
305 "src/f32-vbinary/gen/vminc-wasm-x1.c",
306 "src/f32-vbinary/gen/vminc-wasm-x2.c",
307 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800308 "src/f32-vbinary/gen/vmul-wasm-x1.c",
309 "src/f32-vbinary/gen/vmul-wasm-x2.c",
310 "src/f32-vbinary/gen/vmul-wasm-x4.c",
311 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
312 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
313 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800314 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
315 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
316 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800317 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
318 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
319 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
320 "src/f32-vbinary/gen/vsub-wasm-x1.c",
321 "src/f32-vbinary/gen/vsub-wasm-x2.c",
322 "src/f32-vbinary/gen/vsub-wasm-x4.c",
323 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
324 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
325 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
326 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
327 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
328 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
329]
330
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800331PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800332 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800333 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800334 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700335 "src/f32-avgpool/mp9p8q-psimd.c",
336 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800337 "src/f32-bilinear/gen/psimd-c4.c",
338 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700339 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800340 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
341 "src/f32-dwconv/gen/up4x25-psimd.c",
342 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
343 "src/f32-dwconv/gen/up4x4-psimd.c",
344 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
345 "src/f32-dwconv/gen/up4x9-psimd.c",
346 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
347 "src/f32-dwconv/gen/up8x25-psimd.c",
348 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
349 "src/f32-dwconv/gen/up8x4-psimd.c",
350 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
351 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700352 "src/f32-gavgpool/mp7p7q-psimd.c",
353 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800354 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
355 "src/f32-gemm/gen/1x8-psimd-splat.c",
356 "src/f32-gemm/gen/1x8s4-psimd.c",
357 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
358 "src/f32-gemm/gen/4x8-psimd-splat.c",
359 "src/f32-gemm/gen/4x8s4-psimd.c",
360 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
361 "src/f32-gemm/gen/6x8-psimd-splat.c",
362 "src/f32-gemm/gen/6x8s4-psimd.c",
363 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
364 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
365 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
366 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
367 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
368 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
369 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
370 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
371 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800372 "src/f32-hswish/gen/psimd-x4.c",
373 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800374 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
375 "src/f32-igemm/gen/1x8-psimd-splat.c",
376 "src/f32-igemm/gen/1x8s4-psimd.c",
377 "src/f32-igemm/gen/4x2c4-psimd.c",
378 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
379 "src/f32-igemm/gen/4x8-psimd-splat.c",
380 "src/f32-igemm/gen/4x8s4-psimd.c",
381 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
382 "src/f32-igemm/gen/6x8-psimd-splat.c",
383 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800384 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700385 "src/f32-pavgpool/mp9p8q-psimd.c",
386 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800387 "src/f32-ppmm/gen/4x8-psimd.c",
388 "src/f32-prelu/gen/psimd-2x4.c",
389 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800390 "src/f32-rmax/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800391 "src/f32-vbinary/gen/vadd-psimd-x4.c",
392 "src/f32-vbinary/gen/vadd-psimd-x8.c",
393 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
394 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800395 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
396 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
397 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
398 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800399 "src/f32-vbinary/gen/vmax-psimd-x4.c",
400 "src/f32-vbinary/gen/vmax-psimd-x8.c",
401 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
402 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
403 "src/f32-vbinary/gen/vmin-psimd-x4.c",
404 "src/f32-vbinary/gen/vmin-psimd-x8.c",
405 "src/f32-vbinary/gen/vminc-psimd-x4.c",
406 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800407 "src/f32-vbinary/gen/vmul-psimd-x4.c",
408 "src/f32-vbinary/gen/vmul-psimd-x8.c",
409 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
410 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800411 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
412 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800413 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
414 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
415 "src/f32-vbinary/gen/vsub-psimd-x4.c",
416 "src/f32-vbinary/gen/vsub-psimd-x8.c",
417 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
418 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
419 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
420 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421 "src/x32-packx/x4-psimd.c",
422 "src/x32-pad/x2-psimd.c",
423 "src/x32-unpool/psimd.c",
424 "src/x32-zip/x2-psimd.c",
425 "src/x32-zip/x3-psimd.c",
426 "src/x32-zip/x4-psimd.c",
427 "src/x32-zip/xm-psimd.c",
428]
429
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800430PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800431 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
432 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
433 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
434 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
435 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
436 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
437 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
438 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
439 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
440 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
441 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
442 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800443 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
444 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
445 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
446 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
447 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
448 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
449 "src/math/sigmoid-psimd-p5-div.c",
450]
451
Marat Dukhan08c4a432019-10-03 09:29:21 -0700452# ISA-specific micro-kernels
453NEON_UKERNELS = [
454 "src/f32-avgpool/mp9p8q-neon.c",
455 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800456 "src/f32-bilinear/gen/neon-c4.c",
457 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800459 "src/f32-dwconv/gen/up4x9-neon.c",
460 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
461 "src/f32-dwconv/gen/up8x9-neon.c",
462 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/f32-gavgpool-spchw/neon-x4.c",
464 "src/f32-gavgpool/mp7p7q-neon.c",
465 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800466 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
467 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
468 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
469 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
470 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
471 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800472 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800473 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
474 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
475 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
476 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800477 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800478 "src/f32-gemm/gen/1x8s4-neon.c",
479 "src/f32-gemm/gen/4x8s4-neon.c",
480 "src/f32-gemm/gen/6x8s4-neon.c",
481 "src/f32-gemm/gen/8x8s4-neon.c",
482 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
483 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
484 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
485 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
486 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800487 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800488 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
489 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
490 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
491 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800492 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800493 "src/f32-gemm/gen-inc/1x8s4-neon.c",
494 "src/f32-gemm/gen-inc/4x8s4-neon.c",
495 "src/f32-gemm/gen-inc/6x8s4-neon.c",
496 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800497 "src/f32-hswish/gen/neon-x4.c",
498 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800499 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
500 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
501 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
502 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
503 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
504 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800505 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800506 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
507 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
508 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
509 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800510 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800511 "src/f32-igemm/gen/1x8s4-neon.c",
512 "src/f32-igemm/gen/4x8s4-neon.c",
513 "src/f32-igemm/gen/6x8s4-neon.c",
514 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/f32-pavgpool/mp9p8q-neon.c",
516 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800517 "src/f32-ppmm/gen/4x8-neon.c",
518 "src/f32-ppmm/gen/8x8-neon.c",
519 "src/f32-prelu/gen/neon-2x4.c",
520 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800521 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
522 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
523 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
524 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
525 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
526 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
527 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
528 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
529 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
530 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
531 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
532 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
533 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
534 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
535 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
536 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
537 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
538 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
539 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
540 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
541 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
542 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
543 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
544 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800546 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800547 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
548 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
549 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
550 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
551 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
552 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
553 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
554 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
555 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
556 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
557 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
558 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
559 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
560 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
561 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
562 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
563 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
564 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800565 "src/f32-vbinary/gen/vadd-neon-x4.c",
566 "src/f32-vbinary/gen/vadd-neon-x8.c",
567 "src/f32-vbinary/gen/vaddc-neon-x4.c",
568 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800569 "src/f32-vbinary/gen/vmax-neon-x4.c",
570 "src/f32-vbinary/gen/vmax-neon-x8.c",
571 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
572 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
573 "src/f32-vbinary/gen/vmin-neon-x4.c",
574 "src/f32-vbinary/gen/vmin-neon-x8.c",
575 "src/f32-vbinary/gen/vminc-neon-x4.c",
576 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800577 "src/f32-vbinary/gen/vmul-neon-x4.c",
578 "src/f32-vbinary/gen/vmul-neon-x8.c",
579 "src/f32-vbinary/gen/vmulc-neon-x4.c",
580 "src/f32-vbinary/gen/vmulc-neon-x8.c",
581 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
582 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
583 "src/f32-vbinary/gen/vsub-neon-x4.c",
584 "src/f32-vbinary/gen/vsub-neon-x8.c",
585 "src/f32-vbinary/gen/vsubc-neon-x4.c",
586 "src/f32-vbinary/gen/vsubc-neon-x8.c",
587 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
588 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700589 "src/q8-avgpool/mp9p8q-neon.c",
590 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/q8-dwconv/up8x9-neon.c",
592 "src/q8-gavgpool/mp7p7q-neon.c",
593 "src/q8-gavgpool/up7-neon.c",
594 "src/q8-gemm/4x8-neon.c",
595 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800596 "src/q8-igemm/4x8-neon.c",
597 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700598 "src/q8-vadd/neon.c",
599 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800600 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700601 "src/u8-rmax/neon.c",
602 "src/x32-packx/x4-neon-st4.c",
603 "src/x32-pad/x2-neon.c",
604 "src/x32-zip/x2-neon.c",
605 "src/x32-zip/x3-neon.c",
606 "src/x32-zip/x4-neon.c",
607 "src/x32-zip/xm-neon.c",
608 "src/x8-zip/x2-neon.c",
609 "src/x8-zip/x3-neon.c",
610 "src/x8-zip/x4-neon.c",
611 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800612 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800613 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
614 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
615 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
616 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
617 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
618 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700619]
620
621NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800622 "src/f32-bilinear/gen/neonfma-c4.c",
623 "src/f32-bilinear/gen/neonfma-c8.c",
624 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
625 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
626 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
627 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800628 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800629 "src/f32-igemm/gen/1x8s4-neonfma.c",
630 "src/f32-igemm/gen/4x8s4-neonfma.c",
631 "src/f32-igemm/gen/6x8s4-neonfma.c",
632 "src/f32-igemm/gen/8x8s4-neonfma.c",
633 "src/f32-dwconv/gen/up4x9-neonfma.c",
634 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
635 "src/f32-dwconv/gen/up8x9-neonfma.c",
636 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
637 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
638 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
639 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
640 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800641 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800642 "src/f32-gemm/gen/1x8s4-neonfma.c",
643 "src/f32-gemm/gen/4x8s4-neonfma.c",
644 "src/f32-gemm/gen/6x8s4-neonfma.c",
645 "src/f32-gemm/gen/8x8s4-neonfma.c",
646 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
647 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
648 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
649 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800650 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800651 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
652 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
653 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
654 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800655 "src/f32-hswish/gen/neonfma-x4.c",
656 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800657 "src/f32-ppmm/gen/4x8-neonfma.c",
658 "src/f32-ppmm/gen/8x8-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800659 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
660 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
661 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
662 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
663 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
664 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
665 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
666 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
667 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
668 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
669 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
670 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
671 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
672 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
673 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
674 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
675 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
676 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
677 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
678 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
679 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
680 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
681 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
682 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800683 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
684 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
685 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
686 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
687 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
688 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
689 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
690 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
691 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
692 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
693 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
694 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
695 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
696 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
697 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
698 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
699 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
700 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
701 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
702 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
703 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
704 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
705 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
706 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
707 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
708 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
709 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
710 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
711 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
712 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
713 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
714 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
715 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
716 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
718 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
719 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
720 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
721 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
722 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
723 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
724 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
725 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
726 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
727 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
728 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
729 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
730 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
731 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
732 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
733 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
734 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
735 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
736 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800737 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
738 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800739 "src/math/exp-neonfma-lut64-p2.c",
740 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800741 "src/math/expminus-neonfma-lut2048-p1.c",
742 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800743 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800744 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
745 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
746 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
747 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
748 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
749 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
750 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
751 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
752 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
753 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
754 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
755 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
756 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
757 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
758 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
759 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
760 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
761 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700762]
763
764AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800765 "src/f32-vbinary/gen/vdiv-neon-x4.c",
766 "src/f32-vbinary/gen/vdiv-neon-x8.c",
767 "src/f32-vbinary/gen/vdivc-neon-x4.c",
768 "src/f32-vbinary/gen/vdivc-neon-x8.c",
769 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
770 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800771 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
772 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
773 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
774 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
775 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
776 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800777 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800778 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
779 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
780 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
781 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
782 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800783 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800784 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
785 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
786 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
787 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
788 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
789 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800790 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
792 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
793 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
794 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
795 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
796 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
797 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800798 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
799 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
800 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
801 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
802 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
803 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
804 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
805 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
806 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
807 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
808 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
809 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
810 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
811 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
812 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
813 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
814 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
815 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800816 "src/f32-spmm/gen/12x1-neonfma.c",
817 "src/f32-spmm/gen/12x2-neonfma.c",
818 "src/f32-spmm/gen/12x4-neonfma.c",
819 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
820 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
821 "src/f32-spmm/gen/16x1-neonfma.c",
822 "src/f32-spmm/gen/16x2-neonfma.c",
823 "src/f32-spmm/gen/16x4-neonfma.c",
824 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
825 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
826 "src/f32-spmm/gen/4x1-neonfma.c",
827 "src/f32-spmm/gen/4x2-neonfma.c",
828 "src/f32-spmm/gen/4x4-neonfma.c",
829 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
830 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
831 "src/f32-spmm/gen/8x1-neonfma.c",
832 "src/f32-spmm/gen/8x2-neonfma.c",
833 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800834 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
835 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
836 "src/math/sigmoid-neonfma-rr1-p5-div.c",
837 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
838 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
839 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700840]
841
842AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800843 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
844 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
845 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhanbdb56f52020-02-05 21:42:49 -0800846 "src/f16-spmm/gen/8x1-neonfp16arith.c",
847 "src/f16-spmm/gen/8x1-neonfp16arith-unroll2.c",
848 "src/f16-spmm/gen/16x1-neonfp16arith.c",
849 "src/f16-spmm/gen/16x1-neonfp16arith-unroll2.c",
850 "src/f16-spmm/gen/24x1-neonfp16arith.c",
851 "src/f16-spmm/gen/24x1-neonfp16arith-unroll2.c",
852 "src/f16-spmm/gen/32x1-neonfp16arith.c",
853 "src/f16-spmm/gen/32x1-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700854]
855
856SSE_UKERNELS = [
857 "src/f32-avgpool/mp9p8q-sse.c",
858 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800859 "src/f32-bilinear/gen/sse-c4.c",
860 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700861 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800862 "src/f32-dwconv-spchw/3x3p1-sse.c",
863 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800864 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
865 "src/f32-dwconv/gen/up4x25-sse.c",
866 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
867 "src/f32-dwconv/gen/up4x4-sse.c",
868 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
869 "src/f32-dwconv/gen/up4x9-sse.c",
870 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
871 "src/f32-dwconv/gen/up8x25-sse.c",
872 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
873 "src/f32-dwconv/gen/up8x4-sse.c",
874 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
875 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876 "src/f32-gavgpool-spchw/sse-x4.c",
877 "src/f32-gavgpool/mp7p7q-sse.c",
878 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800879 "src/f32-gemm/gen/1x8-sse-dup.c",
880 "src/f32-gemm/gen/1x8-sse-load1.c",
881 "src/f32-gemm/gen/1x8s4-sse.c",
882 "src/f32-gemm/gen/4x8-sse-dup.c",
883 "src/f32-gemm/gen/4x8-sse-load1.c",
884 "src/f32-gemm/gen/4x8s4-sse.c",
885 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
886 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
887 "src/f32-gemm/gen-inc/1x8s4-sse.c",
888 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
889 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
890 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800891 "src/f32-hswish/gen/sse-x4.c",
892 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800893 "src/f32-igemm/gen/1x8-sse-dup.c",
894 "src/f32-igemm/gen/1x8-sse-load1.c",
895 "src/f32-igemm/gen/1x8s4-sse.c",
896 "src/f32-igemm/gen/4x2c4-sse.c",
897 "src/f32-igemm/gen/4x8-sse-dup.c",
898 "src/f32-igemm/gen/4x8-sse-load1.c",
899 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800900 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700901 "src/f32-pavgpool/mp9p8q-sse.c",
902 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800903 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700904 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800905 "src/f32-spmm/gen/4x1-sse.c",
906 "src/f32-spmm/gen/8x1-sse.c",
907 "src/f32-vbinary/gen/vadd-sse-x4.c",
908 "src/f32-vbinary/gen/vadd-sse-x8.c",
909 "src/f32-vbinary/gen/vaddc-sse-x4.c",
910 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800911 "src/f32-vbinary/gen/vdiv-sse-x4.c",
912 "src/f32-vbinary/gen/vdiv-sse-x8.c",
913 "src/f32-vbinary/gen/vdivc-sse-x4.c",
914 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800915 "src/f32-vbinary/gen/vmax-sse-x4.c",
916 "src/f32-vbinary/gen/vmax-sse-x8.c",
917 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
918 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
919 "src/f32-vbinary/gen/vmin-sse-x4.c",
920 "src/f32-vbinary/gen/vmin-sse-x8.c",
921 "src/f32-vbinary/gen/vminc-sse-x4.c",
922 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800923 "src/f32-vbinary/gen/vmul-sse-x4.c",
924 "src/f32-vbinary/gen/vmul-sse-x8.c",
925 "src/f32-vbinary/gen/vmulc-sse-x4.c",
926 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800927 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
928 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800929 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
930 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
931 "src/f32-vbinary/gen/vsub-sse-x4.c",
932 "src/f32-vbinary/gen/vsub-sse-x8.c",
933 "src/f32-vbinary/gen/vsubc-sse-x4.c",
934 "src/f32-vbinary/gen/vsubc-sse-x8.c",
935 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
936 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937 "src/x32-packx/x4-sse.c",
938]
939
940SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800941 "src/f32-argmaxpool/9p8x-sse2-c4.c",
942 "src/f32-argmaxpool/4x-sse2-c4.c",
943 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800944 "src/f32-prelu/gen/sse2-2x4.c",
945 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800946 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
947 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
948 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
949 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
950 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
951 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
952 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
953 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
954 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
955 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
956 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
957 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800958 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800959 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800960 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800961 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800962 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
963 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700964 "src/q8-avgpool/mp9p8q-sse2.c",
965 "src/q8-avgpool/up9-sse2.c",
966 "src/q8-igemm/4x4c2-sse2.c",
967 "src/q8-dwconv/up8x9-sse2.c",
968 "src/q8-gavgpool/mp7p7q-sse2.c",
969 "src/q8-gavgpool/up7-sse2.c",
970 "src/q8-gemm/2x4c8-sse2.c",
971 "src/q8-gemm/4x4c2-sse2.c",
972 "src/q8-vadd/sse2.c",
973 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800974 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700975 "src/u8-rmax/sse2.c",
976 "src/x32-pad/x2-sse2.c",
977 "src/x32-zip/x2-sse2.c",
978 "src/x32-zip/x3-sse2.c",
979 "src/x32-zip/x4-sse2.c",
980 "src/x32-zip/xm-sse2.c",
981 "src/x8-zip/x2-sse2.c",
982 "src/x8-zip/x3-sse2.c",
983 "src/x8-zip/x4-sse2.c",
984 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800985 "src/math/exp-sse2-p5.c",
986 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800987 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988]
989
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800990SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800991 "src/f32-prelu/gen/sse41-2x4.c",
992 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800993 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800994 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800995 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800996 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800997 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
998 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800999]
1000
Marat Dukhan08c4a432019-10-03 09:29:21 -07001001AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001002 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001003 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
1004 "src/f32-dwconv/gen/up16x4-avx.c",
1005 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
1006 "src/f32-dwconv/gen/up8x4-avx.c",
1007 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
1008 "src/f32-dwconv/gen/up16x9-avx.c",
1009 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
1010 "src/f32-dwconv/gen/up8x9-avx.c",
1011 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
1012 "src/f32-dwconv/gen/up16x25-avx.c",
1013 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
1014 "src/f32-dwconv/gen/up8x25-avx.c",
1015 "src/f32-gemm/gen/1x8-avx-broadcast.c",
1016 "src/f32-gemm/gen/4x8-avx-broadcast.c",
1017 "src/f32-gemm/gen/5x8-avx-broadcast.c",
1018 "src/f32-gemm/gen/6x8-avx-broadcast.c",
1019 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001020 "src/f32-gemm/gen/1x16-avx-broadcast.c",
1021 "src/f32-gemm/gen/3x16-avx-broadcast.c",
1022 "src/f32-gemm/gen/4x16-avx-broadcast.c",
1023 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001024 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
1025 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
1026 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
1027 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
1028 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001029 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
1030 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
1031 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
1032 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001033 "src/f32-hswish/gen/avx-x8.c",
1034 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001035 "src/f32-igemm/gen/1x8-avx-broadcast.c",
1036 "src/f32-igemm/gen/4x8-avx-broadcast.c",
1037 "src/f32-igemm/gen/5x8-avx-broadcast.c",
1038 "src/f32-igemm/gen/6x8-avx-broadcast.c",
1039 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001040 "src/f32-igemm/gen/1x16-avx-broadcast.c",
1041 "src/f32-igemm/gen/3x16-avx-broadcast.c",
1042 "src/f32-igemm/gen/4x16-avx-broadcast.c",
1043 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001044 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001045 "src/f32-vbinary/gen/vadd-avx-x8.c",
1046 "src/f32-vbinary/gen/vadd-avx-x16.c",
1047 "src/f32-vbinary/gen/vaddc-avx-x8.c",
1048 "src/f32-vbinary/gen/vaddc-avx-x16.c",
1049 "src/f32-vbinary/gen/vdiv-avx-x8.c",
1050 "src/f32-vbinary/gen/vdiv-avx-x16.c",
1051 "src/f32-vbinary/gen/vdivc-avx-x8.c",
1052 "src/f32-vbinary/gen/vdivc-avx-x16.c",
1053 "src/f32-vbinary/gen/vmax-avx-x8.c",
1054 "src/f32-vbinary/gen/vmax-avx-x16.c",
1055 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1056 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1057 "src/f32-vbinary/gen/vmin-avx-x8.c",
1058 "src/f32-vbinary/gen/vmin-avx-x16.c",
1059 "src/f32-vbinary/gen/vminc-avx-x8.c",
1060 "src/f32-vbinary/gen/vminc-avx-x16.c",
1061 "src/f32-vbinary/gen/vmul-avx-x8.c",
1062 "src/f32-vbinary/gen/vmul-avx-x16.c",
1063 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1064 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1065 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1066 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1067 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1068 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1069 "src/f32-vbinary/gen/vsub-avx-x8.c",
1070 "src/f32-vbinary/gen/vsub-avx-x16.c",
1071 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1072 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001073 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001074]
1075
Marat Dukhanfda12b82019-11-21 12:27:59 -08001076FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001077 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
1078 "src/f32-dwconv/gen/up16x4-fma3.c",
1079 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
1080 "src/f32-dwconv/gen/up8x4-fma3.c",
1081 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
1082 "src/f32-dwconv/gen/up16x9-fma3.c",
1083 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
1084 "src/f32-dwconv/gen/up8x9-fma3.c",
1085 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
1086 "src/f32-dwconv/gen/up16x25-fma3.c",
1087 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
1088 "src/f32-dwconv/gen/up8x25-fma3.c",
1089 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
1090 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
1091 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
1092 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
1093 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
1094 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001095 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
1096 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
1097 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
1098 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001099 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
1100 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
1101 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
1102 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001103 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
1104 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
1105 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
1106 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
1107 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
1108 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001109 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
1110 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
1111 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
1112 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001113 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
1114 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
1115 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
1116 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001117 "src/f32-hswish/gen/fma3-x8.c",
1118 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001119 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
1120 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
1121 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
1122 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
1123 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
1124 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001125 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
1126 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
1127 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
1128 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001129 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
1130 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
1131 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
1132 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001133]
1134
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001135AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001136 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1137 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1138 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1139 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1140 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1141 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1142 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1143 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1144 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1145 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1146 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1147 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1148 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1149 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1150 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1151 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1152 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1153 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1154 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1155 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1156 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1157 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1158 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1159 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1160 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1161 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1162 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1163 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1164 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1165 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1166 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1167 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1168 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1169 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1170 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1171 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001172 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1173 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1174 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1175 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1176 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1177 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1178 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1179 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1180 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1181 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1182 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1185 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1186 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1187 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1188 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1189 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1190 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1191 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1192 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1193 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1194 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1195 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1196 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1197 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1198 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1199 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1200 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1201 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001202 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1203 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1204 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1205 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1206 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1207 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1208 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1209 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1210 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1211 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1212 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1213 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1214 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1215 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1216 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1217 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1218 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1219 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1220 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1221 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1222 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1223 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1224 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1225 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001226 "src/math/exp-avx2-p5.c",
1227 "src/math/exp-avx2-perm-p3.c",
1228 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001229 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001230 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001231 "src/math/sigmoid-avx2-rr2-p5-div.c",
1232 "src/math/sigmoid-avx2-rr1-p5-div.c",
1233 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1234 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1235 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1236 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001237]
1238
Marat Dukhan08c4a432019-10-03 09:29:21 -07001239AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001240 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -08001241 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
1242 "src/f32-dwconv/gen/up32x4-avx512f.c",
1243 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
1244 "src/f32-dwconv/gen/up16x4-avx512f.c",
1245 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
1246 "src/f32-dwconv/gen/up32x9-avx512f.c",
1247 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
1248 "src/f32-dwconv/gen/up16x9-avx512f.c",
1249 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
1250 "src/f32-dwconv/gen/up32x25-avx512f.c",
1251 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
1252 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001253 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
1254 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
1255 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
1256 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
1257 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
1258 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
1259 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
1260 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
1261 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
1262 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
1263 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
1264 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001265 "src/f32-hswish/gen/avx512f-x16.c",
1266 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001267 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
1268 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
1269 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
1270 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
1271 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
1272 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001273 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1274 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1275 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1276 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1277 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1278 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1279 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1280 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1281 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1282 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1283 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1284 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1285 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1286 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1287 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1288 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1289 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1290 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1291 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1292 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1293 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1294 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1295 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1296 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1297 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1298 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1299 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1300 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1301 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1302 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1303 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1304 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1305 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1306 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1307 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1308 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001309 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001310 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1311 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1312 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1313 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1314 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1315 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1316 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1317 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1318 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1319 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1320 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1321 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1322 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1323 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1324 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1325 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1326 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1327 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1328 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1329 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1330 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1331 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1332 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1333 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1334 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1335 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1336 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1337 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001338 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001339 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1340 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1341 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1342 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1343 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1344 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1345 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1346 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1347 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1348 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1349 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1350 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1351 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1352 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1353 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1354 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1355 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1356 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1357 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1358 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1359 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1360 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1361 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1362 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001363 "src/math/exp-avx512f-p5-scalef.c",
1364 "src/math/exp-avx512f-p5.c",
1365 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001366 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001367 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368]
1369
1370AARCH32_ASM_UKERNELS = [
1371 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001372 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001373 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1374 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001375 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Miao Wang3fa1f012020-02-17 22:45:06 +00001376 "src/f32-igemm/4x8-aarch32-neon-ld64.S",
1377 "src/f32-igemm/gen/4x8-aarch32-neon-cortex-a75.S",
1378 "src/f32-igemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001379]
1380
1381AARCH64_ASM_UKERNELS = [
1382 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1383 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001384 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1385 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1386 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1387 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1388 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1389 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1390 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1391 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1392 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1393 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001394 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001395 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1396 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001397 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001398 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001399 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1400 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1401 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1402 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1403 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1404 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1405 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1406 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1407 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1408 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1409 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1410 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1411 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001412 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001413 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1414 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001415 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001416 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001417 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1418 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1419 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001420 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001421 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001422 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1423 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001424 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001425 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001426 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1427 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1428 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1429 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001430 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001431 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001432 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1433 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001434]
1435
1436INTERNAL_MICROKERNEL_HDRS = [
1437 "src/xnnpack/argmaxpool.h",
1438 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001439 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001440 "src/xnnpack/clamp.h",
1441 "src/xnnpack/common.h",
1442 "src/xnnpack/conv.h",
1443 "src/xnnpack/dwconv.h",
1444 "src/xnnpack/gavgpool.h",
1445 "src/xnnpack/gemm.h",
1446 "src/xnnpack/hswish.h",
1447 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001448 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001449 "src/xnnpack/lut.h",
1450 "src/xnnpack/math.h",
1451 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001452 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001453 "src/xnnpack/packx.h",
1454 "src/xnnpack/pad.h",
1455 "src/xnnpack/params.h",
1456 "src/xnnpack/pavgpool.h",
1457 "src/xnnpack/ppmm.h",
1458 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001459 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001460 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001461 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462 "src/xnnpack/rmax.h",
1463 "src/xnnpack/scalar-utils.h",
1464 "src/xnnpack/spmm.h",
1465 "src/xnnpack/unpool.h",
1466 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001467 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001468 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001469 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001470 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001471 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001472 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001473 "src/xnnpack/zip.h",
1474]
1475
1476INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1477 "include/xnnpack.h",
1478 "src/xnnpack/allocator.h",
1479 "src/xnnpack/compute.h",
1480 "src/xnnpack/im2col.h",
1481 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001482 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001483 "src/xnnpack/operator.h",
1484 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001485 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001486 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001487 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001488 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001489]
1490
1491ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1492 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001493]
1494
1495MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001496 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001497 "include/xnnpack.h",
1498]
1499
1500MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1501 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001502 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001503 "src/xnnpack/requantization.h",
1504 "include/xnnpack.h",
1505]
1506
1507OPERATOR_TEST_PARAMS_HDRS = [
1508 "src/xnnpack/params.h",
1509 "src/xnnpack/common.h",
1510]
1511
1512WEIGHTS_PACK_HDRS = [
1513 "src/xnnpack/pack.h",
1514 "src/xnnpack/operator.h",
1515 "src/xnnpack/compute.h",
1516]
1517
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001518LOGGING_COPTS = select({
1519 # No logging in optimized mode
1520 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1521 # Full logging in debug mode
1522 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1523 # Error-only logging in default (fastbuild) mode
1524 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1525})
1526
1527LOGGING_HDRS = [
1528 "src/xnnpack/log.h",
1529]
1530
Marat Dukhan08c4a432019-10-03 09:29:21 -07001531xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001532 name = "tables",
1533 srcs = TABLE_SRCS,
1534 hdrs = INTERNAL_HDRS,
1535 copts = xnnpack_std_copts(),
1536)
1537
1538xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001539 name = "scalar_ukernels",
1540 srcs = SCALAR_UKERNELS,
1541 hdrs = INTERNAL_HDRS,
1542 aarch32_copts = ["-marm"],
1543 copts = xnnpack_std_copts(),
1544 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001545 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001546 "@FP16",
1547 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001548 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001549 ],
1550)
1551
1552xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001553 name = "wasm_ukernels",
1554 hdrs = INTERNAL_HDRS,
1555 copts = xnnpack_std_copts(),
1556 wasm_srcs = WASM_UKERNELS,
1557 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001558 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001559 "@FP16",
1560 "@FXdiv",
1561 "@pthreadpool",
1562 ],
1563)
1564
1565xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001566 name = "psimd_fastmath_ukernels",
1567 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001568 hdrs = INTERNAL_HDRS,
1569 aarch32_copts = [
1570 "-marm",
1571 "-mfpu=neon",
1572 ],
1573 copts = xnnpack_std_copts(),
1574 optimized_copts = [
1575 "-O3",
1576 "-ffast-math",
1577 ],
1578 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001579 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580 "@FP16",
1581 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001582 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001583 ],
1584)
1585
1586xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001587 name = "psimd_accmath_ukernels",
1588 srcs = PSIMD_ACCMATH_UKERNELS,
1589 hdrs = INTERNAL_HDRS,
1590 aarch32_copts = [
1591 "-marm",
1592 "-mfpu=neon",
1593 ],
1594 copts = xnnpack_std_copts(),
1595 optimized_copts = [
1596 "-O3",
1597 ],
1598 deps = [
1599 ":tables",
1600 "@FP16",
1601 "@psimd",
1602 "@pthreadpool",
1603 ],
1604)
1605
1606xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001607 name = "neon_ukernels",
1608 hdrs = INTERNAL_HDRS,
1609 aarch32_copts = [
1610 "-marm",
1611 "-mfpu=neon",
1612 ],
1613 aarch32_srcs = NEON_UKERNELS,
1614 aarch64_srcs = NEON_UKERNELS,
1615 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001616 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001617 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001618 "@FP16",
1619 "@pthreadpool",
1620 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001621)
1622
1623xnnpack_cc_library(
1624 name = "neonfma_ukernels",
1625 hdrs = INTERNAL_HDRS,
1626 aarch32_copts = [
1627 "-marm",
1628 "-mfpu=neon-vfpv4",
1629 ],
1630 aarch32_srcs = NEONFMA_UKERNELS,
1631 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1632 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001633 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001634 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001635 "@FP16",
1636 "@pthreadpool",
1637 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638)
1639
1640xnnpack_cc_library(
1641 name = "neonfp16arith_ukernels",
1642 hdrs = INTERNAL_HDRS,
1643 aarch64_copts = ["-march=armv8.2-a+fp16"],
1644 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1645 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001646 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001647 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001648 "@FP16",
1649 "@pthreadpool",
1650 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001651)
1652
1653xnnpack_cc_library(
1654 name = "sse2_ukernels",
1655 hdrs = INTERNAL_HDRS,
1656 copts = xnnpack_std_copts(),
1657 x86_copts = ["-msse2"],
1658 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001659 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001660 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001661 "@FP16",
1662 "@pthreadpool",
1663 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664)
1665
1666xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001667 name = "sse41_ukernels",
1668 hdrs = INTERNAL_HDRS,
1669 copts = xnnpack_std_copts(),
1670 x86_copts = ["-msse4.1"],
1671 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001672 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001673 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001674 "@FP16",
1675 "@pthreadpool",
1676 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001677)
1678
1679xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001680 name = "avx_ukernels",
1681 hdrs = INTERNAL_HDRS,
1682 copts = xnnpack_std_copts(),
1683 x86_copts = ["-mavx"],
1684 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001685 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001686 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001687 "@FP16",
1688 "@pthreadpool",
1689 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001690)
1691
1692xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001693 name = "fma3_ukernels",
1694 hdrs = INTERNAL_HDRS,
1695 copts = xnnpack_std_copts(),
1696 x86_copts = [
1697 "-mfma",
1698 ],
1699 x86_srcs = FMA3_UKERNELS,
1700 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001701 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001702 "@FP16",
1703 "@pthreadpool",
1704 ],
1705)
1706
1707xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001708 name = "avx2_ukernels",
1709 hdrs = INTERNAL_HDRS,
1710 copts = xnnpack_std_copts(),
1711 x86_copts = [
1712 "-mfma",
1713 "-mavx2",
1714 ],
1715 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001716 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001717 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001718 "@FP16",
1719 "@pthreadpool",
1720 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001721)
1722
1723xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001724 name = "avx512f_ukernels",
1725 hdrs = INTERNAL_HDRS,
1726 copts = xnnpack_std_copts(),
1727 x86_copts = ["-mavx512f"],
1728 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001729 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001730 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001731 "@FP16",
1732 "@pthreadpool",
1733 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001734)
1735
1736xnnpack_cc_library(
1737 name = "asm_ukernels",
1738 hdrs = ["src/xnnpack/assembly.h"],
1739 aarch32_srcs = AARCH32_ASM_UKERNELS,
1740 aarch64_srcs = AARCH64_ASM_UKERNELS,
1741)
1742
1743xnnpack_aggregate_library(
1744 name = "ukernels",
1745 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001746 ":psimd_fastmath_ukernels",
1747 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001748 ":neon_ukernels",
1749 ":neonfma_ukernels",
1750 ":asm_ukernels",
1751 ],
1752 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001753 ":psimd_fastmath_ukernels",
1754 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001755 ":neon_ukernels",
1756 ":neonfma_ukernels",
1757 ":neonfp16arith_ukernels",
1758 ":asm_ukernels",
1759 ],
1760 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001761 wasm_deps = [
1762 ":wasm_ukernels",
1763 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001764 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001765 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001766 ":psimd_fastmath_ukernels",
1767 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768 ],
1769 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001770 ":psimd_fastmath_ukernels",
1771 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001772 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001773 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001774 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001775 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001776 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 ":avx512f_ukernels",
1778 ],
1779)
1780
1781xnnpack_cc_library(
1782 name = "im2col",
1783 srcs = ["src/im2col.c"],
1784 hdrs = [
1785 "src/xnnpack/common.h",
1786 "src/xnnpack/im2col.h",
1787 ],
1788 copts = xnnpack_std_copts(),
1789)
1790
1791xnnpack_cc_library(
1792 name = "indirection",
1793 srcs = ["src/indirection.c"],
1794 hdrs = INTERNAL_HDRS,
1795 copts = xnnpack_std_copts(),
1796 deps = [
1797 "@FP16",
1798 "@FXdiv",
1799 "@pthreadpool",
1800 ],
1801)
1802
1803xnnpack_cc_library(
1804 name = "operator_run",
1805 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001806 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1807 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001808 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1809 "-Wno-vla",
1810 ],
1811 deps = [
1812 "@FP16",
1813 "@FXdiv",
1814 "@clog",
1815 "@pthreadpool",
1816 ],
1817)
1818
1819cc_library(
1820 name = "enable_assembly",
1821 defines = select({
1822 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1823 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001824 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825 }),
1826)
1827
Marat Dukhancf056b22019-10-07 10:26:29 -07001828xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001829 name = "operators",
1830 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001831 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001832 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001833 ],
1834 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001835 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001836 "-Isrc",
1837 "-Iinclude",
1838 ] + select({
1839 ":debug_build": [],
1840 "//conditions:default": xnnpack_min_size_copts(),
1841 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001842 wasm_srcs = ["src/wasm-stubs.c"],
1843 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001845 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001846 "@FP16",
1847 "@FXdiv",
1848 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001849 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001850 ],
1851)
1852
1853cc_library(
1854 name = "XNNPACK",
1855 srcs = [
1856 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08001857 "src/runtime.c",
1858 "src/subgraph.c",
1859 "src/tensor.c",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001860 ],
1861 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1862 "-Isrc",
1863 "-Iinclude",
1864 ] + select({
1865 ":debug_build": [],
1866 "//conditions:default": xnnpack_min_size_copts(),
1867 }),
1868 includes = ["include"],
1869 linkstatic = True,
1870 textual_hdrs = ["include/xnnpack.h"],
1871 visibility = xnnpack_visibility(),
1872 deps = [
1873 ":enable_assembly",
1874 ":ukernels",
1875 ":operator_run",
1876 ":operators",
1877 "@clog",
1878 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001879 ] + select({
1880 ":emscripten": [],
1881 "//conditions:default": ["@cpuinfo"],
1882 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001883)
1884
1885cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001886 name = "xnnpack_operators_nhwc_f32",
1887 srcs = [
1888 "src/init.c",
1889 ],
1890 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1891 "-Isrc",
1892 "-Iinclude",
1893 ] + select({
1894 ":debug_build": [],
1895 "//conditions:default": xnnpack_min_size_copts(),
1896 }),
1897 defines = [
1898 "XNN_NO_Q8_OPERATORS",
1899 "XNN_NO_U8_OPERATORS",
1900 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001901 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001902 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001903 includes = ["include"],
1904 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001905 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001906 visibility = xnnpack_visibility(),
1907 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001908 ":enable_assembly",
1909 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001910 ":operator_run",
1911 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001912 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001913 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001914 ] + select({
1915 ":emscripten": [],
1916 "//conditions:default": ["@cpuinfo"],
1917 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918)
1919
Marat Dukhancf056b22019-10-07 10:26:29 -07001920xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001921 name = "bench_utils",
1922 srcs = ["bench/utils.cc"],
1923 hdrs = ["bench/utils.h"],
1924 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001925 deps = [
1926 "@com_google_benchmark//:benchmark",
1927 "@cpuinfo",
1928 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001929)
1930
Frank Barchard7e955972019-10-11 10:34:25 -07001931######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001932
1933xnnpack_benchmark(
1934 name = "q8_gemm_bench",
1935 srcs = [
1936 "bench/gemm.h",
1937 "bench/q8-gemm.cc",
1938 "src/xnnpack/AlignedAllocator.h",
1939 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1940 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1941 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1942)
1943
1944xnnpack_benchmark(
1945 name = "f16_gemm_bench",
1946 srcs = [
1947 "bench/f16-gemm.cc",
1948 "bench/gemm.h",
1949 "src/xnnpack/AlignedAllocator.h",
1950 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1951 copts = ["-Wno-unused-function"],
1952 deps = MICROKERNEL_BENCHMARK_DEPS,
1953)
1954
1955xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08001956 name = "f16_spmm_bench",
1957 srcs = [
1958 "bench/f16-spmm.cc",
1959 "bench/gemm.h",
1960 "bench/google/gemm.h",
1961 "src/xnnpack/AlignedAllocator.h",
1962 ] + MICROKERNEL_BENCHMARK_HDRS,
1963 copts = ["-Wno-unused-function"],
1964 deps = MICROKERNEL_BENCHMARK_DEPS,
1965)
1966
1967xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001968 name = "f32_igemm_bench",
1969 srcs = [
1970 "bench/f32-igemm.cc",
1971 "bench/conv.h",
1972 "src/xnnpack/AlignedAllocator.h",
1973 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001974 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001975)
1976
1977xnnpack_benchmark(
1978 name = "f32_conv_hwc_bench",
1979 srcs = [
1980 "bench/f32-conv-hwc.cc",
1981 "bench/dconv.h",
1982 "src/xnnpack/AlignedAllocator.h",
1983 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1984 copts = ["-Wno-unused-function"],
1985 deps = MICROKERNEL_BENCHMARK_DEPS,
1986)
1987
1988xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001989 name = "f32_conv_hwc2spchw_bench",
1990 srcs = [
1991 "bench/f32-conv-hwc2spchw.cc",
1992 "bench/dconv.h",
1993 "src/xnnpack/AlignedAllocator.h",
1994 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1995 copts = ["-Wno-unused-function"],
1996 deps = MICROKERNEL_BENCHMARK_DEPS,
1997)
1998
1999xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002000 name = "f32_dwconv_bench",
2001 srcs = [
2002 "bench/f32-dwconv.cc",
2003 "bench/dwconv.h",
2004 "src/xnnpack/AlignedAllocator.h",
2005 ] + MICROKERNEL_BENCHMARK_HDRS,
2006 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2007)
2008
2009xnnpack_benchmark(
2010 name = "f32_dwconv_spchw_bench",
2011 srcs = [
2012 "bench/f32-dwconv-spchw.cc",
2013 "bench/dwconv.h",
2014 "src/xnnpack/AlignedAllocator.h",
2015 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2016 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2017)
2018
2019xnnpack_benchmark(
2020 name = "f32_gemm_bench",
2021 srcs = [
2022 "bench/f32-gemm.cc",
2023 "bench/gemm.h",
2024 "src/xnnpack/AlignedAllocator.h",
2025 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2026 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07002027 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002028)
2029
2030xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002031 name = "f32_raddexpminusmax_bench",
2032 srcs = [
2033 "bench/f32-raddexpminusmax.cc",
2034 "src/xnnpack/AlignedAllocator.h",
2035 ] + MICROKERNEL_BENCHMARK_HDRS,
2036 deps = MICROKERNEL_BENCHMARK_DEPS,
2037)
2038
2039xnnpack_benchmark(
2040 name = "f32_raddextexp_bench",
2041 srcs = [
2042 "bench/f32-raddextexp.cc",
2043 "src/xnnpack/AlignedAllocator.h",
2044 ] + MICROKERNEL_BENCHMARK_HDRS,
2045 deps = MICROKERNEL_BENCHMARK_DEPS,
2046)
2047
2048xnnpack_benchmark(
2049 name = "f32_raddstoreexpminusmax_bench",
2050 srcs = [
2051 "bench/f32-raddstoreexpminusmax.cc",
2052 "src/xnnpack/AlignedAllocator.h",
2053 ] + MICROKERNEL_BENCHMARK_HDRS,
2054 deps = MICROKERNEL_BENCHMARK_DEPS,
2055)
2056
2057xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002058 name = "f32_rmax_bench",
2059 srcs = [
2060 "bench/f32-rmax.cc",
2061 "src/xnnpack/AlignedAllocator.h",
2062 ] + MICROKERNEL_BENCHMARK_HDRS,
2063 deps = MICROKERNEL_BENCHMARK_DEPS,
2064)
2065
2066xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08002067 name = "f32_sigmoid_bench",
2068 srcs = [
2069 "bench/f32-sigmoid.cc",
2070 "src/xnnpack/AlignedAllocator.h",
2071 ] + MICROKERNEL_BENCHMARK_HDRS,
2072 copts = ["-Wno-unused-function"],
2073 deps = MICROKERNEL_BENCHMARK_DEPS,
2074)
2075
2076xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002077 name = "f32_spmm_bench",
2078 srcs = [
2079 "bench/f32-spmm.cc",
2080 "bench/gemm.h",
2081 "src/xnnpack/AlignedAllocator.h",
2082 ] + MICROKERNEL_BENCHMARK_HDRS,
2083 copts = ["-Wno-unused-function"],
2084 deps = MICROKERNEL_BENCHMARK_DEPS,
2085)
2086
2087xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002088 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002089 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002090 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002091 ] + MICROKERNEL_BENCHMARK_HDRS,
2092 copts = ["-Wno-unused-function"],
2093 deps = MICROKERNEL_BENCHMARK_DEPS,
2094)
2095
2096xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002097 name = "f32_vscaleexpminusmax_bench",
2098 srcs = [
2099 "bench/f32-vscaleexpminusmax.cc",
2100 "src/xnnpack/AlignedAllocator.h",
2101 ] + MICROKERNEL_BENCHMARK_HDRS,
2102 deps = MICROKERNEL_BENCHMARK_DEPS,
2103)
2104
2105xnnpack_benchmark(
2106 name = "f32_vscaleextexp_bench",
2107 srcs = [
2108 "bench/f32-vscaleextexp.cc",
2109 "src/xnnpack/AlignedAllocator.h",
2110 ] + MICROKERNEL_BENCHMARK_HDRS,
2111 deps = MICROKERNEL_BENCHMARK_DEPS,
2112)
2113
2114xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002115 name = "f32_im2col_gemm_bench",
2116 srcs = [
2117 "bench/f32-im2col-gemm.cc",
2118 "bench/conv.h",
2119 "src/xnnpack/AlignedAllocator.h",
2120 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2121 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2122)
2123
2124########################### Benchmarks for operators ###########################
2125
2126xnnpack_benchmark(
2127 name = "add_bench",
2128 srcs = ["bench/add.cc"],
2129 deps = OPERATOR_BENCHMARK_DEPS,
2130)
2131
2132xnnpack_benchmark(
2133 name = "average_pooling_bench",
2134 srcs = ["bench/average-pooling.cc"],
2135 deps = OPERATOR_BENCHMARK_DEPS,
2136)
2137
2138xnnpack_benchmark(
2139 name = "channel_shuffle_bench",
2140 srcs = ["bench/channel-shuffle.cc"],
2141 deps = OPERATOR_BENCHMARK_DEPS,
2142)
2143
2144xnnpack_benchmark(
2145 name = "convolution_bench",
2146 srcs = ["bench/convolution.cc"],
2147 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2148 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2149)
2150
2151xnnpack_benchmark(
2152 name = "deconvolution_bench",
2153 srcs = ["bench/deconvolution.cc"],
2154 copts = xnnpack_optional_tflite_copts(),
2155 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2156)
2157
2158xnnpack_benchmark(
2159 name = "global_average_pooling_bench",
2160 srcs = ["bench/global-average-pooling.cc"],
2161 deps = OPERATOR_BENCHMARK_DEPS,
2162)
2163
2164xnnpack_benchmark(
2165 name = "max_pooling_bench",
2166 srcs = ["bench/max-pooling.cc"],
2167 deps = OPERATOR_BENCHMARK_DEPS,
2168)
2169
2170xnnpack_benchmark(
2171 name = "sigmoid_bench",
2172 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002173 copts = xnnpack_optional_tflite_copts(),
2174 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002175)
2176
2177xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002178 name = "prelu_bench",
2179 srcs = ["bench/prelu.cc"],
2180 copts = xnnpack_optional_tflite_copts(),
2181 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2182)
2183
2184xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08002185 name = "softmax_bench",
2186 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08002187 copts = xnnpack_optional_tflite_copts(),
2188 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002189)
2190
Marat Dukhanc068bb62019-10-04 13:24:39 -07002191############################# End-to-end benchmarks ############################
2192
2193cc_library(
2194 name = "mobilenet_v1",
2195 srcs = ["models/mobilenet-v1.cc"],
2196 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002197 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002198 linkstatic = True,
2199 deps = [
2200 ":XNNPACK",
2201 "@pthreadpool",
2202 ],
2203)
2204
2205cc_library(
2206 name = "mobilenet_v2",
2207 srcs = ["models/mobilenet-v2.cc"],
2208 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002209 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002210 linkstatic = True,
2211 deps = [
2212 ":XNNPACK",
2213 "@pthreadpool",
2214 ],
2215)
2216
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002217cc_library(
2218 name = "mobilenet_v3_large",
2219 srcs = ["models/mobilenet-v3-large.cc"],
2220 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002221 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002222 linkstatic = True,
2223 deps = [
2224 ":XNNPACK",
2225 "@pthreadpool",
2226 ],
2227)
2228
2229cc_library(
2230 name = "mobilenet_v3_small",
2231 srcs = ["models/mobilenet-v3-small.cc"],
2232 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002233 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002234 linkstatic = True,
2235 deps = [
2236 ":XNNPACK",
2237 "@pthreadpool",
2238 ],
2239)
2240
Marat Dukhanc068bb62019-10-04 13:24:39 -07002241xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002242 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002243 srcs = [
2244 "bench/f32-dwconv-e2e.cc",
2245 "bench/end2end.h",
2246 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002247 copts = ["-Wno-unused-function"],
2248 deps = MICROKERNEL_BENCHMARK_DEPS + [
2249 ":XNNPACK",
2250 ":mobilenet_v1",
2251 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002252 ":mobilenet_v3_large",
2253 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002254 ],
2255)
2256
2257xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002258 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002259 srcs = [
2260 "bench/f32-gemm-e2e.cc",
2261 "bench/end2end.h",
2262 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002263 copts = ["-Wno-unused-function"],
2264 deps = MICROKERNEL_BENCHMARK_DEPS + [
2265 ":XNNPACK",
2266 ":mobilenet_v1",
2267 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002268 ":mobilenet_v3_large",
2269 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002270 ],
2271)
2272
2273xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002274 name = "end2end_bench",
2275 srcs = ["bench/end2end.cc"],
2276 deps = [
2277 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002278 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002279 ":mobilenet_v1",
2280 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002281 ":mobilenet_v3_large",
2282 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002283 "@pthreadpool",
2284 ],
2285)
2286
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002287#################### Accuracy evaluation for math functions ####################
2288
2289xnnpack_benchmark(
2290 name = "f32_exp_eval",
2291 srcs = [
2292 "eval/f32-exp.cc",
2293 "src/xnnpack/AlignedAllocator.h",
2294 ] + ACCURACY_EVAL_HDRS,
2295 deps = ACCURACY_EVAL_DEPS,
2296)
2297
Marat Dukhan515c9772019-10-17 18:07:57 -07002298xnnpack_benchmark(
2299 name = "f32_expminus_eval",
2300 srcs = [
2301 "eval/f32-expminus.cc",
2302 "src/xnnpack/AlignedAllocator.h",
2303 ] + ACCURACY_EVAL_HDRS,
2304 deps = ACCURACY_EVAL_DEPS,
2305)
2306
Marat Dukhan98ba4412019-10-23 02:14:28 -07002307xnnpack_benchmark(
2308 name = "f32_extexp_eval",
2309 srcs = [
2310 "eval/f32-extexp.cc",
2311 "src/xnnpack/AlignedAllocator.h",
2312 ] + ACCURACY_EVAL_HDRS,
2313 deps = ACCURACY_EVAL_DEPS,
2314)
2315
Marat Dukhan346a9e52019-11-15 09:06:30 -08002316xnnpack_benchmark(
2317 name = "f32_sigmoid_eval",
2318 srcs = [
2319 "eval/f32-sigmoid.cc",
2320 "src/xnnpack/AlignedAllocator.h",
2321 ] + ACCURACY_EVAL_HDRS,
2322 deps = ACCURACY_EVAL_DEPS,
2323)
2324
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325######################### Unit tests for micro-kernels #########################
2326
2327xnnpack_unit_test(
2328 name = "f16_gemm_test",
2329 srcs = [
2330 "test/f16-gemm.cc",
2331 "test/gemm-microkernel-tester.h",
2332 "src/xnnpack/AlignedAllocator.h",
2333 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2334 deps = MICROKERNEL_TEST_DEPS,
2335)
2336
2337xnnpack_unit_test(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002338 name = "f16_spmm_test",
2339 srcs = [
2340 "test/f16-spmm.cc",
2341 "test/spmm-microkernel-tester.h",
2342 "src/xnnpack/AlignedAllocator.h",
2343 ] + MICROKERNEL_TEST_HDRS,
2344 deps = MICROKERNEL_TEST_DEPS,
2345)
2346
2347xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002348 name = "f32_argmaxpool_test",
2349 srcs = [
2350 "test/f32-argmaxpool.cc",
2351 "test/argmaxpool-microkernel-tester.h",
2352 "src/xnnpack/AlignedAllocator.h",
2353 ] + MICROKERNEL_TEST_HDRS,
2354 deps = MICROKERNEL_TEST_DEPS,
2355)
2356
2357xnnpack_unit_test(
2358 name = "f32_avgpool_test",
2359 srcs = [
2360 "test/f32-avgpool.cc",
2361 "test/avgpool-microkernel-tester.h",
2362 "src/xnnpack/AlignedAllocator.h",
2363 ] + MICROKERNEL_TEST_HDRS,
2364 deps = MICROKERNEL_TEST_DEPS,
2365)
2366
2367xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002368 name = "f32_bilinear_test",
2369 srcs = [
2370 "test/f32-bilinear.cc",
2371 "test/bilinear-microkernel-tester.h",
2372 "src/xnnpack/AlignedAllocator.h",
2373 ] + MICROKERNEL_TEST_HDRS,
2374 deps = MICROKERNEL_TEST_DEPS,
2375)
2376
2377xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002378 name = "f32_clamp_test",
2379 srcs = [
2380 "test/f32-clamp.cc",
2381 "test/clamp-microkernel-tester.h",
2382 ] + MICROKERNEL_TEST_HDRS,
2383 deps = MICROKERNEL_TEST_DEPS,
2384)
2385
2386xnnpack_unit_test(
2387 name = "f32_igemm_test",
2388 srcs = [
2389 "test/f32-igemm.cc",
2390 "test/gemm-microkernel-tester.h",
2391 "src/xnnpack/AlignedAllocator.h",
2392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2393 deps = MICROKERNEL_TEST_DEPS,
2394)
2395
2396xnnpack_unit_test(
2397 name = "f32_conv_hwc_test",
2398 srcs = [
2399 "test/f32-conv-hwc.cc",
2400 "test/conv-hwc-microkernel-tester.h",
2401 "src/xnnpack/AlignedAllocator.h",
2402 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2403 deps = MICROKERNEL_TEST_DEPS,
2404)
2405
2406xnnpack_unit_test(
2407 name = "f32_conv_hwc2spchw_test",
2408 srcs = [
2409 "test/f32-conv-hwc2spchw.cc",
2410 "test/conv-hwc2spchw-microkernel-tester.h",
2411 "src/xnnpack/AlignedAllocator.h",
2412 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2413 deps = MICROKERNEL_TEST_DEPS,
2414)
2415
2416xnnpack_unit_test(
2417 name = "f32_dwconv_test",
2418 srcs = [
2419 "test/f32-dwconv.cc",
2420 "test/dwconv-microkernel-tester.h",
2421 "src/xnnpack/AlignedAllocator.h",
2422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2423 deps = MICROKERNEL_TEST_DEPS,
2424)
2425
2426xnnpack_unit_test(
2427 name = "f32_dwconv_spchw_test",
2428 srcs = [
2429 "test/f32-dwconv-spchw.cc",
2430 "test/dwconv-spchw-microkernel-tester.h",
2431 "src/xnnpack/AlignedAllocator.h",
2432 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2433 deps = MICROKERNEL_TEST_DEPS,
2434)
2435
2436xnnpack_unit_test(
2437 name = "f32_gavgpool_test",
2438 srcs = [
2439 "test/f32-gavgpool.cc",
2440 "test/gavgpool-microkernel-tester.h",
2441 "src/xnnpack/AlignedAllocator.h",
2442 ] + MICROKERNEL_TEST_HDRS,
2443 deps = MICROKERNEL_TEST_DEPS,
2444)
2445
2446xnnpack_unit_test(
2447 name = "f32_gavgpool_spchw_test",
2448 srcs = [
2449 "test/f32-gavgpool-spchw.cc",
2450 "test/gavgpool-spchw-microkernel-tester.h",
2451 "src/xnnpack/AlignedAllocator.h",
2452 ] + MICROKERNEL_TEST_HDRS,
2453 deps = MICROKERNEL_TEST_DEPS,
2454)
2455
2456xnnpack_unit_test(
2457 name = "f32_gemm_test",
2458 srcs = [
2459 "test/f32-gemm.cc",
2460 "test/gemm-microkernel-tester.h",
2461 "src/xnnpack/AlignedAllocator.h",
2462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2463 deps = MICROKERNEL_TEST_DEPS,
2464)
2465
2466xnnpack_unit_test(
2467 name = "f32_gemminc_test",
2468 srcs = [
2469 "test/f32-gemminc.cc",
2470 "test/gemm-microkernel-tester.h",
2471 "src/xnnpack/AlignedAllocator.h",
2472 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2473 deps = MICROKERNEL_TEST_DEPS,
2474)
2475
2476xnnpack_unit_test(
2477 name = "f32_hswish_test",
2478 srcs = [
2479 "test/f32-hswish.cc",
2480 "test/hswish-microkernel-tester.h",
2481 ] + MICROKERNEL_TEST_HDRS,
2482 deps = MICROKERNEL_TEST_DEPS,
2483)
2484
2485xnnpack_unit_test(
2486 name = "f32_maxpool_test",
2487 srcs = [
2488 "test/f32-maxpool.cc",
2489 "test/maxpool-microkernel-tester.h",
2490 ] + MICROKERNEL_TEST_HDRS,
2491 deps = MICROKERNEL_TEST_DEPS,
2492)
2493
2494xnnpack_unit_test(
2495 name = "f32_pavgpool_test",
2496 srcs = [
2497 "test/f32-pavgpool.cc",
2498 "test/avgpool-microkernel-tester.h",
2499 "src/xnnpack/AlignedAllocator.h",
2500 ] + MICROKERNEL_TEST_HDRS,
2501 deps = MICROKERNEL_TEST_DEPS,
2502)
2503
2504xnnpack_unit_test(
2505 name = "f32_ppmm_test",
2506 srcs = [
2507 "test/f32-ppmm.cc",
2508 "test/gemm-microkernel-tester.h",
2509 "src/xnnpack/AlignedAllocator.h",
2510 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2511 deps = MICROKERNEL_TEST_DEPS,
2512)
2513
2514xnnpack_unit_test(
2515 name = "f32_prelu_test",
2516 srcs = [
2517 "test/f32-prelu.cc",
2518 "test/prelu-microkernel-tester.h",
2519 "src/xnnpack/AlignedAllocator.h",
2520 ] + MICROKERNEL_TEST_HDRS,
2521 deps = MICROKERNEL_TEST_DEPS,
2522)
2523
2524xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002525 name = "f32_raddexpminusmax_test",
2526 srcs = [
2527 "test/f32-raddexpminusmax.cc",
2528 "test/raddexpminusmax-microkernel-tester.h",
2529 ] + MICROKERNEL_TEST_HDRS,
2530 deps = MICROKERNEL_TEST_DEPS,
2531)
2532
2533xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002534 name = "f32_raddextexp_test",
2535 srcs = [
2536 "test/f32-raddextexp.cc",
2537 "test/raddextexp-microkernel-tester.h",
2538 ] + MICROKERNEL_TEST_HDRS,
2539 deps = MICROKERNEL_TEST_DEPS,
2540)
2541
2542xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002543 name = "f32_raddstoreexpminusmax_test",
2544 srcs = [
2545 "test/f32-raddstoreexpminusmax.cc",
2546 "test/raddstoreexpminusmax-microkernel-tester.h",
2547 ] + MICROKERNEL_TEST_HDRS,
2548 deps = MICROKERNEL_TEST_DEPS,
2549)
2550
2551xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002552 name = "f32_rmax_test",
2553 srcs = [
2554 "test/f32-rmax.cc",
2555 "test/rmax-microkernel-tester.h",
2556 ] + MICROKERNEL_TEST_HDRS,
2557 deps = MICROKERNEL_TEST_DEPS,
2558)
2559
2560xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002561 name = "f32_sigmoid_test",
2562 srcs = [
2563 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002564 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002565 ] + MICROKERNEL_TEST_HDRS,
2566 deps = MICROKERNEL_TEST_DEPS,
2567)
2568
2569xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002570 name = "f32_spmm_test",
2571 srcs = [
2572 "test/f32-spmm.cc",
2573 "test/spmm-microkernel-tester.h",
2574 "src/xnnpack/AlignedAllocator.h",
2575 ] + MICROKERNEL_TEST_HDRS,
2576 deps = MICROKERNEL_TEST_DEPS,
2577)
2578
2579xnnpack_unit_test(
2580 name = "f32_vadd_test",
2581 srcs = [
2582 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002583 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002584 ] + MICROKERNEL_TEST_HDRS,
2585 deps = MICROKERNEL_TEST_DEPS,
2586)
2587
2588xnnpack_unit_test(
2589 name = "f32_vaddc_test",
2590 srcs = [
2591 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002592 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002593 ] + MICROKERNEL_TEST_HDRS,
2594 deps = MICROKERNEL_TEST_DEPS,
2595)
2596
2597xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002598 name = "f32_vdiv_test",
2599 srcs = [
2600 "test/f32-vdiv.cc",
2601 "test/vbinary-microkernel-tester.h",
2602 ] + MICROKERNEL_TEST_HDRS,
2603 deps = MICROKERNEL_TEST_DEPS,
2604)
2605
2606xnnpack_unit_test(
2607 name = "f32_vdivc_test",
2608 srcs = [
2609 "test/f32-vdivc.cc",
2610 "test/vbinaryc-microkernel-tester.h",
2611 ] + MICROKERNEL_TEST_HDRS,
2612 deps = MICROKERNEL_TEST_DEPS,
2613)
2614
2615xnnpack_unit_test(
2616 name = "f32_vrdivc_test",
2617 srcs = [
2618 "test/f32-vrdivc.cc",
2619 "test/vbinaryc-microkernel-tester.h",
2620 ] + MICROKERNEL_TEST_HDRS,
2621 deps = MICROKERNEL_TEST_DEPS,
2622)
2623
2624xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002625 name = "f32_vmax_test",
2626 srcs = [
2627 "test/f32-vmax.cc",
2628 "test/vbinary-microkernel-tester.h",
2629 ] + MICROKERNEL_TEST_HDRS,
2630 deps = MICROKERNEL_TEST_DEPS,
2631)
2632
2633xnnpack_unit_test(
2634 name = "f32_vmaxc_test",
2635 srcs = [
2636 "test/f32-vmaxc.cc",
2637 "test/vbinaryc-microkernel-tester.h",
2638 ] + MICROKERNEL_TEST_HDRS,
2639 deps = MICROKERNEL_TEST_DEPS,
2640)
2641
2642xnnpack_unit_test(
2643 name = "f32_vmin_test",
2644 srcs = [
2645 "test/f32-vmin.cc",
2646 "test/vbinary-microkernel-tester.h",
2647 ] + MICROKERNEL_TEST_HDRS,
2648 deps = MICROKERNEL_TEST_DEPS,
2649)
2650
2651xnnpack_unit_test(
2652 name = "f32_vminc_test",
2653 srcs = [
2654 "test/f32-vminc.cc",
2655 "test/vbinaryc-microkernel-tester.h",
2656 ] + MICROKERNEL_TEST_HDRS,
2657 deps = MICROKERNEL_TEST_DEPS,
2658)
2659
2660xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002661 name = "f32_vmul_test",
2662 srcs = [
2663 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002664 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002665 ] + MICROKERNEL_TEST_HDRS,
2666 deps = MICROKERNEL_TEST_DEPS,
2667)
2668
2669xnnpack_unit_test(
2670 name = "f32_vmulc_test",
2671 srcs = [
2672 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002673 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002674 ] + MICROKERNEL_TEST_HDRS,
2675 deps = MICROKERNEL_TEST_DEPS,
2676)
2677
2678xnnpack_unit_test(
2679 name = "f32_vmulcaddc_test",
2680 srcs = [
2681 "test/f32-vmulcaddc.cc",
2682 "test/vmulcaddc-microkernel-tester.h",
2683 "src/xnnpack/AlignedAllocator.h",
2684 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2685 deps = MICROKERNEL_TEST_DEPS,
2686)
2687
2688xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002689 name = "f32_vscale_test",
2690 srcs = [
2691 "test/f32-vscale.cc",
2692 "test/vscale-microkernel-tester.h",
2693 ] + MICROKERNEL_TEST_HDRS,
2694 deps = MICROKERNEL_TEST_DEPS,
2695)
2696
2697xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002698 name = "f32_vscaleexpminusmax_test",
2699 srcs = [
2700 "test/f32-vscaleexpminusmax.cc",
2701 "test/vscaleexpminusmax-microkernel-tester.h",
2702 ] + MICROKERNEL_TEST_HDRS,
2703 deps = MICROKERNEL_TEST_DEPS,
2704)
2705
2706xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002707 name = "f32_vscaleextexp_test",
2708 srcs = [
2709 "test/f32-vscaleextexp.cc",
2710 "test/vscaleextexp-microkernel-tester.h",
2711 ] + MICROKERNEL_TEST_HDRS,
2712 deps = MICROKERNEL_TEST_DEPS,
2713)
2714
2715xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002716 name = "f32_vsub_test",
2717 srcs = [
2718 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002719 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002720 ] + MICROKERNEL_TEST_HDRS,
2721 deps = MICROKERNEL_TEST_DEPS,
2722)
2723
2724xnnpack_unit_test(
2725 name = "f32_vsubc_test",
2726 srcs = [
2727 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002728 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002729 ] + MICROKERNEL_TEST_HDRS,
2730 deps = MICROKERNEL_TEST_DEPS,
2731)
2732
2733xnnpack_unit_test(
2734 name = "f32_vrsubc_test",
2735 srcs = [
2736 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002737 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002738 ] + MICROKERNEL_TEST_HDRS,
2739 deps = MICROKERNEL_TEST_DEPS,
2740)
2741
2742xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 name = "q8_avgpool_test",
2744 srcs = [
2745 "test/q8-avgpool.cc",
2746 "test/avgpool-microkernel-tester.h",
2747 "src/xnnpack/AlignedAllocator.h",
2748 ] + MICROKERNEL_TEST_HDRS,
2749 deps = MICROKERNEL_TEST_DEPS,
2750)
2751
2752xnnpack_unit_test(
2753 name = "q8_igemm_test",
2754 srcs = [
2755 "test/q8-igemm.cc",
2756 "test/gemm-microkernel-tester.h",
2757 "src/xnnpack/AlignedAllocator.h",
2758 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2759 deps = MICROKERNEL_TEST_DEPS,
2760)
2761
2762xnnpack_unit_test(
2763 name = "q8_dwconv_test",
2764 srcs = [
2765 "test/q8-dwconv.cc",
2766 "test/dwconv-microkernel-tester.h",
2767 "src/xnnpack/AlignedAllocator.h",
2768 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2769 deps = MICROKERNEL_TEST_DEPS,
2770)
2771
2772xnnpack_unit_test(
2773 name = "q8_gavgpool_test",
2774 srcs = [
2775 "test/q8-gavgpool.cc",
2776 "test/gavgpool-microkernel-tester.h",
2777 "src/xnnpack/AlignedAllocator.h",
2778 ] + MICROKERNEL_TEST_HDRS,
2779 deps = MICROKERNEL_TEST_DEPS,
2780)
2781
2782xnnpack_unit_test(
2783 name = "q8_gemm_test",
2784 srcs = [
2785 "test/q8-gemm.cc",
2786 "test/gemm-microkernel-tester.h",
2787 "src/xnnpack/AlignedAllocator.h",
2788 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2789 deps = MICROKERNEL_TEST_DEPS,
2790)
2791
2792xnnpack_unit_test(
2793 name = "q8_vadd_test",
2794 srcs = [
2795 "test/q8-vadd.cc",
2796 "test/vadd-microkernel-tester.h",
2797 ] + MICROKERNEL_TEST_HDRS,
2798 deps = MICROKERNEL_TEST_DEPS,
2799)
2800
2801xnnpack_unit_test(
2802 name = "u8_clamp_test",
2803 srcs = [
2804 "test/u8-clamp.cc",
2805 "test/clamp-microkernel-tester.h",
2806 ] + MICROKERNEL_TEST_HDRS,
2807 deps = MICROKERNEL_TEST_DEPS,
2808)
2809
2810xnnpack_unit_test(
2811 name = "u8_lut32norm_test",
2812 srcs = [
2813 "test/u8-lut32norm.cc",
2814 "test/lut-norm-microkernel-tester.h",
2815 ] + MICROKERNEL_TEST_HDRS,
2816 deps = MICROKERNEL_TEST_DEPS,
2817)
2818
2819xnnpack_unit_test(
2820 name = "u8_maxpool_test",
2821 srcs = [
2822 "test/u8-maxpool.cc",
2823 "test/maxpool-microkernel-tester.h",
2824 ] + MICROKERNEL_TEST_HDRS,
2825 deps = MICROKERNEL_TEST_DEPS,
2826)
2827
2828xnnpack_unit_test(
2829 name = "u8_rmax_test",
2830 srcs = [
2831 "test/u8-rmax.cc",
2832 "test/rmax-microkernel-tester.h",
2833 ] + MICROKERNEL_TEST_HDRS,
2834 deps = MICROKERNEL_TEST_DEPS,
2835)
2836
2837xnnpack_unit_test(
2838 name = "x32_packx_test",
2839 srcs = [
2840 "test/x32-packx.cc",
2841 "test/pack-microkernel-tester.h",
2842 "src/xnnpack/AlignedAllocator.h",
2843 ] + MICROKERNEL_TEST_HDRS,
2844 deps = MICROKERNEL_TEST_DEPS,
2845)
2846
2847xnnpack_unit_test(
2848 name = "x32_pad_test",
2849 srcs = [
2850 "test/x32-pad.cc",
2851 "test/pad-microkernel-tester.h",
2852 ] + MICROKERNEL_TEST_HDRS,
2853 deps = MICROKERNEL_TEST_DEPS,
2854)
2855
2856xnnpack_unit_test(
2857 name = "x32_unpool_test",
2858 srcs = [
2859 "test/x32-unpool.cc",
2860 "test/unpool-microkernel-tester.h",
2861 ] + MICROKERNEL_TEST_HDRS,
2862 deps = MICROKERNEL_TEST_DEPS,
2863)
2864
2865xnnpack_unit_test(
2866 name = "x32_zip_test",
2867 srcs = [
2868 "test/x32-zip.cc",
2869 "test/zip-microkernel-tester.h",
2870 ] + MICROKERNEL_TEST_HDRS,
2871 deps = MICROKERNEL_TEST_DEPS,
2872)
2873
2874xnnpack_unit_test(
2875 name = "x8_lut_test",
2876 srcs = [
2877 "test/x8-lut.cc",
2878 "test/lut-microkernel-tester.h",
2879 ] + MICROKERNEL_TEST_HDRS,
2880 deps = MICROKERNEL_TEST_DEPS,
2881)
2882
2883xnnpack_unit_test(
2884 name = "x8_zip_test",
2885 srcs = [
2886 "test/x8-zip.cc",
2887 "test/zip-microkernel-tester.h",
2888 ] + MICROKERNEL_TEST_HDRS,
2889 deps = MICROKERNEL_TEST_DEPS,
2890)
2891
2892########################### Size test for the library ##########################
2893
2894xnnpack_binary(
2895 name = "size_test",
2896 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002897 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898)
2899
2900########################### Unit tests for operators ###########################
2901
2902xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002903 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002904 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002905 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002906 "test/add-operator-tester.h",
2907 ],
2908 deps = OPERATOR_TEST_DEPS,
2909)
2910
2911xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002912 name = "add_nd_test",
2913 srcs = [
2914 "test/add-nd.cc",
2915 "test/binary-elementwise-operator-tester.h",
2916 ],
2917 deps = OPERATOR_TEST_DEPS,
2918)
2919
2920xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002921 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002923 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002924 "test/argmax-pooling-operator-tester.h",
2925 ] + OPERATOR_TEST_PARAMS_HDRS,
2926 deps = OPERATOR_TEST_DEPS,
2927)
2928
2929xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002930 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002932 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002933 "test/average-pooling-operator-tester.h",
2934 ] + OPERATOR_TEST_PARAMS_HDRS,
2935 deps = OPERATOR_TEST_DEPS,
2936)
2937
2938xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002939 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002940 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002941 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 "test/channel-pad-operator-tester.h",
2943 ] + OPERATOR_TEST_PARAMS_HDRS,
2944 deps = OPERATOR_TEST_DEPS,
2945)
2946
2947xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002948 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002950 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002951 "test/channel-shuffle-operator-tester.h",
2952 ],
2953 deps = OPERATOR_TEST_DEPS,
2954)
2955
2956xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002957 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002959 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960 "test/clamp-operator-tester.h",
2961 ],
2962 deps = OPERATOR_TEST_DEPS,
2963)
2964
2965xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002966 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002968 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002969 "test/convolution-operator-tester.h",
2970 ],
2971 deps = OPERATOR_TEST_DEPS,
2972)
2973
2974xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002975 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002977 "test/convolution-nchw.cc",
2978 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 ],
2980 deps = OPERATOR_TEST_DEPS,
2981)
2982
2983xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002984 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002986 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 "test/deconvolution-operator-tester.h",
2988 ] + OPERATOR_TEST_PARAMS_HDRS,
2989 deps = OPERATOR_TEST_DEPS,
2990)
2991
2992xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002993 name = "divide_nd_test",
2994 srcs = [
2995 "test/binary-elementwise-operator-tester.h",
2996 "test/divide-nd.cc",
2997 ],
2998 deps = OPERATOR_TEST_DEPS,
2999)
3000
3001xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003002 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003003 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003004 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 "test/fully-connected-operator-tester.h",
3006 ],
3007 deps = OPERATOR_TEST_DEPS,
3008)
3009
3010xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003011 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003012 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003013 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003014 "test/global-average-pooling-operator-tester.h",
3015 ] + OPERATOR_TEST_PARAMS_HDRS,
3016 deps = OPERATOR_TEST_DEPS,
3017)
3018
3019xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003020 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003022 "test/global-average-pooling-ncw.cc",
3023 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 ],
3025 deps = OPERATOR_TEST_DEPS,
3026)
3027
3028xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003029 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003031 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003032 "test/hardswish-operator-tester.h",
3033 ],
3034 deps = OPERATOR_TEST_DEPS,
3035)
3036
3037xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003038 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003040 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003041 "test/leaky-relu-operator-tester.h",
3042 ],
3043 deps = OPERATOR_TEST_DEPS,
3044)
3045
3046xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003047 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003048 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003049 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003050 "test/max-pooling-operator-tester.h",
3051 ] + OPERATOR_TEST_PARAMS_HDRS,
3052 deps = OPERATOR_TEST_DEPS,
3053)
3054
3055xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08003056 name = "maximum_nd_test",
3057 srcs = [
3058 "test/binary-elementwise-operator-tester.h",
3059 "test/maximum-nd.cc",
3060 ],
3061 deps = OPERATOR_TEST_DEPS,
3062)
3063
3064xnnpack_unit_test(
3065 name = "minimum_nd_test",
3066 srcs = [
3067 "test/binary-elementwise-operator-tester.h",
3068 "test/minimum-nd.cc",
3069 ],
3070 deps = OPERATOR_TEST_DEPS,
3071)
3072
3073xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003074 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003075 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08003076 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003077 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08003078 ],
3079 deps = OPERATOR_TEST_DEPS,
3080)
3081
3082xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003083 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003084 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003085 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 "test/prelu-operator-tester.h",
3087 ] + OPERATOR_TEST_PARAMS_HDRS,
3088 deps = OPERATOR_TEST_DEPS,
3089)
3090
3091xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003092 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003093 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003094 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003095 "test/resize-bilinear-operator-tester.h",
3096 ] + OPERATOR_TEST_PARAMS_HDRS,
3097 deps = OPERATOR_TEST_DEPS,
3098)
3099
3100xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003101 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003102 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003103 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104 "test/sigmoid-operator-tester.h",
3105 ],
3106 deps = OPERATOR_TEST_DEPS,
3107)
3108
3109xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003110 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003112 "test/softmax-nc.cc",
3113 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114 ],
3115 deps = OPERATOR_TEST_DEPS,
3116)
3117
3118xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003119 name = "subtract_nd_test",
3120 srcs = [
3121 "test/binary-elementwise-operator-tester.h",
3122 "test/subtract-nd.cc",
3123 ],
3124 deps = OPERATOR_TEST_DEPS,
3125)
3126
3127xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003128 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003129 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003130 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131 "test/unpooling-operator-tester.h",
3132 ],
3133 deps = OPERATOR_TEST_DEPS,
3134)
3135
3136############################# Build configurations #############################
3137
Marat Dukhanb8642352019-10-30 15:43:02 -07003138# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003139config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003140 name = "xnn_enable_assembly_explicit_true",
3141 define_values = {"xnn_enable_assembly": "true"},
3142)
3143
3144# Disables usage of assembly kernels.
3145config_setting(
3146 name = "xnn_enable_assembly_explicit_false",
3147 define_values = {"xnn_enable_assembly": "false"},
3148)
3149
3150# Builds with -c dbg
3151config_setting(
3152 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003154 "compilation_mode": "dbg",
3155 },
3156)
3157
3158# Builds with -c opt
3159config_setting(
3160 name = "optimized_build",
3161 values = {
3162 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003163 },
3164)
3165
3166config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003167 name = "linux_k8",
3168 values = {"cpu": "k8"},
3169)
3170
3171config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003172 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003173 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003174)
3175
3176config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003177 name = "android",
3178 values = {"crosstool_top": "//external:android/crosstool"},
3179)
3180
3181config_setting(
3182 name = "android_armv7",
3183 values = {
3184 "crosstool_top": "//external:android/crosstool",
3185 "cpu": "armeabi-v7a",
3186 },
3187)
3188
3189config_setting(
3190 name = "android_arm64",
3191 values = {
3192 "crosstool_top": "//external:android/crosstool",
3193 "cpu": "arm64-v8a",
3194 },
3195)
3196
3197config_setting(
3198 name = "android_x86",
3199 values = {
3200 "crosstool_top": "//external:android/crosstool",
3201 "cpu": "x86",
3202 },
3203)
3204
3205config_setting(
3206 name = "android_x86_64",
3207 values = {
3208 "crosstool_top": "//external:android/crosstool",
3209 "cpu": "x86_64",
3210 },
3211)
3212
3213config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003214 name = "macos_x86_64",
3215 values = {
3216 "apple_platform_type": "macos",
3217 "cpu": "darwin",
3218 },
3219)
3220
3221config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003222 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003223 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003224)
3225
3226config_setting(
3227 name = "emscripten_wasm",
3228 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003229 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 "cpu": "wasm",
3231 },
3232)
3233
3234config_setting(
3235 name = "emscripten_wasmsimd",
3236 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003237 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003238 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003239 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 },
3241)
3242
3243config_setting(
3244 name = "emscripten_asmjs",
3245 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003246 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247 "cpu": "asmjs",
3248 },
3249)
Miao Wang3fa1f012020-02-17 22:45:06 +00003250
3251config_setting(
3252 name = "ios_armv7",
3253 values = {
3254 "crosstool_top": "//tools/osx/crosstool:crosstool",
3255 "cpu": "ios_armv7",
3256 },
3257)
3258
3259config_setting(
3260 name = "ios_arm64",
3261 values = {
3262 "crosstool_top": "//tools/osx/crosstool:crosstool",
3263 "cpu": "ios_arm64",
3264 },
3265)
3266
3267config_setting(
3268 name = "ios_arm64e",
3269 values = {
3270 "crosstool_top": "//tools/osx/crosstool:crosstool",
3271 "cpu": "ios_arm64e",
3272 },
3273)
3274
3275config_setting(
3276 name = "ios_x86",
3277 values = {
3278 "crosstool_top": "//tools/osx/crosstool:crosstool",
3279 "cpu": "ios_i386",
3280 },
3281)
3282
3283config_setting(
3284 name = "ios_x86_64",
3285 values = {
3286 "crosstool_top": "//tools/osx/crosstool:crosstool",
3287 "cpu": "ios_x86_64",
3288 },
3289)
3290
3291config_setting(
3292 name = "watchos_armv7k",
3293 values = {
3294 "crosstool_top": "//tools/osx/crosstool:crosstool",
3295 "cpu": "watchos_armv7k",
3296 },
3297)
3298
3299config_setting(
3300 name = "watchos_arm64_32",
3301 values = {
3302 "crosstool_top": "//tools/osx/crosstool:crosstool",
3303 "cpu": "watchos_arm64_32",
3304 },
3305)
3306
3307config_setting(
3308 name = "watchos_x86",
3309 values = {
3310 "crosstool_top": "//tools/osx/crosstool:crosstool",
3311 "cpu": "watchos_i386",
3312 },
3313)
3314
3315config_setting(
3316 name = "watchos_x86_64",
3317 values = {
3318 "crosstool_top": "//tools/osx/crosstool:crosstool",
3319 "cpu": "watchos_x86_64",
3320 },
3321)
3322
3323config_setting(
3324 name = "tvos_arm64",
3325 values = {
3326 "crosstool_top": "//tools/osx/crosstool:crosstool",
3327 "cpu": "tvos_arm64",
3328 },
3329)
3330
3331config_setting(
3332 name = "tvos_x86_64",
3333 values = {
3334 "crosstool_top": "//tools/osx/crosstool:crosstool",
3335 "cpu": "tvos_x86_64",
3336 },
3337)