blob: b418e281d185667cc5d81a0ac7252e3dc2bc2c8c [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",
72 "src/softargmax-nc.c",
73 "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 Dukhan08c4a432019-10-03 09:29:21 -0700521 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800522 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800523 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
524 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
525 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
526 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
527 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
528 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
529 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
530 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
531 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
532 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
533 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
534 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
535 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
536 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
537 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
538 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
539 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
540 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800541 "src/f32-vbinary/gen/vadd-neon-x4.c",
542 "src/f32-vbinary/gen/vadd-neon-x8.c",
543 "src/f32-vbinary/gen/vaddc-neon-x4.c",
544 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800545 "src/f32-vbinary/gen/vmax-neon-x4.c",
546 "src/f32-vbinary/gen/vmax-neon-x8.c",
547 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
548 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
549 "src/f32-vbinary/gen/vmin-neon-x4.c",
550 "src/f32-vbinary/gen/vmin-neon-x8.c",
551 "src/f32-vbinary/gen/vminc-neon-x4.c",
552 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800553 "src/f32-vbinary/gen/vmul-neon-x4.c",
554 "src/f32-vbinary/gen/vmul-neon-x8.c",
555 "src/f32-vbinary/gen/vmulc-neon-x4.c",
556 "src/f32-vbinary/gen/vmulc-neon-x8.c",
557 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
558 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
559 "src/f32-vbinary/gen/vsub-neon-x4.c",
560 "src/f32-vbinary/gen/vsub-neon-x8.c",
561 "src/f32-vbinary/gen/vsubc-neon-x4.c",
562 "src/f32-vbinary/gen/vsubc-neon-x8.c",
563 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
564 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700565 "src/q8-avgpool/mp9p8q-neon.c",
566 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700567 "src/q8-dwconv/up8x9-neon.c",
568 "src/q8-gavgpool/mp7p7q-neon.c",
569 "src/q8-gavgpool/up7-neon.c",
570 "src/q8-gemm/4x8-neon.c",
571 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800572 "src/q8-igemm/4x8-neon.c",
573 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700574 "src/q8-vadd/neon.c",
575 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800576 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700577 "src/u8-rmax/neon.c",
578 "src/x32-packx/x4-neon-st4.c",
579 "src/x32-pad/x2-neon.c",
580 "src/x32-zip/x2-neon.c",
581 "src/x32-zip/x3-neon.c",
582 "src/x32-zip/x4-neon.c",
583 "src/x32-zip/xm-neon.c",
584 "src/x8-zip/x2-neon.c",
585 "src/x8-zip/x3-neon.c",
586 "src/x8-zip/x4-neon.c",
587 "src/x8-zip/xm-neon.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800588 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800589 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
590 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
591 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
592 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
593 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
594 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700595]
596
597NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800598 "src/f32-bilinear/gen/neonfma-c4.c",
599 "src/f32-bilinear/gen/neonfma-c8.c",
600 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
601 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
602 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
603 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800604 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800605 "src/f32-igemm/gen/1x8s4-neonfma.c",
606 "src/f32-igemm/gen/4x8s4-neonfma.c",
607 "src/f32-igemm/gen/6x8s4-neonfma.c",
608 "src/f32-igemm/gen/8x8s4-neonfma.c",
609 "src/f32-dwconv/gen/up4x9-neonfma.c",
610 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
611 "src/f32-dwconv/gen/up8x9-neonfma.c",
612 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
613 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
614 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
615 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
616 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800617 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800618 "src/f32-gemm/gen/1x8s4-neonfma.c",
619 "src/f32-gemm/gen/4x8s4-neonfma.c",
620 "src/f32-gemm/gen/6x8s4-neonfma.c",
621 "src/f32-gemm/gen/8x8s4-neonfma.c",
622 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
623 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
624 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
625 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800626 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800627 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
628 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
629 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
630 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800631 "src/f32-hswish/gen/neonfma-x4.c",
632 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800633 "src/f32-ppmm/gen/4x8-neonfma.c",
634 "src/f32-ppmm/gen/8x8-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800635 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
636 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
637 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
638 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
639 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
640 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
641 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
642 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
643 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
644 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
645 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
646 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
647 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
648 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
649 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
650 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
651 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
652 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
653 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
654 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
655 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
656 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
657 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
658 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
659 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
660 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
661 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
662 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
663 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
664 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
665 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
666 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
667 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
668 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
669 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
670 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
671 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
672 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
673 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
674 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
675 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
676 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
677 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
678 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
679 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
680 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
681 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
682 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
683 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
684 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
685 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
686 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
687 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
688 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800689 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
690 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800691 "src/math/exp-neonfma-lut64-p2.c",
692 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800693 "src/math/expminus-neonfma-lut2048-p1.c",
694 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800695 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800696 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
697 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
698 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
699 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
700 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
701 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
702 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
703 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
704 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
705 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
706 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
707 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
708 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
709 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
710 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
711 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
712 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
713 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700714]
715
716AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800717 "src/f32-vbinary/gen/vdiv-neon-x4.c",
718 "src/f32-vbinary/gen/vdiv-neon-x8.c",
719 "src/f32-vbinary/gen/vdivc-neon-x4.c",
720 "src/f32-vbinary/gen/vdivc-neon-x8.c",
721 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
722 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800723 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
724 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
725 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
726 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
727 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
728 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800729 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800730 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
731 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
732 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
733 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
734 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800735 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800736 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
737 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
738 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
739 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
740 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
741 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800742 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700743 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
744 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
745 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
746 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
747 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
748 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
749 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800750 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
751 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
752 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
753 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
754 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
755 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
756 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
757 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
761 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
762 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
763 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
764 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
767 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800768 "src/f32-spmm/gen/12x1-neonfma.c",
769 "src/f32-spmm/gen/12x2-neonfma.c",
770 "src/f32-spmm/gen/12x4-neonfma.c",
771 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
772 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
773 "src/f32-spmm/gen/16x1-neonfma.c",
774 "src/f32-spmm/gen/16x2-neonfma.c",
775 "src/f32-spmm/gen/16x4-neonfma.c",
776 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
777 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
778 "src/f32-spmm/gen/4x1-neonfma.c",
779 "src/f32-spmm/gen/4x2-neonfma.c",
780 "src/f32-spmm/gen/4x4-neonfma.c",
781 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
782 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
783 "src/f32-spmm/gen/8x1-neonfma.c",
784 "src/f32-spmm/gen/8x2-neonfma.c",
785 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800786 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
787 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
788 "src/math/sigmoid-neonfma-rr1-p5-div.c",
789 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
790 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
791 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700792]
793
794AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800795 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
796 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
797 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700798]
799
800SSE_UKERNELS = [
801 "src/f32-avgpool/mp9p8q-sse.c",
802 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800803 "src/f32-bilinear/gen/sse-c4.c",
804 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700805 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800806 "src/f32-dwconv-spchw/3x3p1-sse.c",
807 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800808 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
809 "src/f32-dwconv/gen/up4x25-sse.c",
810 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
811 "src/f32-dwconv/gen/up4x4-sse.c",
812 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
813 "src/f32-dwconv/gen/up4x9-sse.c",
814 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
815 "src/f32-dwconv/gen/up8x25-sse.c",
816 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
817 "src/f32-dwconv/gen/up8x4-sse.c",
818 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
819 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700820 "src/f32-gavgpool-spchw/sse-x4.c",
821 "src/f32-gavgpool/mp7p7q-sse.c",
822 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800823 "src/f32-gemm/gen/1x8-sse-dup.c",
824 "src/f32-gemm/gen/1x8-sse-load1.c",
825 "src/f32-gemm/gen/1x8s4-sse.c",
826 "src/f32-gemm/gen/4x8-sse-dup.c",
827 "src/f32-gemm/gen/4x8-sse-load1.c",
828 "src/f32-gemm/gen/4x8s4-sse.c",
829 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
830 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
831 "src/f32-gemm/gen-inc/1x8s4-sse.c",
832 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
833 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
834 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800835 "src/f32-hswish/gen/sse-x4.c",
836 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800837 "src/f32-igemm/gen/1x8-sse-dup.c",
838 "src/f32-igemm/gen/1x8-sse-load1.c",
839 "src/f32-igemm/gen/1x8s4-sse.c",
840 "src/f32-igemm/gen/4x2c4-sse.c",
841 "src/f32-igemm/gen/4x8-sse-dup.c",
842 "src/f32-igemm/gen/4x8-sse-load1.c",
843 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800844 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700845 "src/f32-pavgpool/mp9p8q-sse.c",
846 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800847 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800849 "src/f32-spmm/gen/4x1-sse.c",
850 "src/f32-spmm/gen/8x1-sse.c",
851 "src/f32-vbinary/gen/vadd-sse-x4.c",
852 "src/f32-vbinary/gen/vadd-sse-x8.c",
853 "src/f32-vbinary/gen/vaddc-sse-x4.c",
854 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800855 "src/f32-vbinary/gen/vdiv-sse-x4.c",
856 "src/f32-vbinary/gen/vdiv-sse-x8.c",
857 "src/f32-vbinary/gen/vdivc-sse-x4.c",
858 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800859 "src/f32-vbinary/gen/vmax-sse-x4.c",
860 "src/f32-vbinary/gen/vmax-sse-x8.c",
861 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
862 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
863 "src/f32-vbinary/gen/vmin-sse-x4.c",
864 "src/f32-vbinary/gen/vmin-sse-x8.c",
865 "src/f32-vbinary/gen/vminc-sse-x4.c",
866 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800867 "src/f32-vbinary/gen/vmul-sse-x4.c",
868 "src/f32-vbinary/gen/vmul-sse-x8.c",
869 "src/f32-vbinary/gen/vmulc-sse-x4.c",
870 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800871 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
872 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800873 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
874 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
875 "src/f32-vbinary/gen/vsub-sse-x4.c",
876 "src/f32-vbinary/gen/vsub-sse-x8.c",
877 "src/f32-vbinary/gen/vsubc-sse-x4.c",
878 "src/f32-vbinary/gen/vsubc-sse-x8.c",
879 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
880 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700881 "src/x32-packx/x4-sse.c",
882]
883
884SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800885 "src/f32-argmaxpool/9p8x-sse2-c4.c",
886 "src/f32-argmaxpool/4x-sse2-c4.c",
887 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800888 "src/f32-prelu/gen/sse2-2x4.c",
889 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
892 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
893 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
896 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
897 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
898 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
899 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
900 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
901 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800902 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800903 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800904 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800905 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800906 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
907 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700908 "src/q8-avgpool/mp9p8q-sse2.c",
909 "src/q8-avgpool/up9-sse2.c",
910 "src/q8-igemm/4x4c2-sse2.c",
911 "src/q8-dwconv/up8x9-sse2.c",
912 "src/q8-gavgpool/mp7p7q-sse2.c",
913 "src/q8-gavgpool/up7-sse2.c",
914 "src/q8-gemm/2x4c8-sse2.c",
915 "src/q8-gemm/4x4c2-sse2.c",
916 "src/q8-vadd/sse2.c",
917 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800918 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700919 "src/u8-rmax/sse2.c",
920 "src/x32-pad/x2-sse2.c",
921 "src/x32-zip/x2-sse2.c",
922 "src/x32-zip/x3-sse2.c",
923 "src/x32-zip/x4-sse2.c",
924 "src/x32-zip/xm-sse2.c",
925 "src/x8-zip/x2-sse2.c",
926 "src/x8-zip/x3-sse2.c",
927 "src/x8-zip/x4-sse2.c",
928 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800929 "src/math/exp-sse2-p5.c",
930 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800931 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700932]
933
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800934SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800935 "src/f32-prelu/gen/sse41-2x4.c",
936 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800937 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800938 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800939 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800940 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800941 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
942 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800943]
944
Marat Dukhan08c4a432019-10-03 09:29:21 -0700945AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800946 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800947 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
948 "src/f32-dwconv/gen/up16x4-avx.c",
949 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
950 "src/f32-dwconv/gen/up8x4-avx.c",
951 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
952 "src/f32-dwconv/gen/up16x9-avx.c",
953 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
954 "src/f32-dwconv/gen/up8x9-avx.c",
955 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
956 "src/f32-dwconv/gen/up16x25-avx.c",
957 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
958 "src/f32-dwconv/gen/up8x25-avx.c",
959 "src/f32-gemm/gen/1x8-avx-broadcast.c",
960 "src/f32-gemm/gen/4x8-avx-broadcast.c",
961 "src/f32-gemm/gen/5x8-avx-broadcast.c",
962 "src/f32-gemm/gen/6x8-avx-broadcast.c",
963 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800964 "src/f32-gemm/gen/1x16-avx-broadcast.c",
965 "src/f32-gemm/gen/3x16-avx-broadcast.c",
966 "src/f32-gemm/gen/4x16-avx-broadcast.c",
967 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800968 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
969 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
970 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
971 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
972 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800973 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
974 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
975 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
976 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800977 "src/f32-hswish/gen/avx-x8.c",
978 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800979 "src/f32-igemm/gen/1x8-avx-broadcast.c",
980 "src/f32-igemm/gen/4x8-avx-broadcast.c",
981 "src/f32-igemm/gen/5x8-avx-broadcast.c",
982 "src/f32-igemm/gen/6x8-avx-broadcast.c",
983 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800984 "src/f32-igemm/gen/1x16-avx-broadcast.c",
985 "src/f32-igemm/gen/3x16-avx-broadcast.c",
986 "src/f32-igemm/gen/4x16-avx-broadcast.c",
987 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -0800989 "src/f32-vbinary/gen/vadd-avx-x8.c",
990 "src/f32-vbinary/gen/vadd-avx-x16.c",
991 "src/f32-vbinary/gen/vaddc-avx-x8.c",
992 "src/f32-vbinary/gen/vaddc-avx-x16.c",
993 "src/f32-vbinary/gen/vdiv-avx-x8.c",
994 "src/f32-vbinary/gen/vdiv-avx-x16.c",
995 "src/f32-vbinary/gen/vdivc-avx-x8.c",
996 "src/f32-vbinary/gen/vdivc-avx-x16.c",
997 "src/f32-vbinary/gen/vmax-avx-x8.c",
998 "src/f32-vbinary/gen/vmax-avx-x16.c",
999 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1000 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1001 "src/f32-vbinary/gen/vmin-avx-x8.c",
1002 "src/f32-vbinary/gen/vmin-avx-x16.c",
1003 "src/f32-vbinary/gen/vminc-avx-x8.c",
1004 "src/f32-vbinary/gen/vminc-avx-x16.c",
1005 "src/f32-vbinary/gen/vmul-avx-x8.c",
1006 "src/f32-vbinary/gen/vmul-avx-x16.c",
1007 "src/f32-vbinary/gen/vmulc-avx-x8.c",
1008 "src/f32-vbinary/gen/vmulc-avx-x16.c",
1009 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
1010 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
1011 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
1012 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
1013 "src/f32-vbinary/gen/vsub-avx-x8.c",
1014 "src/f32-vbinary/gen/vsub-avx-x16.c",
1015 "src/f32-vbinary/gen/vsubc-avx-x8.c",
1016 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001017 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001018]
1019
Marat Dukhanfda12b82019-11-21 12:27:59 -08001020FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001021 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
1022 "src/f32-dwconv/gen/up16x4-fma3.c",
1023 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
1024 "src/f32-dwconv/gen/up8x4-fma3.c",
1025 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
1026 "src/f32-dwconv/gen/up16x9-fma3.c",
1027 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
1028 "src/f32-dwconv/gen/up8x9-fma3.c",
1029 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
1030 "src/f32-dwconv/gen/up16x25-fma3.c",
1031 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
1032 "src/f32-dwconv/gen/up8x25-fma3.c",
1033 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
1034 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
1035 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
1036 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
1037 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
1038 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001039 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
1040 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
1041 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
1042 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001043 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
1044 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
1045 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
1046 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001047 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
1048 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
1049 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
1050 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
1051 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
1052 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001053 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
1054 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
1055 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
1056 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001057 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
1058 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
1059 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
1060 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001061 "src/f32-hswish/gen/fma3-x8.c",
1062 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001063 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
1064 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
1065 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
1066 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
1067 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
1068 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -08001069 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
1070 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
1071 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
1072 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -08001073 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
1074 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
1075 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
1076 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001077]
1078
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001079AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001080 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1081 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1082 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1083 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1084 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1085 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1086 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1087 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1088 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1089 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1090 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1091 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1092 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1093 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1094 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1095 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1096 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1097 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1098 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1099 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1100 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1101 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1102 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1103 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1104 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1105 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1106 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1107 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1108 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1109 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1110 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1111 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1112 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1113 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1114 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1115 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001116 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1117 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1118 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1119 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1120 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1121 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1122 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1123 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1124 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1125 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1126 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1127 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1128 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1129 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1130 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1131 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1132 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1133 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1134 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1135 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1136 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1137 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1138 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1139 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1140 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1141 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1142 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1143 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1144 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1145 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001146 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1147 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1148 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1149 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1150 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1151 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1152 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1153 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1154 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1155 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1156 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1157 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1158 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1159 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1160 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1161 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1162 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1163 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1164 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1165 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1166 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1167 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1168 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1169 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001170 "src/math/exp-avx2-p5.c",
1171 "src/math/exp-avx2-perm-p3.c",
1172 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001173 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001174 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001175 "src/math/sigmoid-avx2-rr2-p5-div.c",
1176 "src/math/sigmoid-avx2-rr1-p5-div.c",
1177 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1178 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1179 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1180 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001181]
1182
Marat Dukhan08c4a432019-10-03 09:29:21 -07001183AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -08001184 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -08001185 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
1186 "src/f32-dwconv/gen/up32x4-avx512f.c",
1187 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
1188 "src/f32-dwconv/gen/up16x4-avx512f.c",
1189 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
1190 "src/f32-dwconv/gen/up32x9-avx512f.c",
1191 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
1192 "src/f32-dwconv/gen/up16x9-avx512f.c",
1193 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
1194 "src/f32-dwconv/gen/up32x25-avx512f.c",
1195 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
1196 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001197 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
1198 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
1199 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
1200 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
1201 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
1202 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
1203 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
1204 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
1205 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
1206 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
1207 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
1208 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001209 "src/f32-hswish/gen/avx512f-x16.c",
1210 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -08001211 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
1212 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
1213 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
1214 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
1215 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
1216 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001217 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1218 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1219 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1220 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1221 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1222 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1223 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1224 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1225 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1226 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1227 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1228 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1229 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1230 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1231 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1232 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1233 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1234 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1235 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1236 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1237 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1238 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1239 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1240 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1241 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1242 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1243 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1244 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1245 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1246 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1247 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1248 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1249 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1250 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1251 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1252 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001253 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001254 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1255 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1256 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1257 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1258 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1259 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1260 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1261 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1262 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1263 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1264 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1265 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1266 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1267 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1268 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1269 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1270 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1271 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1272 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1273 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1274 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1275 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1276 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1277 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1278 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1279 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1280 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1281 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001282 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001283 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1284 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1285 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1286 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1287 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1288 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1289 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1290 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1291 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1292 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1293 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1294 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1295 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1296 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1297 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1298 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1299 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1300 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1301 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1302 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1303 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1304 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1305 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1306 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001307 "src/math/exp-avx512f-p5-scalef.c",
1308 "src/math/exp-avx512f-p5.c",
1309 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001310 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001311 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001312]
1313
1314AARCH32_ASM_UKERNELS = [
1315 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001316 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001317 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1318 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001319 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001320]
1321
1322AARCH64_ASM_UKERNELS = [
1323 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1324 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001325 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1326 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1327 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1328 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1329 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1330 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1331 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1332 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1333 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1334 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001335 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001336 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1337 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001338 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001339 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001340 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1341 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1342 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1343 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1344 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1345 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1346 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1347 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1348 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1349 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1350 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1351 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1352 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001353 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001354 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1355 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001356 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001357 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001358 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1359 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1360 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001361 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001362 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001363 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1364 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001365 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001366 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001367 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1368 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1369 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1370 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001371 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001372 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001373 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1374 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001375]
1376
1377INTERNAL_MICROKERNEL_HDRS = [
1378 "src/xnnpack/argmaxpool.h",
1379 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001380 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001381 "src/xnnpack/clamp.h",
1382 "src/xnnpack/common.h",
1383 "src/xnnpack/conv.h",
1384 "src/xnnpack/dwconv.h",
1385 "src/xnnpack/gavgpool.h",
1386 "src/xnnpack/gemm.h",
1387 "src/xnnpack/hswish.h",
1388 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001389 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001390 "src/xnnpack/lut.h",
1391 "src/xnnpack/math.h",
1392 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001393 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001394 "src/xnnpack/packx.h",
1395 "src/xnnpack/pad.h",
1396 "src/xnnpack/params.h",
1397 "src/xnnpack/pavgpool.h",
1398 "src/xnnpack/ppmm.h",
1399 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001400 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001401 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001402 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403 "src/xnnpack/rmax.h",
1404 "src/xnnpack/scalar-utils.h",
1405 "src/xnnpack/spmm.h",
1406 "src/xnnpack/unpool.h",
1407 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001408 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001409 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001410 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001411 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001412 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001413 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414 "src/xnnpack/zip.h",
1415]
1416
1417INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1418 "include/xnnpack.h",
1419 "src/xnnpack/allocator.h",
1420 "src/xnnpack/compute.h",
1421 "src/xnnpack/im2col.h",
1422 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001423 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001424 "src/xnnpack/operator.h",
1425 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001426 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001427 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001428 "src/xnnpack/requantization.h",
1429]
1430
1431ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1432 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001433]
1434
1435MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001436 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001437 "include/xnnpack.h",
1438]
1439
1440MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1441 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001442 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001443 "src/xnnpack/requantization.h",
1444 "include/xnnpack.h",
1445]
1446
1447OPERATOR_TEST_PARAMS_HDRS = [
1448 "src/xnnpack/params.h",
1449 "src/xnnpack/common.h",
1450]
1451
1452WEIGHTS_PACK_HDRS = [
1453 "src/xnnpack/pack.h",
1454 "src/xnnpack/operator.h",
1455 "src/xnnpack/compute.h",
1456]
1457
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001458LOGGING_COPTS = select({
1459 # No logging in optimized mode
1460 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1461 # Full logging in debug mode
1462 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1463 # Error-only logging in default (fastbuild) mode
1464 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1465})
1466
1467LOGGING_HDRS = [
1468 "src/xnnpack/log.h",
1469]
1470
Marat Dukhan08c4a432019-10-03 09:29:21 -07001471xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001472 name = "tables",
1473 srcs = TABLE_SRCS,
1474 hdrs = INTERNAL_HDRS,
1475 copts = xnnpack_std_copts(),
1476)
1477
1478xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001479 name = "scalar_ukernels",
1480 srcs = SCALAR_UKERNELS,
1481 hdrs = INTERNAL_HDRS,
1482 aarch32_copts = ["-marm"],
1483 copts = xnnpack_std_copts(),
1484 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001485 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001486 "@FP16",
1487 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001488 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001489 ],
1490)
1491
1492xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001493 name = "wasm_ukernels",
1494 hdrs = INTERNAL_HDRS,
1495 copts = xnnpack_std_copts(),
1496 wasm_srcs = WASM_UKERNELS,
1497 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001498 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001499 "@FP16",
1500 "@FXdiv",
1501 "@pthreadpool",
1502 ],
1503)
1504
1505xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001506 name = "psimd_fastmath_ukernels",
1507 srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508 hdrs = INTERNAL_HDRS,
1509 aarch32_copts = [
1510 "-marm",
1511 "-mfpu=neon",
1512 ],
1513 copts = xnnpack_std_copts(),
1514 optimized_copts = [
1515 "-O3",
1516 "-ffast-math",
1517 ],
1518 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001519 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001520 "@FP16",
1521 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001522 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001523 ],
1524)
1525
1526xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001527 name = "psimd_accmath_ukernels",
1528 srcs = PSIMD_ACCMATH_UKERNELS,
1529 hdrs = INTERNAL_HDRS,
1530 aarch32_copts = [
1531 "-marm",
1532 "-mfpu=neon",
1533 ],
1534 copts = xnnpack_std_copts(),
1535 optimized_copts = [
1536 "-O3",
1537 ],
1538 deps = [
1539 ":tables",
1540 "@FP16",
1541 "@psimd",
1542 "@pthreadpool",
1543 ],
1544)
1545
1546xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001547 name = "neon_ukernels",
1548 hdrs = INTERNAL_HDRS,
1549 aarch32_copts = [
1550 "-marm",
1551 "-mfpu=neon",
1552 ],
1553 aarch32_srcs = NEON_UKERNELS,
1554 aarch64_srcs = NEON_UKERNELS,
1555 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001556 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001557 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001558 "@FP16",
1559 "@pthreadpool",
1560 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001561)
1562
1563xnnpack_cc_library(
1564 name = "neonfma_ukernels",
1565 hdrs = INTERNAL_HDRS,
1566 aarch32_copts = [
1567 "-marm",
1568 "-mfpu=neon-vfpv4",
1569 ],
1570 aarch32_srcs = NEONFMA_UKERNELS,
1571 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1572 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001573 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001574 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001575 "@FP16",
1576 "@pthreadpool",
1577 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001578)
1579
1580xnnpack_cc_library(
1581 name = "neonfp16arith_ukernels",
1582 hdrs = INTERNAL_HDRS,
1583 aarch64_copts = ["-march=armv8.2-a+fp16"],
1584 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1585 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001586 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001587 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001588 "@FP16",
1589 "@pthreadpool",
1590 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001591)
1592
1593xnnpack_cc_library(
1594 name = "sse2_ukernels",
1595 hdrs = INTERNAL_HDRS,
1596 copts = xnnpack_std_copts(),
1597 x86_copts = ["-msse2"],
1598 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001599 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001600 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001601 "@FP16",
1602 "@pthreadpool",
1603 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604)
1605
1606xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001607 name = "sse41_ukernels",
1608 hdrs = INTERNAL_HDRS,
1609 copts = xnnpack_std_copts(),
1610 x86_copts = ["-msse4.1"],
1611 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001612 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001613 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001614 "@FP16",
1615 "@pthreadpool",
1616 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001617)
1618
1619xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001620 name = "avx_ukernels",
1621 hdrs = INTERNAL_HDRS,
1622 copts = xnnpack_std_copts(),
1623 x86_copts = ["-mavx"],
1624 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001625 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001626 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001627 "@FP16",
1628 "@pthreadpool",
1629 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630)
1631
1632xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001633 name = "fma3_ukernels",
1634 hdrs = INTERNAL_HDRS,
1635 copts = xnnpack_std_copts(),
1636 x86_copts = [
1637 "-mfma",
1638 ],
1639 x86_srcs = FMA3_UKERNELS,
1640 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001641 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001642 "@FP16",
1643 "@pthreadpool",
1644 ],
1645)
1646
1647xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001648 name = "avx2_ukernels",
1649 hdrs = INTERNAL_HDRS,
1650 copts = xnnpack_std_copts(),
1651 x86_copts = [
1652 "-mfma",
1653 "-mavx2",
1654 ],
1655 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001656 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001657 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001658 "@FP16",
1659 "@pthreadpool",
1660 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001661)
1662
1663xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664 name = "avx512f_ukernels",
1665 hdrs = INTERNAL_HDRS,
1666 copts = xnnpack_std_copts(),
1667 x86_copts = ["-mavx512f"],
1668 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001669 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001670 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001671 "@FP16",
1672 "@pthreadpool",
1673 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001674)
1675
1676xnnpack_cc_library(
1677 name = "asm_ukernels",
1678 hdrs = ["src/xnnpack/assembly.h"],
1679 aarch32_srcs = AARCH32_ASM_UKERNELS,
1680 aarch64_srcs = AARCH64_ASM_UKERNELS,
1681)
1682
1683xnnpack_aggregate_library(
1684 name = "ukernels",
1685 aarch32_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001686 ":psimd_fastmath_ukernels",
1687 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001688 ":neon_ukernels",
1689 ":neonfma_ukernels",
1690 ":asm_ukernels",
1691 ],
1692 aarch64_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001693 ":psimd_fastmath_ukernels",
1694 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001695 ":neon_ukernels",
1696 ":neonfma_ukernels",
1697 ":neonfp16arith_ukernels",
1698 ":asm_ukernels",
1699 ],
1700 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001701 wasm_deps = [
1702 ":wasm_ukernels",
1703 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001704 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001705 ":wasm_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001706 ":psimd_fastmath_ukernels",
1707 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001708 ],
1709 x86_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001710 ":psimd_fastmath_ukernels",
1711 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001712 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001713 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001714 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001715 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001716 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001717 ":avx512f_ukernels",
1718 ],
1719)
1720
1721xnnpack_cc_library(
1722 name = "im2col",
1723 srcs = ["src/im2col.c"],
1724 hdrs = [
1725 "src/xnnpack/common.h",
1726 "src/xnnpack/im2col.h",
1727 ],
1728 copts = xnnpack_std_copts(),
1729)
1730
1731xnnpack_cc_library(
1732 name = "indirection",
1733 srcs = ["src/indirection.c"],
1734 hdrs = INTERNAL_HDRS,
1735 copts = xnnpack_std_copts(),
1736 deps = [
1737 "@FP16",
1738 "@FXdiv",
1739 "@pthreadpool",
1740 ],
1741)
1742
1743xnnpack_cc_library(
1744 name = "operator_run",
1745 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001746 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1747 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001748 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1749 "-Wno-vla",
1750 ],
1751 deps = [
1752 "@FP16",
1753 "@FXdiv",
1754 "@clog",
1755 "@pthreadpool",
1756 ],
1757)
1758
1759cc_library(
1760 name = "enable_assembly",
1761 defines = select({
1762 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1763 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001764 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001765 }),
1766)
1767
Marat Dukhancf056b22019-10-07 10:26:29 -07001768xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 name = "operators",
1770 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001771 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001772 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001773 ],
1774 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001775 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776 "-Isrc",
1777 "-Iinclude",
1778 ] + select({
1779 ":debug_build": [],
1780 "//conditions:default": xnnpack_min_size_copts(),
1781 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001782 wasm_srcs = ["src/wasm-stubs.c"],
1783 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001784 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001785 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001786 "@FP16",
1787 "@FXdiv",
1788 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001789 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001790 ],
1791)
1792
1793cc_library(
1794 name = "XNNPACK",
1795 srcs = [
1796 "src/init.c",
1797 ],
1798 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1799 "-Isrc",
1800 "-Iinclude",
1801 ] + select({
1802 ":debug_build": [],
1803 "//conditions:default": xnnpack_min_size_copts(),
1804 }),
1805 includes = ["include"],
1806 linkstatic = True,
1807 textual_hdrs = ["include/xnnpack.h"],
1808 visibility = xnnpack_visibility(),
1809 deps = [
1810 ":enable_assembly",
1811 ":ukernels",
1812 ":operator_run",
1813 ":operators",
1814 "@clog",
1815 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001816 ] + select({
1817 ":emscripten": [],
1818 "//conditions:default": ["@cpuinfo"],
1819 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820)
1821
1822cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001823 name = "xnnpack_operators_nhwc_f32",
1824 srcs = [
1825 "src/init.c",
1826 ],
1827 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1828 "-Isrc",
1829 "-Iinclude",
1830 ] + select({
1831 ":debug_build": [],
1832 "//conditions:default": xnnpack_min_size_copts(),
1833 }),
1834 defines = [
1835 "XNN_NO_Q8_OPERATORS",
1836 "XNN_NO_U8_OPERATORS",
1837 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001838 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001839 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001840 includes = ["include"],
1841 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001842 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001843 visibility = xnnpack_visibility(),
1844 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001845 ":enable_assembly",
1846 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001847 ":operator_run",
1848 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001849 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001850 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001851 ] + select({
1852 ":emscripten": [],
1853 "//conditions:default": ["@cpuinfo"],
1854 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855)
1856
Marat Dukhancf056b22019-10-07 10:26:29 -07001857xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001858 name = "bench_utils",
1859 srcs = ["bench/utils.cc"],
1860 hdrs = ["bench/utils.h"],
1861 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001862 deps = [
1863 "@com_google_benchmark//:benchmark",
1864 "@cpuinfo",
1865 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001866)
1867
Frank Barchard7e955972019-10-11 10:34:25 -07001868######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001869
1870xnnpack_benchmark(
1871 name = "q8_gemm_bench",
1872 srcs = [
1873 "bench/gemm.h",
1874 "bench/q8-gemm.cc",
1875 "src/xnnpack/AlignedAllocator.h",
1876 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1877 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1878 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1879)
1880
1881xnnpack_benchmark(
1882 name = "f16_gemm_bench",
1883 srcs = [
1884 "bench/f16-gemm.cc",
1885 "bench/gemm.h",
1886 "src/xnnpack/AlignedAllocator.h",
1887 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1888 copts = ["-Wno-unused-function"],
1889 deps = MICROKERNEL_BENCHMARK_DEPS,
1890)
1891
1892xnnpack_benchmark(
1893 name = "f32_igemm_bench",
1894 srcs = [
1895 "bench/f32-igemm.cc",
1896 "bench/conv.h",
1897 "src/xnnpack/AlignedAllocator.h",
1898 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001899 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900)
1901
1902xnnpack_benchmark(
1903 name = "f32_conv_hwc_bench",
1904 srcs = [
1905 "bench/f32-conv-hwc.cc",
1906 "bench/dconv.h",
1907 "src/xnnpack/AlignedAllocator.h",
1908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1909 copts = ["-Wno-unused-function"],
1910 deps = MICROKERNEL_BENCHMARK_DEPS,
1911)
1912
1913xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001914 name = "f32_conv_hwc2spchw_bench",
1915 srcs = [
1916 "bench/f32-conv-hwc2spchw.cc",
1917 "bench/dconv.h",
1918 "src/xnnpack/AlignedAllocator.h",
1919 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1920 copts = ["-Wno-unused-function"],
1921 deps = MICROKERNEL_BENCHMARK_DEPS,
1922)
1923
1924xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925 name = "f32_dwconv_bench",
1926 srcs = [
1927 "bench/f32-dwconv.cc",
1928 "bench/dwconv.h",
1929 "src/xnnpack/AlignedAllocator.h",
1930 ] + MICROKERNEL_BENCHMARK_HDRS,
1931 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1932)
1933
1934xnnpack_benchmark(
1935 name = "f32_dwconv_spchw_bench",
1936 srcs = [
1937 "bench/f32-dwconv-spchw.cc",
1938 "bench/dwconv.h",
1939 "src/xnnpack/AlignedAllocator.h",
1940 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1941 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1942)
1943
1944xnnpack_benchmark(
1945 name = "f32_gemm_bench",
1946 srcs = [
1947 "bench/f32-gemm.cc",
1948 "bench/gemm.h",
1949 "src/xnnpack/AlignedAllocator.h",
1950 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1951 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001952 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001953)
1954
1955xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001956 name = "f32_raddexpminusmax_bench",
1957 srcs = [
1958 "bench/f32-raddexpminusmax.cc",
1959 "src/xnnpack/AlignedAllocator.h",
1960 ] + MICROKERNEL_BENCHMARK_HDRS,
1961 deps = MICROKERNEL_BENCHMARK_DEPS,
1962)
1963
1964xnnpack_benchmark(
1965 name = "f32_raddextexp_bench",
1966 srcs = [
1967 "bench/f32-raddextexp.cc",
1968 "src/xnnpack/AlignedAllocator.h",
1969 ] + MICROKERNEL_BENCHMARK_HDRS,
1970 deps = MICROKERNEL_BENCHMARK_DEPS,
1971)
1972
1973xnnpack_benchmark(
1974 name = "f32_raddstoreexpminusmax_bench",
1975 srcs = [
1976 "bench/f32-raddstoreexpminusmax.cc",
1977 "src/xnnpack/AlignedAllocator.h",
1978 ] + MICROKERNEL_BENCHMARK_HDRS,
1979 deps = MICROKERNEL_BENCHMARK_DEPS,
1980)
1981
1982xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001983 name = "f32_rmax_bench",
1984 srcs = [
1985 "bench/f32-rmax.cc",
1986 "src/xnnpack/AlignedAllocator.h",
1987 ] + MICROKERNEL_BENCHMARK_HDRS,
1988 deps = MICROKERNEL_BENCHMARK_DEPS,
1989)
1990
1991xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001992 name = "f32_sigmoid_bench",
1993 srcs = [
1994 "bench/f32-sigmoid.cc",
1995 "src/xnnpack/AlignedAllocator.h",
1996 ] + MICROKERNEL_BENCHMARK_HDRS,
1997 copts = ["-Wno-unused-function"],
1998 deps = MICROKERNEL_BENCHMARK_DEPS,
1999)
2000
2001xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002002 name = "f32_spmm_bench",
2003 srcs = [
2004 "bench/f32-spmm.cc",
2005 "bench/gemm.h",
2006 "src/xnnpack/AlignedAllocator.h",
2007 ] + MICROKERNEL_BENCHMARK_HDRS,
2008 copts = ["-Wno-unused-function"],
2009 deps = MICROKERNEL_BENCHMARK_DEPS,
2010)
2011
2012xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07002013 name = "f32_softargmax_bench",
2014 srcs = [
2015 "bench/f32-softargmax.cc",
2016 ] + MICROKERNEL_BENCHMARK_HDRS,
2017 copts = ["-Wno-unused-function"],
2018 deps = MICROKERNEL_BENCHMARK_DEPS,
2019)
2020
2021xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002022 name = "f32_vscaleexpminusmax_bench",
2023 srcs = [
2024 "bench/f32-vscaleexpminusmax.cc",
2025 "src/xnnpack/AlignedAllocator.h",
2026 ] + MICROKERNEL_BENCHMARK_HDRS,
2027 deps = MICROKERNEL_BENCHMARK_DEPS,
2028)
2029
2030xnnpack_benchmark(
2031 name = "f32_vscaleextexp_bench",
2032 srcs = [
2033 "bench/f32-vscaleextexp.cc",
2034 "src/xnnpack/AlignedAllocator.h",
2035 ] + MICROKERNEL_BENCHMARK_HDRS,
2036 deps = MICROKERNEL_BENCHMARK_DEPS,
2037)
2038
2039xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002040 name = "f32_im2col_gemm_bench",
2041 srcs = [
2042 "bench/f32-im2col-gemm.cc",
2043 "bench/conv.h",
2044 "src/xnnpack/AlignedAllocator.h",
2045 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
2046 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
2047)
2048
2049########################### Benchmarks for operators ###########################
2050
2051xnnpack_benchmark(
2052 name = "add_bench",
2053 srcs = ["bench/add.cc"],
2054 deps = OPERATOR_BENCHMARK_DEPS,
2055)
2056
2057xnnpack_benchmark(
2058 name = "average_pooling_bench",
2059 srcs = ["bench/average-pooling.cc"],
2060 deps = OPERATOR_BENCHMARK_DEPS,
2061)
2062
2063xnnpack_benchmark(
2064 name = "channel_shuffle_bench",
2065 srcs = ["bench/channel-shuffle.cc"],
2066 deps = OPERATOR_BENCHMARK_DEPS,
2067)
2068
2069xnnpack_benchmark(
2070 name = "convolution_bench",
2071 srcs = ["bench/convolution.cc"],
2072 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
2073 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
2074)
2075
2076xnnpack_benchmark(
2077 name = "deconvolution_bench",
2078 srcs = ["bench/deconvolution.cc"],
2079 copts = xnnpack_optional_tflite_copts(),
2080 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2081)
2082
2083xnnpack_benchmark(
2084 name = "global_average_pooling_bench",
2085 srcs = ["bench/global-average-pooling.cc"],
2086 deps = OPERATOR_BENCHMARK_DEPS,
2087)
2088
2089xnnpack_benchmark(
2090 name = "max_pooling_bench",
2091 srcs = ["bench/max-pooling.cc"],
2092 deps = OPERATOR_BENCHMARK_DEPS,
2093)
2094
2095xnnpack_benchmark(
2096 name = "sigmoid_bench",
2097 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08002098 copts = xnnpack_optional_tflite_copts(),
2099 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002100)
2101
2102xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07002103 name = "prelu_bench",
2104 srcs = ["bench/prelu.cc"],
2105 copts = xnnpack_optional_tflite_copts(),
2106 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
2107)
2108
2109xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002110 name = "softargmax_bench",
2111 srcs = ["bench/softargmax.cc"],
2112 deps = OPERATOR_BENCHMARK_DEPS,
2113)
2114
Marat Dukhanc068bb62019-10-04 13:24:39 -07002115############################# End-to-end benchmarks ############################
2116
2117cc_library(
2118 name = "mobilenet_v1",
2119 srcs = ["models/mobilenet-v1.cc"],
2120 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002121 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002122 linkstatic = True,
2123 deps = [
2124 ":XNNPACK",
2125 "@pthreadpool",
2126 ],
2127)
2128
2129cc_library(
2130 name = "mobilenet_v2",
2131 srcs = ["models/mobilenet-v2.cc"],
2132 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002133 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07002134 linkstatic = True,
2135 deps = [
2136 ":XNNPACK",
2137 "@pthreadpool",
2138 ],
2139)
2140
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002141cc_library(
2142 name = "mobilenet_v3_large",
2143 srcs = ["models/mobilenet-v3-large.cc"],
2144 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002145 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002146 linkstatic = True,
2147 deps = [
2148 ":XNNPACK",
2149 "@pthreadpool",
2150 ],
2151)
2152
2153cc_library(
2154 name = "mobilenet_v3_small",
2155 srcs = ["models/mobilenet-v3-small.cc"],
2156 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08002157 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002158 linkstatic = True,
2159 deps = [
2160 ":XNNPACK",
2161 "@pthreadpool",
2162 ],
2163)
2164
Marat Dukhanc068bb62019-10-04 13:24:39 -07002165xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07002166 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002167 srcs = [
2168 "bench/f32-dwconv-e2e.cc",
2169 "bench/end2end.h",
2170 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07002171 copts = ["-Wno-unused-function"],
2172 deps = MICROKERNEL_BENCHMARK_DEPS + [
2173 ":XNNPACK",
2174 ":mobilenet_v1",
2175 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002176 ":mobilenet_v3_large",
2177 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07002178 ],
2179)
2180
2181xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07002182 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002183 srcs = [
2184 "bench/f32-gemm-e2e.cc",
2185 "bench/end2end.h",
2186 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07002187 copts = ["-Wno-unused-function"],
2188 deps = MICROKERNEL_BENCHMARK_DEPS + [
2189 ":XNNPACK",
2190 ":mobilenet_v1",
2191 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002192 ":mobilenet_v3_large",
2193 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07002194 ],
2195)
2196
2197xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07002198 name = "end2end_bench",
2199 srcs = ["bench/end2end.cc"],
2200 deps = [
2201 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07002202 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002203 ":mobilenet_v1",
2204 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08002205 ":mobilenet_v3_large",
2206 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07002207 "@pthreadpool",
2208 ],
2209)
2210
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002211#################### Accuracy evaluation for math functions ####################
2212
2213xnnpack_benchmark(
2214 name = "f32_exp_eval",
2215 srcs = [
2216 "eval/f32-exp.cc",
2217 "src/xnnpack/AlignedAllocator.h",
2218 ] + ACCURACY_EVAL_HDRS,
2219 deps = ACCURACY_EVAL_DEPS,
2220)
2221
Marat Dukhan515c9772019-10-17 18:07:57 -07002222xnnpack_benchmark(
2223 name = "f32_expminus_eval",
2224 srcs = [
2225 "eval/f32-expminus.cc",
2226 "src/xnnpack/AlignedAllocator.h",
2227 ] + ACCURACY_EVAL_HDRS,
2228 deps = ACCURACY_EVAL_DEPS,
2229)
2230
Marat Dukhan98ba4412019-10-23 02:14:28 -07002231xnnpack_benchmark(
2232 name = "f32_extexp_eval",
2233 srcs = [
2234 "eval/f32-extexp.cc",
2235 "src/xnnpack/AlignedAllocator.h",
2236 ] + ACCURACY_EVAL_HDRS,
2237 deps = ACCURACY_EVAL_DEPS,
2238)
2239
Marat Dukhan346a9e52019-11-15 09:06:30 -08002240xnnpack_benchmark(
2241 name = "f32_sigmoid_eval",
2242 srcs = [
2243 "eval/f32-sigmoid.cc",
2244 "src/xnnpack/AlignedAllocator.h",
2245 ] + ACCURACY_EVAL_HDRS,
2246 deps = ACCURACY_EVAL_DEPS,
2247)
2248
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249######################### Unit tests for micro-kernels #########################
2250
2251xnnpack_unit_test(
2252 name = "f16_gemm_test",
2253 srcs = [
2254 "test/f16-gemm.cc",
2255 "test/gemm-microkernel-tester.h",
2256 "src/xnnpack/AlignedAllocator.h",
2257 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2258 deps = MICROKERNEL_TEST_DEPS,
2259)
2260
2261xnnpack_unit_test(
2262 name = "f32_argmaxpool_test",
2263 srcs = [
2264 "test/f32-argmaxpool.cc",
2265 "test/argmaxpool-microkernel-tester.h",
2266 "src/xnnpack/AlignedAllocator.h",
2267 ] + MICROKERNEL_TEST_HDRS,
2268 deps = MICROKERNEL_TEST_DEPS,
2269)
2270
2271xnnpack_unit_test(
2272 name = "f32_avgpool_test",
2273 srcs = [
2274 "test/f32-avgpool.cc",
2275 "test/avgpool-microkernel-tester.h",
2276 "src/xnnpack/AlignedAllocator.h",
2277 ] + MICROKERNEL_TEST_HDRS,
2278 deps = MICROKERNEL_TEST_DEPS,
2279)
2280
2281xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002282 name = "f32_bilinear_test",
2283 srcs = [
2284 "test/f32-bilinear.cc",
2285 "test/bilinear-microkernel-tester.h",
2286 "src/xnnpack/AlignedAllocator.h",
2287 ] + MICROKERNEL_TEST_HDRS,
2288 deps = MICROKERNEL_TEST_DEPS,
2289)
2290
2291xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292 name = "f32_clamp_test",
2293 srcs = [
2294 "test/f32-clamp.cc",
2295 "test/clamp-microkernel-tester.h",
2296 ] + MICROKERNEL_TEST_HDRS,
2297 deps = MICROKERNEL_TEST_DEPS,
2298)
2299
2300xnnpack_unit_test(
2301 name = "f32_igemm_test",
2302 srcs = [
2303 "test/f32-igemm.cc",
2304 "test/gemm-microkernel-tester.h",
2305 "src/xnnpack/AlignedAllocator.h",
2306 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2307 deps = MICROKERNEL_TEST_DEPS,
2308)
2309
2310xnnpack_unit_test(
2311 name = "f32_conv_hwc_test",
2312 srcs = [
2313 "test/f32-conv-hwc.cc",
2314 "test/conv-hwc-microkernel-tester.h",
2315 "src/xnnpack/AlignedAllocator.h",
2316 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2317 deps = MICROKERNEL_TEST_DEPS,
2318)
2319
2320xnnpack_unit_test(
2321 name = "f32_conv_hwc2spchw_test",
2322 srcs = [
2323 "test/f32-conv-hwc2spchw.cc",
2324 "test/conv-hwc2spchw-microkernel-tester.h",
2325 "src/xnnpack/AlignedAllocator.h",
2326 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2327 deps = MICROKERNEL_TEST_DEPS,
2328)
2329
2330xnnpack_unit_test(
2331 name = "f32_dwconv_test",
2332 srcs = [
2333 "test/f32-dwconv.cc",
2334 "test/dwconv-microkernel-tester.h",
2335 "src/xnnpack/AlignedAllocator.h",
2336 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2337 deps = MICROKERNEL_TEST_DEPS,
2338)
2339
2340xnnpack_unit_test(
2341 name = "f32_dwconv_spchw_test",
2342 srcs = [
2343 "test/f32-dwconv-spchw.cc",
2344 "test/dwconv-spchw-microkernel-tester.h",
2345 "src/xnnpack/AlignedAllocator.h",
2346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2347 deps = MICROKERNEL_TEST_DEPS,
2348)
2349
2350xnnpack_unit_test(
2351 name = "f32_gavgpool_test",
2352 srcs = [
2353 "test/f32-gavgpool.cc",
2354 "test/gavgpool-microkernel-tester.h",
2355 "src/xnnpack/AlignedAllocator.h",
2356 ] + MICROKERNEL_TEST_HDRS,
2357 deps = MICROKERNEL_TEST_DEPS,
2358)
2359
2360xnnpack_unit_test(
2361 name = "f32_gavgpool_spchw_test",
2362 srcs = [
2363 "test/f32-gavgpool-spchw.cc",
2364 "test/gavgpool-spchw-microkernel-tester.h",
2365 "src/xnnpack/AlignedAllocator.h",
2366 ] + MICROKERNEL_TEST_HDRS,
2367 deps = MICROKERNEL_TEST_DEPS,
2368)
2369
2370xnnpack_unit_test(
2371 name = "f32_gemm_test",
2372 srcs = [
2373 "test/f32-gemm.cc",
2374 "test/gemm-microkernel-tester.h",
2375 "src/xnnpack/AlignedAllocator.h",
2376 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2377 deps = MICROKERNEL_TEST_DEPS,
2378)
2379
2380xnnpack_unit_test(
2381 name = "f32_gemminc_test",
2382 srcs = [
2383 "test/f32-gemminc.cc",
2384 "test/gemm-microkernel-tester.h",
2385 "src/xnnpack/AlignedAllocator.h",
2386 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2387 deps = MICROKERNEL_TEST_DEPS,
2388)
2389
2390xnnpack_unit_test(
2391 name = "f32_hswish_test",
2392 srcs = [
2393 "test/f32-hswish.cc",
2394 "test/hswish-microkernel-tester.h",
2395 ] + MICROKERNEL_TEST_HDRS,
2396 deps = MICROKERNEL_TEST_DEPS,
2397)
2398
2399xnnpack_unit_test(
2400 name = "f32_maxpool_test",
2401 srcs = [
2402 "test/f32-maxpool.cc",
2403 "test/maxpool-microkernel-tester.h",
2404 ] + MICROKERNEL_TEST_HDRS,
2405 deps = MICROKERNEL_TEST_DEPS,
2406)
2407
2408xnnpack_unit_test(
2409 name = "f32_pavgpool_test",
2410 srcs = [
2411 "test/f32-pavgpool.cc",
2412 "test/avgpool-microkernel-tester.h",
2413 "src/xnnpack/AlignedAllocator.h",
2414 ] + MICROKERNEL_TEST_HDRS,
2415 deps = MICROKERNEL_TEST_DEPS,
2416)
2417
2418xnnpack_unit_test(
2419 name = "f32_ppmm_test",
2420 srcs = [
2421 "test/f32-ppmm.cc",
2422 "test/gemm-microkernel-tester.h",
2423 "src/xnnpack/AlignedAllocator.h",
2424 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2425 deps = MICROKERNEL_TEST_DEPS,
2426)
2427
2428xnnpack_unit_test(
2429 name = "f32_prelu_test",
2430 srcs = [
2431 "test/f32-prelu.cc",
2432 "test/prelu-microkernel-tester.h",
2433 "src/xnnpack/AlignedAllocator.h",
2434 ] + MICROKERNEL_TEST_HDRS,
2435 deps = MICROKERNEL_TEST_DEPS,
2436)
2437
2438xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002439 name = "f32_raddexpminusmax_test",
2440 srcs = [
2441 "test/f32-raddexpminusmax.cc",
2442 "test/raddexpminusmax-microkernel-tester.h",
2443 ] + MICROKERNEL_TEST_HDRS,
2444 deps = MICROKERNEL_TEST_DEPS,
2445)
2446
2447xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002448 name = "f32_raddextexp_test",
2449 srcs = [
2450 "test/f32-raddextexp.cc",
2451 "test/raddextexp-microkernel-tester.h",
2452 ] + MICROKERNEL_TEST_HDRS,
2453 deps = MICROKERNEL_TEST_DEPS,
2454)
2455
2456xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002457 name = "f32_raddstoreexpminusmax_test",
2458 srcs = [
2459 "test/f32-raddstoreexpminusmax.cc",
2460 "test/raddstoreexpminusmax-microkernel-tester.h",
2461 ] + MICROKERNEL_TEST_HDRS,
2462 deps = MICROKERNEL_TEST_DEPS,
2463)
2464
2465xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002466 name = "f32_rmax_test",
2467 srcs = [
2468 "test/f32-rmax.cc",
2469 "test/rmax-microkernel-tester.h",
2470 ] + MICROKERNEL_TEST_HDRS,
2471 deps = MICROKERNEL_TEST_DEPS,
2472)
2473
2474xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002475 name = "f32_sigmoid_test",
2476 srcs = [
2477 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002478 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002479 ] + MICROKERNEL_TEST_HDRS,
2480 deps = MICROKERNEL_TEST_DEPS,
2481)
2482
2483xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002484 name = "f32_spmm_test",
2485 srcs = [
2486 "test/f32-spmm.cc",
2487 "test/spmm-microkernel-tester.h",
2488 "src/xnnpack/AlignedAllocator.h",
2489 ] + MICROKERNEL_TEST_HDRS,
2490 deps = MICROKERNEL_TEST_DEPS,
2491)
2492
2493xnnpack_unit_test(
2494 name = "f32_vadd_test",
2495 srcs = [
2496 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002497 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002498 ] + MICROKERNEL_TEST_HDRS,
2499 deps = MICROKERNEL_TEST_DEPS,
2500)
2501
2502xnnpack_unit_test(
2503 name = "f32_vaddc_test",
2504 srcs = [
2505 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002506 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507 ] + MICROKERNEL_TEST_HDRS,
2508 deps = MICROKERNEL_TEST_DEPS,
2509)
2510
2511xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002512 name = "f32_vdiv_test",
2513 srcs = [
2514 "test/f32-vdiv.cc",
2515 "test/vbinary-microkernel-tester.h",
2516 ] + MICROKERNEL_TEST_HDRS,
2517 deps = MICROKERNEL_TEST_DEPS,
2518)
2519
2520xnnpack_unit_test(
2521 name = "f32_vdivc_test",
2522 srcs = [
2523 "test/f32-vdivc.cc",
2524 "test/vbinaryc-microkernel-tester.h",
2525 ] + MICROKERNEL_TEST_HDRS,
2526 deps = MICROKERNEL_TEST_DEPS,
2527)
2528
2529xnnpack_unit_test(
2530 name = "f32_vrdivc_test",
2531 srcs = [
2532 "test/f32-vrdivc.cc",
2533 "test/vbinaryc-microkernel-tester.h",
2534 ] + MICROKERNEL_TEST_HDRS,
2535 deps = MICROKERNEL_TEST_DEPS,
2536)
2537
2538xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002539 name = "f32_vmax_test",
2540 srcs = [
2541 "test/f32-vmax.cc",
2542 "test/vbinary-microkernel-tester.h",
2543 ] + MICROKERNEL_TEST_HDRS,
2544 deps = MICROKERNEL_TEST_DEPS,
2545)
2546
2547xnnpack_unit_test(
2548 name = "f32_vmaxc_test",
2549 srcs = [
2550 "test/f32-vmaxc.cc",
2551 "test/vbinaryc-microkernel-tester.h",
2552 ] + MICROKERNEL_TEST_HDRS,
2553 deps = MICROKERNEL_TEST_DEPS,
2554)
2555
2556xnnpack_unit_test(
2557 name = "f32_vmin_test",
2558 srcs = [
2559 "test/f32-vmin.cc",
2560 "test/vbinary-microkernel-tester.h",
2561 ] + MICROKERNEL_TEST_HDRS,
2562 deps = MICROKERNEL_TEST_DEPS,
2563)
2564
2565xnnpack_unit_test(
2566 name = "f32_vminc_test",
2567 srcs = [
2568 "test/f32-vminc.cc",
2569 "test/vbinaryc-microkernel-tester.h",
2570 ] + MICROKERNEL_TEST_HDRS,
2571 deps = MICROKERNEL_TEST_DEPS,
2572)
2573
2574xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002575 name = "f32_vmul_test",
2576 srcs = [
2577 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002578 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002579 ] + MICROKERNEL_TEST_HDRS,
2580 deps = MICROKERNEL_TEST_DEPS,
2581)
2582
2583xnnpack_unit_test(
2584 name = "f32_vmulc_test",
2585 srcs = [
2586 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002587 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002588 ] + MICROKERNEL_TEST_HDRS,
2589 deps = MICROKERNEL_TEST_DEPS,
2590)
2591
2592xnnpack_unit_test(
2593 name = "f32_vmulcaddc_test",
2594 srcs = [
2595 "test/f32-vmulcaddc.cc",
2596 "test/vmulcaddc-microkernel-tester.h",
2597 "src/xnnpack/AlignedAllocator.h",
2598 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2599 deps = MICROKERNEL_TEST_DEPS,
2600)
2601
2602xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002603 name = "f32_vscale_test",
2604 srcs = [
2605 "test/f32-vscale.cc",
2606 "test/vscale-microkernel-tester.h",
2607 ] + MICROKERNEL_TEST_HDRS,
2608 deps = MICROKERNEL_TEST_DEPS,
2609)
2610
2611xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002612 name = "f32_vscaleexpminusmax_test",
2613 srcs = [
2614 "test/f32-vscaleexpminusmax.cc",
2615 "test/vscaleexpminusmax-microkernel-tester.h",
2616 ] + MICROKERNEL_TEST_HDRS,
2617 deps = MICROKERNEL_TEST_DEPS,
2618)
2619
2620xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002621 name = "f32_vscaleextexp_test",
2622 srcs = [
2623 "test/f32-vscaleextexp.cc",
2624 "test/vscaleextexp-microkernel-tester.h",
2625 ] + MICROKERNEL_TEST_HDRS,
2626 deps = MICROKERNEL_TEST_DEPS,
2627)
2628
2629xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002630 name = "f32_vsub_test",
2631 srcs = [
2632 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002633 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002634 ] + MICROKERNEL_TEST_HDRS,
2635 deps = MICROKERNEL_TEST_DEPS,
2636)
2637
2638xnnpack_unit_test(
2639 name = "f32_vsubc_test",
2640 srcs = [
2641 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002642 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002643 ] + MICROKERNEL_TEST_HDRS,
2644 deps = MICROKERNEL_TEST_DEPS,
2645)
2646
2647xnnpack_unit_test(
2648 name = "f32_vrsubc_test",
2649 srcs = [
2650 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002651 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002652 ] + MICROKERNEL_TEST_HDRS,
2653 deps = MICROKERNEL_TEST_DEPS,
2654)
2655
2656xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 name = "q8_avgpool_test",
2658 srcs = [
2659 "test/q8-avgpool.cc",
2660 "test/avgpool-microkernel-tester.h",
2661 "src/xnnpack/AlignedAllocator.h",
2662 ] + MICROKERNEL_TEST_HDRS,
2663 deps = MICROKERNEL_TEST_DEPS,
2664)
2665
2666xnnpack_unit_test(
2667 name = "q8_igemm_test",
2668 srcs = [
2669 "test/q8-igemm.cc",
2670 "test/gemm-microkernel-tester.h",
2671 "src/xnnpack/AlignedAllocator.h",
2672 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2673 deps = MICROKERNEL_TEST_DEPS,
2674)
2675
2676xnnpack_unit_test(
2677 name = "q8_dwconv_test",
2678 srcs = [
2679 "test/q8-dwconv.cc",
2680 "test/dwconv-microkernel-tester.h",
2681 "src/xnnpack/AlignedAllocator.h",
2682 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2683 deps = MICROKERNEL_TEST_DEPS,
2684)
2685
2686xnnpack_unit_test(
2687 name = "q8_gavgpool_test",
2688 srcs = [
2689 "test/q8-gavgpool.cc",
2690 "test/gavgpool-microkernel-tester.h",
2691 "src/xnnpack/AlignedAllocator.h",
2692 ] + MICROKERNEL_TEST_HDRS,
2693 deps = MICROKERNEL_TEST_DEPS,
2694)
2695
2696xnnpack_unit_test(
2697 name = "q8_gemm_test",
2698 srcs = [
2699 "test/q8-gemm.cc",
2700 "test/gemm-microkernel-tester.h",
2701 "src/xnnpack/AlignedAllocator.h",
2702 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2703 deps = MICROKERNEL_TEST_DEPS,
2704)
2705
2706xnnpack_unit_test(
2707 name = "q8_vadd_test",
2708 srcs = [
2709 "test/q8-vadd.cc",
2710 "test/vadd-microkernel-tester.h",
2711 ] + MICROKERNEL_TEST_HDRS,
2712 deps = MICROKERNEL_TEST_DEPS,
2713)
2714
2715xnnpack_unit_test(
2716 name = "u8_clamp_test",
2717 srcs = [
2718 "test/u8-clamp.cc",
2719 "test/clamp-microkernel-tester.h",
2720 ] + MICROKERNEL_TEST_HDRS,
2721 deps = MICROKERNEL_TEST_DEPS,
2722)
2723
2724xnnpack_unit_test(
2725 name = "u8_lut32norm_test",
2726 srcs = [
2727 "test/u8-lut32norm.cc",
2728 "test/lut-norm-microkernel-tester.h",
2729 ] + MICROKERNEL_TEST_HDRS,
2730 deps = MICROKERNEL_TEST_DEPS,
2731)
2732
2733xnnpack_unit_test(
2734 name = "u8_maxpool_test",
2735 srcs = [
2736 "test/u8-maxpool.cc",
2737 "test/maxpool-microkernel-tester.h",
2738 ] + MICROKERNEL_TEST_HDRS,
2739 deps = MICROKERNEL_TEST_DEPS,
2740)
2741
2742xnnpack_unit_test(
2743 name = "u8_rmax_test",
2744 srcs = [
2745 "test/u8-rmax.cc",
2746 "test/rmax-microkernel-tester.h",
2747 ] + MICROKERNEL_TEST_HDRS,
2748 deps = MICROKERNEL_TEST_DEPS,
2749)
2750
2751xnnpack_unit_test(
2752 name = "x32_packx_test",
2753 srcs = [
2754 "test/x32-packx.cc",
2755 "test/pack-microkernel-tester.h",
2756 "src/xnnpack/AlignedAllocator.h",
2757 ] + MICROKERNEL_TEST_HDRS,
2758 deps = MICROKERNEL_TEST_DEPS,
2759)
2760
2761xnnpack_unit_test(
2762 name = "x32_pad_test",
2763 srcs = [
2764 "test/x32-pad.cc",
2765 "test/pad-microkernel-tester.h",
2766 ] + MICROKERNEL_TEST_HDRS,
2767 deps = MICROKERNEL_TEST_DEPS,
2768)
2769
2770xnnpack_unit_test(
2771 name = "x32_unpool_test",
2772 srcs = [
2773 "test/x32-unpool.cc",
2774 "test/unpool-microkernel-tester.h",
2775 ] + MICROKERNEL_TEST_HDRS,
2776 deps = MICROKERNEL_TEST_DEPS,
2777)
2778
2779xnnpack_unit_test(
2780 name = "x32_zip_test",
2781 srcs = [
2782 "test/x32-zip.cc",
2783 "test/zip-microkernel-tester.h",
2784 ] + MICROKERNEL_TEST_HDRS,
2785 deps = MICROKERNEL_TEST_DEPS,
2786)
2787
2788xnnpack_unit_test(
2789 name = "x8_lut_test",
2790 srcs = [
2791 "test/x8-lut.cc",
2792 "test/lut-microkernel-tester.h",
2793 ] + MICROKERNEL_TEST_HDRS,
2794 deps = MICROKERNEL_TEST_DEPS,
2795)
2796
2797xnnpack_unit_test(
2798 name = "x8_zip_test",
2799 srcs = [
2800 "test/x8-zip.cc",
2801 "test/zip-microkernel-tester.h",
2802 ] + MICROKERNEL_TEST_HDRS,
2803 deps = MICROKERNEL_TEST_DEPS,
2804)
2805
2806########################### Size test for the library ##########################
2807
2808xnnpack_binary(
2809 name = "size_test",
2810 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002811 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002812)
2813
2814########################### Unit tests for operators ###########################
2815
2816xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002817 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002818 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002819 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820 "test/add-operator-tester.h",
2821 ],
2822 deps = OPERATOR_TEST_DEPS,
2823)
2824
2825xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002826 name = "add_nd_test",
2827 srcs = [
2828 "test/add-nd.cc",
2829 "test/binary-elementwise-operator-tester.h",
2830 ],
2831 deps = OPERATOR_TEST_DEPS,
2832)
2833
2834xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002835 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002836 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002837 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002838 "test/argmax-pooling-operator-tester.h",
2839 ] + OPERATOR_TEST_PARAMS_HDRS,
2840 deps = OPERATOR_TEST_DEPS,
2841)
2842
2843xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002844 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002845 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002846 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002847 "test/average-pooling-operator-tester.h",
2848 ] + OPERATOR_TEST_PARAMS_HDRS,
2849 deps = OPERATOR_TEST_DEPS,
2850)
2851
2852xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002853 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002854 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002855 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002856 "test/channel-pad-operator-tester.h",
2857 ] + OPERATOR_TEST_PARAMS_HDRS,
2858 deps = OPERATOR_TEST_DEPS,
2859)
2860
2861xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002862 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002863 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002864 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 "test/channel-shuffle-operator-tester.h",
2866 ],
2867 deps = OPERATOR_TEST_DEPS,
2868)
2869
2870xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002871 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002873 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002874 "test/clamp-operator-tester.h",
2875 ],
2876 deps = OPERATOR_TEST_DEPS,
2877)
2878
2879xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002880 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002882 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002883 "test/convolution-operator-tester.h",
2884 ],
2885 deps = OPERATOR_TEST_DEPS,
2886)
2887
2888xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002889 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002891 "test/convolution-nchw.cc",
2892 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002893 ],
2894 deps = OPERATOR_TEST_DEPS,
2895)
2896
2897xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002898 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002899 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002900 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 "test/deconvolution-operator-tester.h",
2902 ] + OPERATOR_TEST_PARAMS_HDRS,
2903 deps = OPERATOR_TEST_DEPS,
2904)
2905
2906xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002907 name = "divide_nd_test",
2908 srcs = [
2909 "test/binary-elementwise-operator-tester.h",
2910 "test/divide-nd.cc",
2911 ],
2912 deps = OPERATOR_TEST_DEPS,
2913)
2914
2915xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002916 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002917 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002918 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 "test/fully-connected-operator-tester.h",
2920 ],
2921 deps = OPERATOR_TEST_DEPS,
2922)
2923
2924xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002925 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002927 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928 "test/global-average-pooling-operator-tester.h",
2929 ] + OPERATOR_TEST_PARAMS_HDRS,
2930 deps = OPERATOR_TEST_DEPS,
2931)
2932
2933xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002934 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002935 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002936 "test/global-average-pooling-ncw.cc",
2937 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002938 ],
2939 deps = OPERATOR_TEST_DEPS,
2940)
2941
2942xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002943 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002944 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002945 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 "test/hardswish-operator-tester.h",
2947 ],
2948 deps = OPERATOR_TEST_DEPS,
2949)
2950
2951xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002952 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002953 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002954 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002955 "test/leaky-relu-operator-tester.h",
2956 ],
2957 deps = OPERATOR_TEST_DEPS,
2958)
2959
2960xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002961 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002963 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 "test/max-pooling-operator-tester.h",
2965 ] + OPERATOR_TEST_PARAMS_HDRS,
2966 deps = OPERATOR_TEST_DEPS,
2967)
2968
2969xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08002970 name = "maximum_nd_test",
2971 srcs = [
2972 "test/binary-elementwise-operator-tester.h",
2973 "test/maximum-nd.cc",
2974 ],
2975 deps = OPERATOR_TEST_DEPS,
2976)
2977
2978xnnpack_unit_test(
2979 name = "minimum_nd_test",
2980 srcs = [
2981 "test/binary-elementwise-operator-tester.h",
2982 "test/minimum-nd.cc",
2983 ],
2984 deps = OPERATOR_TEST_DEPS,
2985)
2986
2987xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002988 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002989 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002990 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002991 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002992 ],
2993 deps = OPERATOR_TEST_DEPS,
2994)
2995
2996xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002997 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002998 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002999 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 "test/prelu-operator-tester.h",
3001 ] + OPERATOR_TEST_PARAMS_HDRS,
3002 deps = OPERATOR_TEST_DEPS,
3003)
3004
3005xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003006 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08003007 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003008 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08003009 "test/resize-bilinear-operator-tester.h",
3010 ] + OPERATOR_TEST_PARAMS_HDRS,
3011 deps = OPERATOR_TEST_DEPS,
3012)
3013
3014xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003015 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003017 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 "test/sigmoid-operator-tester.h",
3019 ],
3020 deps = OPERATOR_TEST_DEPS,
3021)
3022
3023xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003024 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003026 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 "test/softargmax-operator-tester.h",
3028 ],
3029 deps = OPERATOR_TEST_DEPS,
3030)
3031
3032xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08003033 name = "subtract_nd_test",
3034 srcs = [
3035 "test/binary-elementwise-operator-tester.h",
3036 "test/subtract-nd.cc",
3037 ],
3038 deps = OPERATOR_TEST_DEPS,
3039)
3040
3041xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08003042 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08003044 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045 "test/unpooling-operator-tester.h",
3046 ],
3047 deps = OPERATOR_TEST_DEPS,
3048)
3049
3050############################# Build configurations #############################
3051
Marat Dukhanb8642352019-10-30 15:43:02 -07003052# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07003053config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003054 name = "xnn_enable_assembly_explicit_true",
3055 define_values = {"xnn_enable_assembly": "true"},
3056)
3057
3058# Disables usage of assembly kernels.
3059config_setting(
3060 name = "xnn_enable_assembly_explicit_false",
3061 define_values = {"xnn_enable_assembly": "false"},
3062)
3063
3064# Builds with -c dbg
3065config_setting(
3066 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07003068 "compilation_mode": "dbg",
3069 },
3070)
3071
3072# Builds with -c opt
3073config_setting(
3074 name = "optimized_build",
3075 values = {
3076 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077 },
3078)
3079
3080config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07003081 name = "linux_k8",
3082 values = {"cpu": "k8"},
3083)
3084
3085config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07003086 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07003087 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07003088)
3089
3090config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091 name = "android",
3092 values = {"crosstool_top": "//external:android/crosstool"},
3093)
3094
3095config_setting(
3096 name = "android_armv7",
3097 values = {
3098 "crosstool_top": "//external:android/crosstool",
3099 "cpu": "armeabi-v7a",
3100 },
3101)
3102
3103config_setting(
3104 name = "android_arm64",
3105 values = {
3106 "crosstool_top": "//external:android/crosstool",
3107 "cpu": "arm64-v8a",
3108 },
3109)
3110
3111config_setting(
3112 name = "android_x86",
3113 values = {
3114 "crosstool_top": "//external:android/crosstool",
3115 "cpu": "x86",
3116 },
3117)
3118
3119config_setting(
3120 name = "android_x86_64",
3121 values = {
3122 "crosstool_top": "//external:android/crosstool",
3123 "cpu": "x86_64",
3124 },
3125)
3126
3127config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07003128 name = "macos_x86_64",
3129 values = {
3130 "apple_platform_type": "macos",
3131 "cpu": "darwin",
3132 },
3133)
3134
3135config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003137 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07003138)
3139
3140config_setting(
3141 name = "emscripten_wasm",
3142 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003143 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003144 "cpu": "wasm",
3145 },
3146)
3147
3148config_setting(
3149 name = "emscripten_wasmsimd",
3150 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003151 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07003153 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154 },
3155)
3156
3157config_setting(
3158 name = "emscripten_asmjs",
3159 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07003160 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 "cpu": "asmjs",
3162 },
3163)